Thread: unite recovery.conf and postgresql.conf
Hi, http://archives.postgresql.org/pgsql-hackers/2010-12/msg02343.php In previous discussion, we've reached the consensus that we should unite recovery.conf and postgresql.conf. The attached patch does that. The patch is WIP, I'll have to update the document, but if you notice something, please feel free to comment. This patch allows us to specify recovery parameters like primary_conninfo in postgresql.conf. You can change some of them without restarting the server (i.e., by using "pg_ctl reload") and see the current settings by SHOW command. But note that recovery.conf is still required as a status file for archive recovery and standby server even if you put all the settings in postgresql.conf. Recovery parameters in postgresql.conf only have effect when recovery.conf exists. For backward compatibility, this patch still allows us to specify recovery parameters in recovery.conf. But, as in past years, you cannot reload recovery.conf and see the current settings by SHOW command. We should recommend to put all the settings in postgresql.conf and empty recovery.conf, but you can also put all recovery parameters in recovery.conf. If the same parameter is specified in both file, the setting in recovery.conf overrides that in postgresql.conf. In this case, SHOW command displays the settings in postgresql.conf even though they are not used at all. Even if you change the settings in postgresql.conf and reload it, they have no effect because the settings in recovery.conf are used preferentially. These limitations on recovery.conf might confuse users. But since most users will put all the settings in postgresql.conf if we recommend to do that, I don't think that it's worth complicating the source code for getting rid of those limitations. Thought? Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
Attachment
On Fri, Sep 9, 2011 at 11:56, Fujii Masao <masao.fujii@gmail.com> wrote: > Hi, > > http://archives.postgresql.org/pgsql-hackers/2010-12/msg02343.php > > In previous discussion, we've reached the consensus that we should unite > recovery.conf and postgresql.conf. The attached patch does that. The > patch is WIP, I'll have to update the document, but if you notice something, > please feel free to comment. > > This patch allows us to specify recovery parameters like primary_conninfo > in postgresql.conf. You can change some of them without restarting > the server (i.e., by using "pg_ctl reload") and see the current settings by > SHOW command. But note that recovery.conf is still required as a status > file for archive recovery and standby server even if you put all the settings > in postgresql.conf. Recovery parameters in postgresql.conf only have effect > when recovery.conf exists. Why exactly do we need to keep recovery.conf in this case? > For backward compatibility, this patch still allows us to specify recovery > parameters in recovery.conf. But, as in past years, you cannot reload > recovery.conf and see the current settings by SHOW command. We should > recommend to put all the settings in postgresql.conf and empty recovery.conf, > but you can also put all recovery parameters in recovery.conf. Perhaps we should just have people use recovery.conf as an include file in this case? As a whole, I'm not sure I like the idea of having such critical parameters in two different places. Do we really need to care about the backwards compatibility of something like that this much? > If the same parameter is specified in both file, the setting in recovery.conf > overrides that in postgresql.conf. In this case, SHOW command displays > the settings in postgresql.conf even though they are not used at all. Even if Now, *that* is just broken, IMHO. The SHOW command should show what is actually in use, period. Everything is guaranteed to confuse the hell out of anybody trying to use it. > you change the settings in postgresql.conf and reload it, they have no effect > because the settings in recovery.conf are used preferentially. > > These limitations on recovery.conf might confuse users. But since most > users will put all the settings in postgresql.conf if we recommend to do that, > I don't think that it's worth complicating the source code for getting rid of > those limitations. I think we should just get rid of supporting them in recovery.conf completely. Recovery settings are critical, and making that confusing to the user is not a good thing. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/
On Fri, Sep 9, 2011 at 7:21 PM, Magnus Hagander <magnus@hagander.net> wrote: > On Fri, Sep 9, 2011 at 11:56, Fujii Masao <masao.fujii@gmail.com> wrote: >> Hi, >> >> http://archives.postgresql.org/pgsql-hackers/2010-12/msg02343.php >> >> In previous discussion, we've reached the consensus that we should unite >> recovery.conf and postgresql.conf. The attached patch does that. The >> patch is WIP, I'll have to update the document, but if you notice something, >> please feel free to comment. >> >> This patch allows us to specify recovery parameters like primary_conninfo >> in postgresql.conf. You can change some of them without restarting >> the server (i.e., by using "pg_ctl reload") and see the current settings by >> SHOW command. But note that recovery.conf is still required as a status >> file for archive recovery and standby server even if you put all the settings >> in postgresql.conf. Recovery parameters in postgresql.conf only have effect >> when recovery.conf exists. > > Why exactly do we need to keep recovery.conf in this case? PostgreSQL automatically reinitializes after a backend crash. Imagine the case where this happens after archive recovery. Since restore_command is left set in postgresql.conf and it still has effect (if we've completely removed recovery.conf), the reinitialization will makes the server go into archive recovery mode again unexpectedly. We can avoid this problem by using recovery.conf. At the end of archive recovery, recovery.conf is renamed to recovery.done. So when the reinitialization happens, recovery.conf doesn't exist and restore_command has no effect. Which prevents the server from going into archive recovery mode again. >> For backward compatibility, this patch still allows us to specify recovery >> parameters in recovery.conf. But, as in past years, you cannot reload >> recovery.conf and see the current settings by SHOW command. We should >> recommend to put all the settings in postgresql.conf and empty recovery.conf, >> but you can also put all recovery parameters in recovery.conf. > > Perhaps we should just have people use recovery.conf as an include > file in this case? Yeah, that's my first idea, but I've given up adopting that. The problem is that recovery.conf is renamed to recovery.done at the end of recovery. This means that all the settings in recovery.conf are reset when archive recovery ends. If you run "pg_ctl reload" after archive recovery, you'll get something like the following error message and the reload will always fail; LOG: parameter "standby_mode" cannot be changed without restarting the server > As a whole, I'm not sure I like the idea of having such critical > parameters in two different places. Do we really need to care about > the backwards compatibility of something like that this much? I don't have strong opinion about this. But in previous discussion, Simon argued that we should. http://archives.postgresql.org/pgsql-hackers/2010-10/msg00017.php >> If the same parameter is specified in both file, the setting in recovery.conf >> overrides that in postgresql.conf. In this case, SHOW command displays >> the settings in postgresql.conf even though they are not used at all. Even if > > Now, *that* is just broken, IMHO. The SHOW command should show what is > actually in use, period. Everything is guaranteed to confuse the hell > out of anybody trying to use it. I'm afraid that we have to add very complicated code to fix that problem. Though of course we can easily fix the problem if we don't care about the backward compatibility. Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
On Fri, Sep 9, 2011 at 13:15, Fujii Masao <masao.fujii@gmail.com> wrote: > On Fri, Sep 9, 2011 at 7:21 PM, Magnus Hagander <magnus@hagander.net> wrote: >> On Fri, Sep 9, 2011 at 11:56, Fujii Masao <masao.fujii@gmail.com> wrote: >>> Hi, >>> >>> http://archives.postgresql.org/pgsql-hackers/2010-12/msg02343.php >>> >>> In previous discussion, we've reached the consensus that we should unite >>> recovery.conf and postgresql.conf. The attached patch does that. The >>> patch is WIP, I'll have to update the document, but if you notice something, >>> please feel free to comment. >>> >>> This patch allows us to specify recovery parameters like primary_conninfo >>> in postgresql.conf. You can change some of them without restarting >>> the server (i.e., by using "pg_ctl reload") and see the current settings by >>> SHOW command. But note that recovery.conf is still required as a status >>> file for archive recovery and standby server even if you put all the settings >>> in postgresql.conf. Recovery parameters in postgresql.conf only have effect >>> when recovery.conf exists. >> >> Why exactly do we need to keep recovery.conf in this case? > > PostgreSQL automatically reinitializes after a backend crash. Imagine the case > where this happens after archive recovery. Since restore_command is left set > in postgresql.conf and it still has effect (if we've completely removed > recovery.conf), the reinitialization will makes the server go into > archive recovery > mode again unexpectedly. > > We can avoid this problem by using recovery.conf. At the end of > archive recovery, > recovery.conf is renamed to recovery.done. So when the reinitialization happens, > recovery.conf doesn't exist and restore_command has no effect. Which prevents > the server from going into archive recovery mode again. Ah, ugh. Good point. I have to wonder though, if it wouldn't be less confusing to just get rid of recovery.conf and use a *different* file for this. Just to make it clear it's not a config file, but just a boolean exists/notexists state. >>> For backward compatibility, this patch still allows us to specify recovery >>> parameters in recovery.conf. But, as in past years, you cannot reload >>> recovery.conf and see the current settings by SHOW command. We should >>> recommend to put all the settings in postgresql.conf and empty recovery.conf, >>> but you can also put all recovery parameters in recovery.conf. >> >> Perhaps we should just have people use recovery.conf as an include >> file in this case? > > Yeah, that's my first idea, but I've given up adopting that. The problem is that > recovery.conf is renamed to recovery.done at the end of recovery. This means > that all the settings in recovery.conf are reset when archive recovery ends. > If you run "pg_ctl reload" after archive recovery, you'll get something like the > following error message and the reload will always fail; > > LOG: parameter "standby_mode" cannot be changed without > restarting the server Good point. And I believe another argument for actually using completely different files ;) >> As a whole, I'm not sure I like the idea of having such critical >> parameters in two different places. Do we really need to care about >> the backwards compatibility of something like that this much? > > I don't have strong opinion about this. But in previous discussion, Simon argued > that we should. > http://archives.postgresql.org/pgsql-hackers/2010-10/msg00017.php I see his argument, but I think reduced confusion is a more important one. >>> If the same parameter is specified in both file, the setting in recovery.conf >>> overrides that in postgresql.conf. In this case, SHOW command displays >>> the settings in postgresql.conf even though they are not used at all. Even if >> >> Now, *that* is just broken, IMHO. The SHOW command should show what is >> actually in use, period. Everything is guaranteed to confuse the hell >> out of anybody trying to use it. > > I'm afraid that we have to add very complicated code to fix that problem. > Though of course we can easily fix the problem if we don't care about > the backward compatibility. That is an even bigger reason to drop backwards compatibility. Unless someone else can come up with a neat way of fixing it. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/
Magnus Hagander <magnus@hagander.net> writes:
> I have to wonder though, if it wouldn't be less confusing to just get
> rid of recovery.conf and use a *different* file for this. Just to make
> it clear it's not a config file, but just a boolean exists/notexists
> state.
+1. If it's not a configuration file anymore, it shouldn't be called
one.
regards, tom lane
On Fri, Sep 9, 2011 at 3:05 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > Magnus Hagander <magnus@hagander.net> writes: >> I have to wonder though, if it wouldn't be less confusing to just get >> rid of recovery.conf and use a *different* file for this. Just to make >> it clear it's not a config file, but just a boolean exists/notexists >> state. > > +1. If it's not a configuration file anymore, it shouldn't be called > one. +1 to rename file +1 to overall concept, just thinking same myself, not looked at patch yet -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On 9/9/11 7:05 AM, Tom Lane wrote:
> Magnus Hagander <magnus@hagander.net> writes:
>> I have to wonder though, if it wouldn't be less confusing to just get
>> rid of recovery.conf and use a *different* file for this. Just to make
>> it clear it's not a config file, but just a boolean exists/notexists
>> state.
>
> +1. If it's not a configuration file anymore, it shouldn't be called
> one.
I'm in favor of this. People are sufficiently confused by the existing
behavior that we're not going to confuse them further by changing it.
In fact: currently we have a "trigger file" concept in recovery.conf.
Might we unite the trigger file concept with the concept of having a
file to indicate that server is a replica on startup?
i.e. in postgresql.conf we have two settings:
replication_status = {master,standby}
failover_trigger_file = '' #defaults to $PGDATA/master_trigger
If replication_status = master, then the server is a standalone or
master node.
If replication_status = standby, then the server is a PITR, warm
standby, or hot standby replica.
If the failover_trigger_file is created and detected, then PostgreSQL
will reload in master mode. replication_status will get set to "master"
so that it's easy to tell this from the command line.
The above would be a lot easier to comprehend than the current behavior,
and would allow us all of the same options the current behavoir does.
--
Josh Berkus
PostgreSQL Experts Inc.
http://pgexperts.com
On Fri, Sep 9, 2011 at 6:40 PM, Josh Berkus <josh@agliodbs.com> wrote: > I'm in favor of this. People are sufficiently confused by the existing > behavior that we're not going to confuse them further by changing it. > Fwiw as someone who *was* confused previously, it now makes perfect sense to me. "We have postgres.conf which always applies and then recovery.conf which can have all the same options but they only apply during recover". That's much clearer than "we have two configuration files with two disjoint sets of options and good luck remembering which options belong in which file". And it still serves a useful purpose if you have options like recovery_target that you only want to apply during recovery and then plan to remove. -- greg
On Sat, Sep 10, 2011 at 01:07, Greg Stark <stark@mit.edu> wrote: > On Fri, Sep 9, 2011 at 6:40 PM, Josh Berkus <josh@agliodbs.com> wrote: >> I'm in favor of this. People are sufficiently confused by the existing >> behavior that we're not going to confuse them further by changing it. >> > > Fwiw as someone who *was* confused previously, it now makes perfect > sense to me. "We have postgres.conf which always applies and then > recovery.conf which can have all the same options but they only apply > during recover". That's much clearer than "we have two configuration But it only means that for *some* options. Which certainly doesn't reduce confusion. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/
On Fri, Sep 9, 2011 at 8:27 PM, Magnus Hagander <magnus@hagander.net> wrote: >>>> If the same parameter is specified in both file, the setting in recovery.conf >>>> overrides that in postgresql.conf. In this case, SHOW command displays >>>> the settings in postgresql.conf even though they are not used at all. Even if >>> >>> Now, *that* is just broken, IMHO. The SHOW command should show what is >>> actually in use, period. Everything is guaranteed to confuse the hell >>> out of anybody trying to use it. >> >> I'm afraid that we have to add very complicated code to fix that problem. >> Though of course we can easily fix the problem if we don't care about >> the backward compatibility. > > That is an even bigger reason to drop backwards compatibility. Unless > someone else can come up with a neat way of fixing it. Agreed. After a little thought, it's came to mind that we might be able to simply fix that problem by making the postmaster instead of the startup process read recovery.conf; The postmaster reads recovery.conf before it invokes the startup process, and sets the global variables of GUC to the specified values. The child processes including the startup process inherit the global variables of GUC, so SHOW can return the exact setting value and the startup process can use them to do a recovery. Even in EXEC_BACKEND case, we can adopt this idea because those global variables are inherited from parent to children via the save/restore_backend_variables functions. Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
On Sat, Sep 10, 2011 at 12:18 AM, Simon Riggs <simon@2ndquadrant.com> wrote: > On Fri, Sep 9, 2011 at 3:05 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: >> Magnus Hagander <magnus@hagander.net> writes: >>> I have to wonder though, if it wouldn't be less confusing to just get >>> rid of recovery.conf and use a *different* file for this. Just to make >>> it clear it's not a config file, but just a boolean exists/notexists >>> state. >> >> +1. If it's not a configuration file anymore, it shouldn't be called >> one. > > +1 to rename file > > +1 to overall concept, just thinking same myself, not looked at patch yet Are you still thinking the backward-compatibility (i.e., the capability to specify recovery parameters in recovery.conf) is required? Even if we maintain the backward-compatibility, if we rename the file, ISTM that the tools which depends on recovery.conf would need to be changed so that <file-with-new-name> is used instead of recovery.conf. So I'm not sure whether leaving the capability to set parameters in recovery.conf as it is is really worth supporting or not. If we would like to make our effort for the backward-compatibility more worthwhile, we might have to make the path of the file configurable as a user can set it to "recovery.conf". Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
On tis, 2011-09-13 at 14:46 +0900, Fujii Masao wrote: > Are you still thinking the backward-compatibility (i.e., the > capability to specify recovery parameters in recovery.conf) is > required? I think parameters related to a particular recovery, e.g., recovery_target_time, fit better into a recovery.conf that is renamed after the recovery is complete. That was the original idea, after all. Everything that is a permanent setting across multiple recovery attempts, and anything related to replication, better fits elsewhere.
On Tue, Sep 13, 2011 at 2:51 PM, Peter Eisentraut <peter_e@gmx.net> wrote: > On tis, 2011-09-13 at 14:46 +0900, Fujii Masao wrote: >> Are you still thinking the backward-compatibility (i.e., the >> capability to specify recovery parameters in recovery.conf) is >> required? > > I think parameters related to a particular recovery, e.g., > recovery_target_time, fit better into a recovery.conf that is renamed > after the recovery is complete. That was the original idea, after all. > > Everything that is a permanent setting across multiple recovery > attempts, and anything related to replication, better fits elsewhere. I've just been thinking that a better way would be to make recovery.conf an extension of postgresql.conf when we are in archive recovery. So treat postgresql.conf as if it has an automatic "include recovery.conf" in it. The file format is the same. That way we don't need to change existing behaviour, so any software that relies upon this will still work, but we gain the additional ability to reload values in recovery,conf (where appropriate). We can change the .sample files to show parameters that make more sense in one or the other file, rather than making it a hard requirement for them to appear in specific files which will be a real pain in the ass. Internal changes would then be to move existing recovery.conf parameters into guc.c and revise the manual accordingly. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Wed, Sep 14, 2011 at 1:10 AM, Simon Riggs <simon@2ndquadrant.com> wrote: > On Tue, Sep 13, 2011 at 2:51 PM, Peter Eisentraut <peter_e@gmx.net> wrote: >> On tis, 2011-09-13 at 14:46 +0900, Fujii Masao wrote: >>> Are you still thinking the backward-compatibility (i.e., the >>> capability to specify recovery parameters in recovery.conf) is >>> required? >> >> I think parameters related to a particular recovery, e.g., >> recovery_target_time, fit better into a recovery.conf that is renamed >> after the recovery is complete. That was the original idea, after all. >> >> Everything that is a permanent setting across multiple recovery >> attempts, and anything related to replication, better fits elsewhere. > > I've just been thinking that a better way would be to make > recovery.conf an extension of postgresql.conf when we are in archive > recovery. > > So treat postgresql.conf as if it has an automatic "include > recovery.conf" in it. The file format is the same. > > That way we don't need to change existing behaviour, so any software > that relies upon this will still work, but we gain the additional > ability to reload values in recovery,conf (where appropriate). > > We can change the .sample files to show parameters that make more > sense in one or the other file, rather than making it a hard > requirement for them to appear in specific files which will be a real > pain in the ass. > > Internal changes would then be to move existing recovery.conf > parameters into guc.c and revise the manual accordingly. Sounds reasonable. I'll revise the patch that way. Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
On tis, 2011-09-13 at 17:10 +0100, Simon Riggs wrote: > So treat postgresql.conf as if it has an automatic "include > recovery.conf" in it. The file format is the same. Sounds good. That would also have the merit that you could use, say, different memory settings during recovery.
On Wed, Sep 14, 2011 at 6:33 PM, Peter Eisentraut <peter_e@gmx.net> wrote: > On tis, 2011-09-13 at 17:10 +0100, Simon Riggs wrote: >> So treat postgresql.conf as if it has an automatic "include >> recovery.conf" in it. The file format is the same. > > Sounds good. That would also have the merit that you could use, say, > different memory settings during recovery. One problem of this is that recovery.conf is relative to the configuration file directory instead of data directory if we treat it as an "include" file. If your configuration file directory is the same as the data directory, you don't need to worry about this problem. But if you set data_directory to the directory other than ConfigDir, the problem might break your tool which depends on recovery.conf under the data directory. If we'd like to treat recovery.conf as if it's under the data directory, I'm afraid that we should add complicated code to parse recovery.conf after the value of data_directory has been determined from postgresql.conf. Furthermore, what if recovery.conf has another setting of data_directory? Since recovery.conf is a configuration file, it's intuitive for me to put it in configuration file directory rather than data directory. So I'm not inclined to treat recovery.conf as if it's under data directory. Is this OK? Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
Fujii Masao <masao.fujii@gmail.com> writes: > If we'd like to treat recovery.conf as if it's under the data directory, I'm > afraid that we should add complicated code to parse recovery.conf after > the value of data_directory has been determined from postgresql.conf. > Furthermore, what if recovery.conf has another setting of data_directory? We already do this dance for custom_variable_classes. IIRC the only thing you have to care about is setting the GUC before any other one. I guess you could just process the data_directory the same way, before the main loop, and be done with it. > Since recovery.conf is a configuration file, it's intuitive for me to put it > in configuration file directory rather than data directory. So I'm not inclined > to treat recovery.conf as if it's under data directory. Is this OK? I think that if we keep some compatibility with recovery.conf, then we need to include finding it in the data_directory or the configuration file directory, in that order, and with a LOG message that says where we did find it. I think we should *NOT* load both of them with some priority rules. Regards, -- Dimitri Fontaine http://2ndQuadrant.fr PostgreSQL : Expertise, Formation et Support
On tor, 2011-09-15 at 16:54 +0900, Fujii Masao wrote: > On Wed, Sep 14, 2011 at 6:33 PM, Peter Eisentraut <peter_e@gmx.net> wrote: > > On tis, 2011-09-13 at 17:10 +0100, Simon Riggs wrote: > >> So treat postgresql.conf as if it has an automatic "include > >> recovery.conf" in it. The file format is the same. > > > > Sounds good. That would also have the merit that you could use, say, > > different memory settings during recovery. > > One problem of this is that recovery.conf is relative to the configuration > file directory instead of data directory if we treat it as an "include" file. Treat it *as if* it were an included file. A special included file if you will. > If we'd like to treat recovery.conf as if it's under the data directory, I'm > afraid that we should add complicated code to parse recovery.conf after > the value of data_directory has been determined from postgresql.conf. > Furthermore, what if recovery.conf has another setting of data_directory? Perhaps that could be addresses by inventing a new context setting PGC_RECOVERY so that you can only set sane values. > Since recovery.conf is a configuration file, it's intuitive for me to put it > in configuration file directory rather than data directory. So I'm not inclined > to treat recovery.conf as if it's under data directory. Is this OK? It's not a configuration file, because it disappears after use. (And a lot of configuration file management systems would be really upset if we had a configuration file that behaved that way.) The whole point of this exercise is allowing the permanent configuration file parameters to be moved to a real configuration file (postgresql.conf), while leaving the temporary settings separately. Alternatively, we could just forget about the whole thing and move everything to postgresql.conf and treat recovery.conf as a simple empty signal file. I don't know if that's necessarily better.
On Thu, Sep 15, 2011 at 5:58 AM, Peter Eisentraut <peter_e@gmx.net> wrote: > Alternatively, we could just forget about the whole thing and move > everything to postgresql.conf and treat recovery.conf as a simple empty > signal file. I don't know if that's necessarily better. Seems like it might be simpler. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Thu, Sep 15, 2011 at 8:04 PM, Robert Haas <robertmhaas@gmail.com> wrote: > On Thu, Sep 15, 2011 at 5:58 AM, Peter Eisentraut <peter_e@gmx.net> wrote: >> Alternatively, we could just forget about the whole thing and move >> everything to postgresql.conf and treat recovery.conf as a simple empty >> signal file. I don't know if that's necessarily better. > > Seems like it might be simpler. It seems to need a bit more time until we've reached a consensus about the treatment of recovery.conf. How about committing the core patch first, and addressing the recovery.conf issue as a different patch later? The attached patch provides a core part of the feature, i.e., it moves every recovery parameters from recovery.conf to postgresql.conf. Even if you create recovery.conf, the server doesn't read it automatically. The patch renames recovery.conf to recovery.ready, so if you want to enter archive recovery or standby mode, you need to create recovery.ready file in the cluster data directory. Since recovery.ready is just a signal file, its contents have no effect. Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
Attachment
Fujii Masao <masao.fujii@gmail.com> writes:
> It seems to need a bit more time until we've reached a consensus about
> the treatment of recovery.conf. How about committing the core patch
> first, and addressing the recovery.conf issue as a different patch later?
> The attached patch provides a core part of the feature, i.e., it moves
> every recovery parameters from recovery.conf to postgresql.conf.
> Even if you create recovery.conf, the server doesn't read it automatically.
> The patch renames recovery.conf to recovery.ready, so if you want to
> enter archive recovery or standby mode, you need to create
> recovery.ready file in the cluster data directory. Since recovery.ready is
> just a signal file, its contents have no effect.
This seems like it's already predetermining the outcome of the argument
about recovery.conf. Mind you, I'm not unhappy with this choice, but
it's hardly implementing only behavior that's not being debated.
If we're satisfied with not treating recovery-only parameters different
from run-of-the-mill GUCs, this is fine.
regards, tom lane
On Thu, Sep 15, 2011 at 11:37 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > This seems like it's already predetermining the outcome of the argument > about recovery.conf. Mind you, I'm not unhappy with this choice, but > it's hardly implementing only behavior that's not being debated. > > If we're satisfied with not treating recovery-only parameters different > from run-of-the-mill GUCs, this is fine. Okay, we need to reach a consensus about the treatment of recovery.conf. We have three choices. Which do you like the best? #1 Use empty recovery.ready file to enter arhicve recovery. recovery.conf is not read automatically. All recovery parameters are expected to be specified in postgresql.conf. If you must specify them in recovery.conf, you need to add "include 'recovery.conf'" into postgresql.conf. But note that that recovery.conf will not be renamed to recovery.done at the end of recovery. This is what the patch I've posted does. This is simplest approach, but might confuse people who use the tools which depend on recovery.conf. #2 Use empty recovery.ready file to enter archive recovery. recovery.conf is read automatically. You can specify recovery parameters in recovery.conf without adding "include 'recovery.conf'" into postgresql.conf. But note that that recovery.conf will not be renamed to recovery.done at the end of recovery. If we adopt this, we might need to implement what Dimitri suggested. I guess this is not so difficult thing. http://archives.postgresql.org/pgsql-hackers/2011-09/msg00745.php #3 Use recovery.conf file to enter archive recovery. recovery.conf is read automatically, and will be renamed to recovery.done at the end of recovery. This is least confusing approach for the existing users, but I guess we need to add lots of code (e.g., as Peter suggested, we might need to invent new context setting like PGC_RECOVERY) to address the problem I pointed before. http://archives.postgresql.org/pgsql-hackers/2011-09/msg00482.php If we want to use recovery.conf as a temporary configuration file for recovery (i.e., configuration file disappears after use), we must choose #3. If we can live with that recovery.conf is treated as a permanent configuration file but don't want to add "include 'recovery.conf'" into postgresql.conf, #2 is best. If we don't use recovery.conf or mind editing postgresql.conf to use recovery.conf, #1 is best. Which behavior are you expecting? Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
Fujii Masao <masao.fujii@gmail.com> writes:
> We have three choices. Which do you like the best?
I'm in favor of defining a separate, content-free trigger file to enable
archive recovery. Not sure about the name "recovery.ready", though ---
that makes it look like one of the WAL archive transfer trigger files,
which does not seem like a great analogy. The pg_standby documentation
suggests names like "foo.trigger" for failover triggers, which is a bit
better analogy because something external to the database creates the
file. What about "recovery.trigger"?
As far as the other issues go, I think there is actually a prerequisite
discussion to be had here, which is whether we are turning the recovery
parameters into plain old GUCs or not. If they are plain old GUCs, then
they will presumably still have their values when we are *not* doing
recovery. That leads to a couple of questions:
* will seeing these values present in pg_settings confuse anybody?
* can the values be changed when not in recovery, if so what happens, and again will that confuse anybody?
* is there any security hazard from ordinary users being able to see what settings had been used?
If these settings are to be plain old GUCs, then I think we should just
stick them in postgresql.conf and forget about recovery.conf (although
of course someone could "include recovery.conf" if he were bent on
storing them separately). On the other hand, if we think they are *not*
plain old GUCs, maybe they shouldn't be in postgresql.conf. But the
source file isn't the first thing to worry about in that case.
regards, tom lane
On fre, 2011-09-16 at 01:32 -0400, Tom Lane wrote: > As far as the other issues go, I think there is actually a > prerequisite > discussion to be had here, which is whether we are turning the > recovery > parameters into plain old GUCs or not. If they are plain old GUCs, > then > they will presumably still have their values when we are *not* doing > recovery. That leads to a couple of questions: > * will seeing these values present in pg_settings confuse anybody? How so? We add or change the available parameters all the time. > * can the values be changed when not in recovery, if so what happens, > and again will that confuse anybody? Should be similar to archive_command and archive_mode. You can still see and change archive_command when archive_mode is off. > * is there any security hazard from ordinary users being able to see > what settings had been used? Again, not much different from the archive_* settings. They are, after all, almost the same in the opposite direction.
On fre, 2011-09-16 at 11:54 +0900, Fujii Masao wrote:
> #1
> Use empty recovery.ready file to enter arhicve recovery. recovery.conf
> is not read automatically. All recovery parameters are expected to be
> specified in postgresql.conf. If you must specify them in recovery.conf,
> you need to add "include 'recovery.conf'" into postgresql.conf. But note
> that that recovery.conf will not be renamed to recovery.done at the
> end of recovery. This is what the patch I've posted does. This is
> simplest approach, but might confuse people who use the tools which
> depend on recovery.conf.
A small variant to this: When you are actually doing recovery from a
backup, having a recovery trigger and a recovery done file is obviously
quite helpful and necessary for safety. But when you're setting up a
replication slave, it adds extra complexity for the user. The
approximately goal ought to be to be able to do
pg_basebackup -h master -D there
postgres -D there --standby-mode=on --primary-conninfo=master
without the need to touch any obscure recovery trigger files.
So perhaps recovery.{trigger,ready} should not be needed if, say,
standby_mode=on. But then what impact should the presence of
recovery.done have?
On 15-09-2011 23:54, Fujii Masao wrote: > #1 > Use empty recovery.ready file to enter arhicve recovery. recovery.conf > is not read automatically. All recovery parameters are expected to be > specified in postgresql.conf. If you must specify them in recovery.conf, > you need to add "include 'recovery.conf'" into postgresql.conf. But note > that that recovery.conf will not be renamed to recovery.done at the > end of recovery. This is what the patch I've posted does. This is > simplest approach, but might confuse people who use the tools which > depend on recovery.conf. > more or less +1. We don't need two config files.; just one: postgresql.conf. Just turn all recovery.conf parameters to GUCs. As already said, the recovery.conf settings are not different from archive settings, we just need a way to trigger the recovery. And that trigger could be pulled by a GUC (standby_mode) or a file (say recovery -> recovery.done). Also, recovery.done could be filled with recovery information just for DBA record. standby_mode does not create any file, it just trigger the recovery (as it will be used mainly for replication purposes). -- Euler Taveira de Oliveira - Timbira http://www.timbira.com.br/ PostgreSQL: Consultoria, Desenvolvimento, Suporte24x7 e Treinamento
> I'm in favor of defining a separate, content-free trigger file to > enable > archive recovery. Not sure about the name "recovery.ready", though > --- > that makes it look like one of the WAL archive transfer trigger > files, > which does not seem like a great analogy. The pg_standby > documentation > suggests names like "foo.trigger" for failover triggers, which is a > bit > better analogy because something external to the database creates the > file. What about "recovery.trigger"? Do we want a trigger file to enable recovery, or one to *disable* recovery? Or both? Also, I might point out that we're really confusing our users by talking about "recovery" all the time, if they're just usingstreaming replication. Just sayin' > * will seeing these values present in pg_settings confuse anybody? No. pg_settings already has a couple dozen "developer" parameters which nobody not on this mailing list understands. Addingthe recovery parameters to it wouldn't confuse anyone further, and would have the advantage of making the recoveryparameters available by monitoring query on a hot standby. For that matter, I'd suggest that we add a read-only setting called in_recovery. > * can the values be changed when not in recovery, if so what happens, > and again will that confuse anybody? Yes, and no. > * is there any security hazard from ordinary users being able to see > what settings had been used? primary_conninfo could be a problem, since it's possible to set a password there. --Josh
On Fri, Sep 16, 2011 at 3:22 PM, Joshua Berkus <josh@agliodbs.com> wrote: > No. pg_settings already has a couple dozen "developer" parameters which nobody not on this mailing list understands. Addingthe recovery parameters to it wouldn't confuse anyone further, and would have the advantage of making the recoveryparameters available by monitoring query on a hot standby. +1. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Sat, Sep 17, 2011 at 4:22 AM, Joshua Berkus <josh@agliodbs.com> wrote: >> that makes it look like one of the WAL archive transfer trigger >> files, >> which does not seem like a great analogy. The pg_standby >> documentation >> suggests names like "foo.trigger" for failover triggers, which is a >> bit >> better analogy because something external to the database creates the >> file. What about "recovery.trigger"? I'm OK with that name. > Do we want a trigger file to enable recovery, or one to *disable* recovery? Or both? ISTM that only supporting a trigger file to enable recovery is less confusing. >> * will seeing these values present in pg_settings confuse anybody? > > No. pg_settings already has a couple dozen "developer" parameters which nobody not on this mailing list understands. Addingthe recovery parameters to it wouldn't confuse anyone further, and would have the advantage of making the recoveryparameters available by monitoring query on a hot standby. +1 >> * is there any security hazard from ordinary users being able to see >> what settings had been used? > > primary_conninfo could be a problem, since it's possible to set a password there. True. I agree that primary_conninfo should be restricted to superuser. Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
On Fri, Sep 16, 2011 at 9:25 PM, Peter Eisentraut <peter_e@gmx.net> wrote:
> On fre, 2011-09-16 at 11:54 +0900, Fujii Masao wrote:
>> #1
>> Use empty recovery.ready file to enter arhicve recovery. recovery.conf
>> is not read automatically. All recovery parameters are expected to be
>> specified in postgresql.conf. If you must specify them in recovery.conf,
>> you need to add "include 'recovery.conf'" into postgresql.conf. But note
>> that that recovery.conf will not be renamed to recovery.done at the
>> end of recovery. This is what the patch I've posted does. This is
>> simplest approach, but might confuse people who use the tools which
>> depend on recovery.conf.
>
> A small variant to this: When you are actually doing recovery from a
> backup, having a recovery trigger and a recovery done file is obviously
> quite helpful and necessary for safety. But when you're setting up a
> replication slave, it adds extra complexity for the user. The
> approximately goal ought to be to be able to do
>
> pg_basebackup -h master -D there
> postgres -D there --standby-mode=on --primary-conninfo=master
>
> without the need to touch any obscure recovery trigger files.
>
> So perhaps recovery.{trigger,ready} should not be needed if, say,
> standby_mode=on.
Or what about making pg_basebackup automatically create a
recovery trigger file?
Regards,
--
Fujii Masao
NIPPON TELEGRAPH AND TELEPHONE CORPORATION
NTT Open Source Software Center
On Fri, Sep 16, 2011 at 2:46 PM, Euler Taveira de Oliveira <euler@timbira.com> wrote: > On 15-09-2011 23:54, Fujii Masao wrote: >> >> #1 >> Use empty recovery.ready file to enter arhicve recovery. recovery.conf >> is not read automatically. All recovery parameters are expected to be >> specified in postgresql.conf. If you must specify them in recovery.conf, >> you need to add "include 'recovery.conf'" into postgresql.conf. But note >> that that recovery.conf will not be renamed to recovery.done at the >> end of recovery. This is what the patch I've posted does. This is >> simplest approach, but might confuse people who use the tools which >> depend on recovery.conf. >> > more or less +1. We don't need two config files.; just one: postgresql.conf. > Just turn all recovery.conf parameters to GUCs. As already said, the > recovery.conf settings are not different from archive settings, we just need > a way to trigger the recovery. And that trigger could be pulled by a GUC > (standby_mode) or a file (say recovery -> recovery.done). Also, > recovery.done could be filled with recovery information just for DBA record. > standby_mode does not create any file, it just trigger the recovery (as it > will be used mainly for replication purposes). I sympathise with this view, to an extent. If people want to put all parameters in one file, they can do so. So +1 to that. Should they be forced to adopt that new capability by us deliberately breaking their existing setups? No. So -1 to that. If we do an automatic include of recovery.conf first, then follow by reading postgresql,conf then we will preserve the old as well as allowing the new. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Fri, Sep 16, 2011 at 3:54 AM, Fujii Masao <masao.fujii@gmail.com> wrote: > On Thu, Sep 15, 2011 at 11:37 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: >> This seems like it's already predetermining the outcome of the argument >> about recovery.conf. Mind you, I'm not unhappy with this choice, but >> it's hardly implementing only behavior that's not being debated. >> >> If we're satisfied with not treating recovery-only parameters different >> from run-of-the-mill GUCs, this is fine. > > Okay, we need to reach a consensus about the treatment of > recovery.conf. > > We have three choices. What we should focus on is these requirements 1. Don't break software that relies on the existing behaviour 2. Allow parameters to be reloaded at SIGHUP 3. Allow recovery parameters to be handled same way as other GUCs 4. We need to retain recovery.conf/.done style behaviour to mark end of archive recovery Making an automatic include makes (2) and (3) work OK, without breaking (1). I haven't seen another solution that doesnt break (1). Rename of .conf to .done allows us to maintain working code for existing solutions (there are *many* out there...) We should say that the automatic include only works during recovery, so if someone puts recovery.conf back then we ignore it. If we treat recovery,conf as being read *first* then any parameter mentioned twice (i.e. mentioned again in postgresql.conf) will override the setting in recovery.conf and we have no issues. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Fri, Sep 16, 2011 at 1:13 PM, Peter Eisentraut <peter_e@gmx.net> wrote: > On fre, 2011-09-16 at 01:32 -0400, Tom Lane wrote: >> As far as the other issues go, I think there is actually a >> prerequisite >> discussion to be had here, which is whether we are turning the >> recovery >> parameters into plain old GUCs or not. If they are plain old GUCs, >> then >> they will presumably still have their values when we are *not* doing >> recovery. That leads to a couple of questions: >> * will seeing these values present in pg_settings confuse anybody? > > How so? We add or change the available parameters all the time. > >> * can the values be changed when not in recovery, if so what happens, >> and again will that confuse anybody? > > Should be similar to archive_command and archive_mode. You can still > see and change archive_command when archive_mode is off. I do think special handling would be useful here, of some kind. We could reset them as well, if we wished, but that is probably more trouble than is worth. Perhaps we need a new SCOPE attribute on pg_settings to show whether the parameter applies in recovery, in normal or both. >> * is there any security hazard from ordinary users being able to see >> what settings had been used? > > Again, not much different from the archive_* settings. They are, after > all, almost the same in the opposite direction. There is a potential security hole if people hardcode passwords into primary_conninfo. As long as we document not to do that, we're OK. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Simon Riggs <simon@2ndQuadrant.com> writes:
> I sympathise with this view, to an extent.
> If people want to put all parameters in one file, they can do so. So +1 to that.
> Should they be forced to adopt that new capability by us deliberately
> breaking their existing setups? No. So -1 to that.
> If we do an automatic include of recovery.conf first, then follow by
> reading postgresql,conf then we will preserve the old as well as
> allowing the new.
I don't buy this argument at all. I don't believe that recovery.conf is
part of anyone's automated processes at all, let alone to an extent that
they won't be able to cope with a change to rationalize the file layout.
And most especially I don't buy that someone who does want to keep using
it couldn't cope with adding an "include" to postgresql.conf manually.
If we're going to move these parameters into postgresql.conf, we should
just do that and remove all mention of recovery.conf. Anything else
will generate much more confusion than benefit.
regards, tom lane
> I don't buy this argument at all. I don't believe that recovery.conf is > part of anyone's automated processes at all, let alone to an extent that > they won't be able to cope with a change to rationalize the file layout. > And most especially I don't buy that someone who does want to keep using > it couldn't cope with adding an "include" to postgresql.conf manually. Speaking as someone who has a lot of client admin scripts written around recovery.conf, we will be *thrilled* to update those scripts in order to simplify the configuration file situation. Having a third conf file (or fourth, or fifth, depending on which auth features you're using) already adds unwarranted complexity to automation scripts, and each config file we get rid of makes those scripts easier to maintain and troubleshoot. For that matter, I'd love to consolidate pg_hba and pg_ident into a single file. Any chance? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
All, First, if we're going to change behavior, I assert that we should stop calling stuff "recovery" and either call it "replica" or "standby". Our use of the word "recovery" confuses users; it is historical in nature and requires an understanding of PostgreSQL internals to know why it's called that. It's also inconsistent with our use of the word "standby" everywhere else. Second, I haven't seen a response to this: > Do we want a trigger file to enable recovery, or one to *disable* recovery? Or both? I'll go further and say that we only want one trigger file by default, one which either enables or disables recovery. I'll further suggest that we: a) have a standby.on file which puts the server in replica/recovery mode if it's detected on startup, and b) that we poll for the standby.on file going away as a trigger to stop recovery and bring up the server in master mode, and c) that pg_basebackup automatically create a standby.on file. > Perhaps we need a new SCOPE attribute on pg_settings to show whether > the parameter applies in recovery, in normal or both. An overhaul of the category tree would also do that. I've been putting off an overhaul/cleanup of categories for pg_settings, maybe it's about time. > There is a potential security hole if people hardcode passwords into > primary_conninfo. As long as we document not to do that, we're OK. Yeah, I'd almost be inclined to actively prohibit this, but that would draw user complaints. We'll have to be satisfied with a doc plus a comment. Speaking of which, .pgpass could be better documented as an option for handling this sort of thing. Will take a stab, eventually. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Tue, Sep 20, 2011 at 1:01 PM, Josh Berkus <josh@agliodbs.com> wrote: > I'll go further and say that we only want one trigger file by default, > one which either enables or disables recovery. I'll further suggest > that we: > > a) have a standby.on file which puts the server in replica/recovery mode > if it's detected on startup, and > b) that we poll for the standby.on file going away as a trigger to stop > recovery and bring up the server in master mode, and > c) that pg_basebackup automatically create a standby.on file. It seems a bit confusing to me to have a file that takes effect only at startup when created but anytime when removed. I think one of the insufficiently-lauded 9.1 features is Fujii Masao's "pg_ctl promote". Now THAT is a good interface. Unlike trigger_file, it doesn't require any advance preparation, or monkeying with files on disk. You just tell it to promote, and it does. Sweet.Now it turns out that it uses a file to make that happenbehind the scenes, but who cares? From the user's perspective It Just Works. I like the idea of some kind of sentinel file that tells the server to start up in recovery mode. But instead of having the user remove it to cause a promotion, I think the server should remove it when it does promote. That's more like what we've done in the past, and it ties in very nicely with what pg_ctl promote already does. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On 9/20/11 10:09 AM, Robert Haas wrote: > I like the idea of some kind of sentinel file that tells the server to > start up in recovery mode. But instead of having the user remove it > to cause a promotion, I think the server should remove it when it does > promote. That's more like what we've done in the past, and it ties in > very nicely with what pg_ctl promote already does. Yes, I agree that that is a superior approach. And then we could keep the trigger_file configuration parameter for backwards compatibility, with intent to depreciate it after a couple more Postgres versions. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
Josh Berkus <josh@agliodbs.com> writes:
> First, if we're going to change behavior, I assert that we should stop
> calling stuff "recovery" and either call it "replica" or "standby". Our
> use of the word "recovery" confuses users; it is historical in nature
> and requires an understanding of PostgreSQL internals to know why it's
> called that. It's also inconsistent with our use of the word "standby"
> everywhere else.
Are we all talking about the same thing? In my mind recovery.conf is
for configuring a point-in-time archive recovery run. It's got nothing
to do with either replication or standbys. Perhaps part of our problem
here is overloading that case with standby behavior.
> Second, I haven't seen a response to this:
> Do we want a trigger file to enable recovery, or one to *disable*
> recovery? Or both?
As far as the PITR scenario is concerned, only the former can possibly
make any sense; the latter would be downright dangerous.
>> There is a potential security hole if people hardcode passwords into
>> primary_conninfo. As long as we document not to do that, we're OK.
> Yeah, I'd almost be inclined to actively prohibit this, but that would
> draw user complaints. We'll have to be satisfied with a doc plus a comment.
I think that marking the GUC as "only readable by superuser" is a
sufficient fix.
regards, tom lane
On Tue, Sep 20, 2011 at 1:30 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > Josh Berkus <josh@agliodbs.com> writes: >> First, if we're going to change behavior, I assert that we should stop >> calling stuff "recovery" and either call it "replica" or "standby". Our >> use of the word "recovery" confuses users; it is historical in nature >> and requires an understanding of PostgreSQL internals to know why it's >> called that. It's also inconsistent with our use of the word "standby" >> everywhere else. > > Are we all talking about the same thing? In my mind recovery.conf is > for configuring a point-in-time archive recovery run. It's got nothing > to do with either replication or standbys. Huh? How else can you create a standby? I do it by creating a recovery.conf file that says: standby_mode=on primary_conninfo='whatever' I wasn't aware that there is another method. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Robert Haas <robertmhaas@gmail.com> writes:
> On Tue, Sep 20, 2011 at 1:30 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Are we all talking about the same thing? �In my mind recovery.conf is
>> for configuring a point-in-time archive recovery run. �It's got nothing
>> to do with either replication or standbys.
> Huh? How else can you create a standby? I do it by creating a
> recovery.conf file that says:
> standby_mode=on
The point I'm trying to make is that it seems like this discussion is
getting driven entirely by the standby case, without remembering that
recovery.conf was originally designed for, and is still used in,
a significantly different use-case. Maybe we had better take two
steps back and think about the implications for the archive-recovery
case.
regards, tom lane
> The point I'm trying to make is that it seems like this discussion is > getting driven entirely by the standby case, without remembering that > recovery.conf was originally designed for, and is still used in, > a significantly different use-case. Maybe we had better take two > steps back and think about the implications for the archive-recovery > case. I think we should take that into consideration, sure. But it should not be in the driver's seat for things like nomenclature. Far more people use replication than use PITR. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Tue, Sep 20, 2011 at 3:10 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > Robert Haas <robertmhaas@gmail.com> writes: >> On Tue, Sep 20, 2011 at 1:30 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: >>> Are we all talking about the same thing? In my mind recovery.conf is >>> for configuring a point-in-time archive recovery run. It's got nothing >>> to do with either replication or standbys. > >> Huh? How else can you create a standby? I do it by creating a >> recovery.conf file that says: >> standby_mode=on > > The point I'm trying to make is that it seems like this discussion is > getting driven entirely by the standby case, without remembering that > recovery.conf was originally designed for, and is still used in, > a significantly different use-case. Maybe we had better take two > steps back and think about the implications for the archive-recovery > case. I'm not sure there really are any implications that are worth getting excited about. The problem is really one of naming; I'm reminded of our recent discussions of the use of the term "relation". The problem is that the word "recovery" encompasses a number of very different scenarios. First, we have crash recovery. Second, we have archive recovery (which, confusingly enough, no longer requires an archive). Archive recovery can be further subdivided by where the xlog files are coming from (archive, streaming replication, both, or neither) and how long we plan to stay in recovery (forever, if acting as a standby; until end of WAL, if promoting a standby or recovering a hot backup; or until the recovery target is reached, if performing a point in time recovery). I would guess that most of those are not what the typical user thinks of as "recovery", but what else are we gonna call it? Josh is arguing that we ought to use the term "replication", but it seems to me that's just as misleading - maybe moreso, since "recovery" is sufficiently a term of art to make you at least think about reading the manual, whereas you know (or think you know) what replication is. For now, I think we're best off not changing the terminology, and confining the remit of this patch to (a) turning all of the existing recovery.conf parameters into GUCs and (b) replacing recovery.conf with a sentinel file a sentinel file (name TBB) to indicate that the server is to start in recovery mode. The naming isn't great but the more we change at once the less chance of reaching agreement. It seems like we have pretty broad agreement on the basics here, so let's start with that. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Robert, > Josh is arguing that we ought to use the term "replication", but it Actually, no. I'm arguing that we should use the term "standby", since that term is consistent with how we refer to replica servers throughout the docs, and the term "recovery" is not. > seems to me that's just as misleading - maybe moreso, since "recovery" > is sufficiently a term of art to make you at least think about reading > the manual, whereas you know (or think you know) what replication is. Nope. What it means is that users see stuff relating to "recovery" and say "oh, that's not right, the replication stuff must be somewhere else". I've taught a half-dozen classes on PostgreSQL binary replication now, and the "recovery" nomenclature *always* confuses students. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Wed, Sep 21, 2011 at 12:55 PM, Josh Berkus <josh@agliodbs.com> wrote: >> Josh is arguing that we ought to use the term "replication", but it > > Actually, no. I'm arguing that we should use the term "standby", since > that term is consistent with how we refer to replica servers throughout > the docs, and the term "recovery" is not. > >> seems to me that's just as misleading - maybe moreso, since "recovery" >> is sufficiently a term of art to make you at least think about reading >> the manual, whereas you know (or think you know) what replication is. > > Nope. What it means is that users see stuff relating to "recovery" and > say "oh, that's not right, the replication stuff must be somewhere else". > > I've taught a half-dozen classes on PostgreSQL binary replication now, > and the "recovery" nomenclature *always* confuses students. Yeah, I get it. But I think standby would confuse them, too, just in a different set of situations. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
> Yeah, I get it. But I think standby would confuse them, too, just in > a different set of situations. Other than PITR, what situations? PITR is used by a minority of our users. Binary replication, if not already used by a majority, will be in the future (it's certainly the majority of my professional clients). Further, PITR is usually something which is either handled by vendor backup management software, or by professional DBAs, whereas replication is used by developers with little or no DBA support. Why should we make terminology obscure for the majority usecase to make it clear for the minority one? Especially since the majority use-case has almost all the newbies? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Wed, Sep 21, 2011 at 1:03 PM, Josh Berkus <josh@agliodbs.com> wrote: >> Yeah, I get it. But I think standby would confuse them, too, just in >> a different set of situations. > > Other than PITR, what situations? Hot backup? -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On 9/21/11 10:07 AM, Robert Haas wrote: > On Wed, Sep 21, 2011 at 1:03 PM, Josh Berkus <josh@agliodbs.com> wrote: >>> Yeah, I get it. But I think standby would confuse them, too, just in >>> a different set of situations. >> >> Other than PITR, what situations? > > Hot backup? Hot backup == PITR. You're just not bothering to accumulate WAL logs. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Wed, Sep 21, 2011 at 1:08 PM, Josh Berkus <josh@agliodbs.com> wrote: > On 9/21/11 10:07 AM, Robert Haas wrote: >> On Wed, Sep 21, 2011 at 1:03 PM, Josh Berkus <josh@agliodbs.com> wrote: >>>> Yeah, I get it. But I think standby would confuse them, too, just in >>>> a different set of situations. >>> >>> Other than PITR, what situations? >> >> Hot backup? > > Hot backup == PITR. You're just not bothering to accumulate WAL logs. Well, I don't think of it that way, but YMMV, of course. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Wed, Sep 21, 2011 at 1:13 PM, Robert Haas <robertmhaas@gmail.com> wrote: > On Wed, Sep 21, 2011 at 1:08 PM, Josh Berkus <josh@agliodbs.com> wrote: >> On 9/21/11 10:07 AM, Robert Haas wrote: >>> On Wed, Sep 21, 2011 at 1:03 PM, Josh Berkus <josh@agliodbs.com> wrote: >>>>> Yeah, I get it. But I think standby would confuse them, too, just in >>>>> a different set of situations. >>>> >>>> Other than PITR, what situations? >>> >>> Hot backup? >> >> Hot backup == PITR. You're just not bothering to accumulate WAL logs. > > Well, I don't think of it that way, but YMMV, of course. I think that the major differentiating factor is the "intended action when caught up", and the definition of caught up, and trying to use a single term for both of them is going to always cause confusion. So I tend to think of the use cases by their "continuation". A "slave" is intended to "continually keep trying to get more" once it's retrieved and applied all the changes it can. It can be hot, or cold, streaming, or archive, etc... And "recovery" is intended to stop recovering and become "normal" once it's finished retrieving and applying all changes it can. Again, it has multiple ways to retrive it's wal too. And I think Tom touched on this point in the "recovery.conf/recovery.done" thread a bit too. Maybe we need to really start talking about the different "when done do ..." distinctions, and and using that distinction to help our nomenclature. Both recovery/slave (both hot or cold) use the same retrieve/apply machinery (and thus configuration options). But because of the different "caught up action", are different features. a. -- Aidan Van Dyk Create like a god, aidan@highrise.ca command like a king, http://www.highrise.ca/ work like a slave.
On Wed, Sep 21, 2011 at 1:34 PM, Aidan Van Dyk <aidan@highrise.ca> wrote: > And I think Tom touched on this point in the > "recovery.conf/recovery.done" thread a bit too. Doh! That's this thread.... /me slinks away, ashamed for not even taking a close look at the to/cc list... -- Aidan Van Dyk Create like a god, aidan@highrise.ca command like a king, http://www.highrise.ca/ work like a slave.
On tis, 2011-09-20 at 16:38 -0400, Robert Haas wrote: > For now, I think we're best off not changing the terminology, and > confining the remit of this patch to (a) turning all of the existing > recovery.conf parameters into GUCs and (b) replacing recovery.conf > with a sentinel file a sentinel file (name TBB) to indicate that the > server is to start in recovery mode. The naming isn't great but the > more we change at once the less chance of reaching agreement. It > seems like we have pretty broad agreement on the basics here, so let's > start with that. The only thing that's slightly bogus about that is that if you were doing an archive recovery, you'd have to edit the main postgresql.conf with one-shot parameters for that particular recovery (and then delete them again, or leave them in place, confusing the next guy). But perhaps that's worth the overall simplification.
On Tue, Sep 20, 2011 at 5:23 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > Simon Riggs <simon@2ndQuadrant.com> writes: >> I sympathise with this view, to an extent. > >> If people want to put all parameters in one file, they can do so. So +1 to that. > >> Should they be forced to adopt that new capability by us deliberately >> breaking their existing setups? No. So -1 to that. > >> If we do an automatic include of recovery.conf first, then follow by >> reading postgresql,conf then we will preserve the old as well as >> allowing the new. > > I don't buy this argument at all. I don't believe that recovery.conf is > part of anyone's automated processes at all, let alone to an extent that > they won't be able to cope with a change to rationalize the file layout. There are many. Tools I can name include pgpool, 2warm, PITRtools, but there are also various tools from Sun, an IBM reseller I have forgotten the name of, OmniTI and various other backup software providers. Those are just the ones I can recall quickly. We've encouraged people to write software on top and they have done so. Breaking backwards compatibility isn't something we do elsewhere, when its easy to keep it. I don't object to new functionality (and agreed to it upthread), just don't break the old way when there is no need. > And most especially I don't buy that someone who does want to keep using > it couldn't cope with adding an "include" to postgresql.conf manually. This just creates a barrier to upgrade. Most people using PostgreSQL use multiple releases, so its a pain to have to maintain multiple versions of scripts to have things work properly. Even people that don't mind changing won't be able to do it fully. That creates confusion, which leads to mistakes. These things relate to backup and recovery. Any changes to them give risk of data loss. Cosmetic considerations are secondary. There is no command to safely confirm that "include recovery.conf" is added to postgresql.conf, so leaving it optional makes it unclear as to whether the old ways will work or not. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On 09/20/2011 09:23 AM, Tom Lane wrote: > Simon Riggs<simon@2ndQuadrant.com> writes: >> I sympathise with this view, to an extent. > >> If we do an automatic include of recovery.conf first, then follow by >> reading postgresql,conf then we will preserve the old as well as >> allowing the new. > > I don't buy this argument at all. I don't believe that recovery.conf is > part of anyone's automated processes at all, let alone to an extent that > they won't be able to cope with a change to rationalize the file layout. > And most especially I don't buy that someone who does want to keep using > it couldn't cope with adding an "include" to postgresql.conf manually. As Simon has already appropriately posted.... You would be incorrect. Joshua D. Drake -- Command Prompt, Inc. - http://www.commandprompt.com/ PostgreSQL Support, Training, Professional Services and Development The PostgreSQL Conference - http://www.postgresqlconference.org/ @cmdpromptinc - @postgresconf - 509-416-6579
Simon, > There are many. Tools I can name include pgpool, 2warm, PITRtools, but > there are also various tools from Sun, an IBM reseller I have > forgotten the name of, OmniTI and various other backup software > providers. Those are just the ones I can recall quickly. We've > encouraged people to write software on top and they have done so. Actually, just to correct this list: * there are no tools from Sun * pgPool2 does not deal with recovery.conf * there are additional ones: WAL-E, etc., which may or may not need to be edited. > Breaking backwards compatibility isn't something we do elsewhere, when > its easy to keep it. FWIW, I've already found that I have to modify all my backup scripts for 9.0 from 8.4, and again for 9.1 from 9.0. So we do break backwards compatibility, frequently. > I don't object to new functionality (and agreed to it upthread), just > don't break the old way when there is no need. I'm happy to make upgrades easier, but I want a path which eventually ends in recovery.conf going away. It's a bad API, confuses our users, and is difficult to support and maintain. If you think it's easier on our users to do that in stages over several versions rather than in one fell swoop, then let's plan it that way. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Fri, Sep 23, 2011 at 12:51 PM, Josh Berkus <josh@agliodbs.com> wrote: > I'm happy to make upgrades easier, but I want a path which eventually > ends in recovery.conf going away. It's a bad API, confuses our users, > and is difficult to support and maintain. I agree. GUC = Grand Unified Configuration, but recovery.conf is an example of where it's not so unified after all. We've already done a non-trivial amount of work to allow recovery.conf values to be specified without quotes, a random incompatibility with GUCs that resulted from having different parsing code for each file. If that were the last issue, then maybe it wouldn't be worth worrying about, but it's not. For example, it would be nice to have reload behavior on SIGHUP. And we keep talking about having an ALTER SYSTEM SET guc = value or SET PERMANENT guc = value command, and I think ALTER SYSTEM SET recovery_target_time = '...' would be pretty sweet. I don't want us to have to implement such things separately for postgresql.conf and recovery.conf. Now, it's true that Simon's proposal (of having recovery.conf automatically included) if it exists doesn't necessarily preclude those things. But it seems to me that it is adding a lot of complexity to core for a pretty minimal benefit to end-users, and that the semantics are not going to end up being very clean. For example, now you potentially have the situation where recovery.conf has work_mem=128MB and postgresql.conf has work_mem=4MB, and now when you end recovery you've got to make sure that everyone picks up the new setting. Now, in some sense you could say that's a feature addition, and I'm not going to deny that it might be useful to some people, but I think it's also going to require a fairly substantial convolution of the GUC machinery, and it's going to discourage people from moving away from recovery.conf. And like Josh, I think that ought to be the long-term goal, for the reasons he states. I don't want to go willy-nilly breaking third-party tools that work with PostgreSQL, but in this case I think that the reason there are so many tools in the first place is because what we're providing in core is not very good. If we are unwilling to improve it for fear of breaking compatibility with the tools, then we are stuck. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Sep23, 2011, at 17:46 , Peter Eisentraut wrote: > On tis, 2011-09-20 at 16:38 -0400, Robert Haas wrote: >> For now, I think we're best off not changing the terminology, and >> confining the remit of this patch to (a) turning all of the existing >> recovery.conf parameters into GUCs and (b) replacing recovery.conf >> with a sentinel file a sentinel file (name TBB) to indicate that the >> server is to start in recovery mode. The naming isn't great but the >> more we change at once the less chance of reaching agreement. It >> seems like we have pretty broad agreement on the basics here, so let's >> start with that. > > The only thing that's slightly bogus about that is that if you were > doing an archive recovery, you'd have to edit the main postgresql.conf > with one-shot parameters for that particular recovery (and then delete > them again, or leave them in place, confusing the next guy). But > perhaps that's worth the overall simplification. OTOH, if they're GUCs, you can specify them on the postmaster's command line. We could even get crazy and patch pg_ctl to allow <untar base backup> pg_ctl recover -D <dir> --target_xid=<the xid that broke stuff> --target_inclusive=false pg_ctl start-D <dir> best regards, Florian Pflug
Josh, >> There are many. Tools I can name include pgpool, 2warm, PITRtools, but >> there are also various tools from Sun, an IBM reseller I have >> forgotten the name of, OmniTI and various other backup software >> providers. Those are just the ones I can recall quickly. We've >> encouraged people to write software on top and they have done so. > > Actually, just to correct this list: > * there are no tools from Sun > * pgPool2 does not deal with recovery.conf I'm not sure what you mean by "not deal with" but part of pgpool-II's functionality assumes that we can easily generate recovery.conf. If reconf.conf is integrated into postgresql.conf, we need to edit postgresql.conf, which is a little bit harder than generating recovery.conf, I think. > * there are additional ones: WAL-E, etc., which may or may not need to > be edited. > >> Breaking backwards compatibility isn't something we do elsewhere, when >> its easy to keep it. > > FWIW, I've already found that I have to modify all my backup scripts for > 9.0 from 8.4, and again for 9.1 from 9.0. So we do break backwards > compatibility, frequently. > >> I don't object to new functionality (and agreed to it upthread), just >> don't break the old way when there is no need. > > I'm happy to make upgrades easier, but I want a path which eventually > ends in recovery.conf going away. It's a bad API, confuses our users, > and is difficult to support and maintain. If you think it's easier on > our users to do that in stages over several versions rather than in one > fell swoop, then let's plan it that way. > > -- > Josh Berkus > PostgreSQL Experts Inc. > http://pgexperts.com > > -- > Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) > To make changes to your subscription: > http://www.postgresql.org/mailpref/pgsql-hackers
> I'm not sure what you mean by "not deal with" but part of pgpool-II's > functionality assumes that we can easily generate recovery.conf. If > reconf.conf is integrated into postgresql.conf, we need to edit > postgresql.conf, which is a little bit harder than generating > recovery.conf, I think. Oh? Clearly I've been abusing pgPool2 then. Where's the code that generates that? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Fri, Sep 23, 2011 at 5:51 PM, Josh Berkus <josh@agliodbs.com> wrote: > Simon, > >> There are many. Tools I can name include pgpool, 2warm, PITRtools, but >> there are also various tools from Sun, an IBM reseller I have >> forgotten the name of, OmniTI and various other backup software >> providers. Those are just the ones I can recall quickly. We've >> encouraged people to write software on top and they have done so. > > Actually, just to correct this list: > * there are no tools from Sun Just for the record, I sat through a 1 hour presentation at the PostgreSQL UK conference from a Sun employee describing the product and its very clear use of PostgreSQL facilities. Josh definitely wasn't at that presentation, many others here were. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Fri, Sep 23, 2011 at 6:17 PM, Robert Haas <robertmhaas@gmail.com> wrote: > On Fri, Sep 23, 2011 at 12:51 PM, Josh Berkus <josh@agliodbs.com> wrote: >> I'm happy to make upgrades easier, but I want a path which eventually >> ends in recovery.conf going away. It's a bad API, confuses our users, >> and is difficult to support and maintain. > > I agree. > > GUC = Grand Unified Configuration, but recovery.conf is an example of > where it's not so unified after all. We've already done a non-trivial > amount of work to allow recovery.conf values to be specified without > quotes, a random incompatibility with GUCs that resulted from having > different parsing code for each file. If that were the last issue, > then maybe it wouldn't be worth worrying about, but it's not. For > example, it would be nice to have reload behavior on SIGHUP. ... > I don't want us > to have to implement such things separately for postgresql.conf and > recovery.conf. It was always my plan to do exactly the above, and there are code comments that say that from 2004. The time to replace it is now and I welcome that day and have already agreed to it. We all want every word quoted above and nothing there is under debate. > And we > keep talking about having an ALTER SYSTEM SET guc = value or SET > PERMANENT guc = value command, and I think ALTER SYSTEM SET > recovery_target_time = '...' would be pretty sweet. I don't want us > to have to implement such things separately for postgresql.conf and > recovery.conf. There is a reason why it doesn't work that way which you overlook. Please start a separate thread if you wish to discuss that. > Now, it's true that Simon's proposal (of having recovery.conf > automatically included) if it exists doesn't necessarily preclude > those things. But it seems to me that it is adding a lot of > complexity to core for a pretty minimal benefit to end-users, and that > the semantics are not going to end up being very clean. For example, > now you potentially have the situation where recovery.conf has > work_mem=128MB and postgresql.conf has work_mem=4MB, and now when you > end recovery you've got to make sure that everyone picks up the new > setting. Now, in some sense you could say that's a feature addition, > and I'm not going to deny that it might be useful to some people, but > I think it's also going to require a fairly substantial convolution of > the GUC machinery, and it's going to discourage people from moving > away from recovery.conf. And like Josh, I think that ought to be the > long-term goal, for the reasons he states. The semantics are clear: recovery.conf is read first, then postgresql.conf. It's easy to implement (1 line of code) and easy to understand. So we can support the old and the new very, very easily and clearly. "Complexity" - no definitely not. "Minimal benefit for end users" - backwards compatibility isn't minimal benefit. It's a major issue. If you put things in two places, yes that causes problems. You can already add the same parameter twice and cause exactly the same problems. > I don't want to go willy-nilly breaking third-party tools that work > with PostgreSQL, but in this case I think that the reason there are so > many tools in the first place is because what we're providing in core > is not very good. If we are unwilling to improve it for fear of > breaking compatibility with the tools, then we are stuck. No, there are many tools because there are many requirements. A simple, open API has allowed our technology to be widely used. That was by design not by chance. Nobody is unwilling to improve it. The debate is about people being unwilling to provide a simple and easy to understand backwards compatibility feature, which breaks things for no reason and does not interfere with the proposed new features. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
>> I'm not sure what you mean by "not deal with" but part of pgpool-II's >> functionality assumes that we can easily generate recovery.conf. If >> reconf.conf is integrated into postgresql.conf, we need to edit >> postgresql.conf, which is a little bit harder than generating >> recovery.conf, I think. > > Oh? Clearly I've been abusing pgPool2 then. Where's the code that > generates that? pgpool-II itself does not generate the file but scripts for pgpool-II are generating the file as stated in documentation comimg with pgpool-II. -- Tatsuo Ishii SRA OSS, Inc. Japan English: http://www.sraoss.co.jp/index_en.php Japanese: http://www.sraoss.co.jp
On Sat, Sep 24, 2011 at 9:02 AM, Simon Riggs <simon@2ndquadrant.com> wrote: > The semantics are clear: recovery.conf is read first, then > postgresql.conf. It's easy to implement (1 line of code) and easy to > understand. Eh, well, if you can implement it in one line of code, consider my objection withdrawn. I can't see how that would be possible, though. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Fri, Sep 23, 2011 at 6:55 PM, Tatsuo Ishii <ishii@postgresql.org> wrote: >>> There are many. Tools I can name include pgpool, 2warm, PITRtools, but >>> there are also various tools from Sun, an IBM reseller I have >>> forgotten the name of, OmniTI and various other backup software >>> providers. Those are just the ones I can recall quickly. We've >>> encouraged people to write software on top and they have done so. >> >> Actually, just to correct this list: >> * there are no tools from Sun >> * pgPool2 does not deal with recovery.conf > > I'm not sure what you mean by "not deal with" but part of pgpool-II's > functionality assumes that we can easily generate recovery.conf. If > reconf.conf is integrated into postgresql.conf, we need to edit > postgresql.conf, which is a little bit harder than generating > recovery.conf, I think. Since we haven't yet come up with a reasonable way of machine-editing postgresql.conf, this seems like a fairly serious objection to getting rid of recovery.conf. I wonder if there's a way we can work around that... *thinks a little* What if we modified pg_ctl to allow passing configuration parameters through to postmaster, so you could do something like this: pg_ctl start -c work_mem=8MB -c recovery_target_time='...' Would that meet pgpool's needs? (Sadly pg_ctl -c means something else right now, so we'd probably have to pick another option name, but you get the idea.) -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Simon Riggs <simon@2ndQuadrant.com> writes:
> ... The time to replace it is now and I
> welcome that day and have already agreed to it.
Okay, so you do agree that eventually we want to be rid of
recovery.conf? I think everyone else agrees on that. But if we are
going to remove recovery.conf eventually, what is the benefit of
postponing doing so? The pain is going to be inflicted sooner or later,
and the longer we wait, the more third-party code there is likely to be
that expects it to exist. If optionally reading it helped provide a
smoother transition, then maybe there would be some point. But AFAICS
having a temporary third behavior will just make things even more
complicated, not less so, for third-party code that needs to cope with
multiple versions.
> The semantics are clear: recovery.conf is read first, then
> postgresql.conf. It's easy to implement (1 line of code) and easy to
> understand.
It's not clear to me why the override order should be that way; I'd have
expected the other way myself. So this isn't as open-and-shut as you
think.
regards, tom lane
Robert Haas <robertmhaas@gmail.com> writes:
> On Fri, Sep 23, 2011 at 6:55 PM, Tatsuo Ishii <ishii@postgresql.org> wrote:
>> I'm not sure what you mean by "not deal with" but part of pgpool-II's
>> functionality assumes that we can easily generate recovery.conf. If
>> reconf.conf is integrated into postgresql.conf, we need to edit
>> postgresql.conf, which is a little bit harder than generating
>> recovery.conf, I think.
> Since we haven't yet come up with a reasonable way of machine-editing
> postgresql.conf, this seems like a fairly serious objection to getting
> rid of recovery.conf.
I don't exactly buy this argument. If postgresql.conf is hard to
machine-edit, why is recovery.conf any easier?
> What if we modified pg_ctl to allow passing configuration parameters
> through to postmaster,
You mean like pg_ctl -o?
regards, tom lane
On Sep 24, 2011, at 1:04 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > I don't exactly buy this argument. If postgresql.conf is hard to > machine-edit, why is recovery.conf any easier? Because you generally just write a brand-new file, without worrying about preserving existing settings. You aren't reallyediting at all, just writing. > >> What if we modified pg_ctl to allow passing configuration parameters >> through to postmaster, > > You mean like pg_ctl -o? Oh, cool. Yes, like that. ...Robert
On Sat, Sep 24, 2011 at 6:01 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > Simon Riggs <simon@2ndQuadrant.com> writes: >> ... The time to replace it is now and I >> welcome that day and have already agreed to it. > > Okay, so you do agree that eventually we want to be rid of > recovery.conf? I think everyone else agrees on that. But if we are > going to remove recovery.conf eventually, what is the benefit of > postponing doing so? I am happy that we don't recommend the use of recovery.conf in the future, and that the handling of the contents of recovery.conf should be identical to the handling of postgresql.conf. The latter point has always been the plan from day one. (I should not have used "it" in my sentence, since doing so always leads to confusion) My joyous rush into agreeing to removal has since been replaced with the cold reality that we must support backwards compatibility. Emphasise "must". >> The semantics are clear: recovery.conf is read first, then >> postgresql.conf. It's easy to implement (1 line of code) and easy to >> understand. > > It's not clear to me why the override order should be that way; I'd have > expected the other way myself. So this isn't as open-and-shut as you > think. I agree with you that recovery.conf as an override makes more sense, though am happy with either way. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
> Since we haven't yet come up with a reasonable way of machine-editing > postgresql.conf, this seems like a fairly serious objection to > getting > rid of recovery.conf. I wonder if there's a way we can work around > that... Well, we *did* actually come up with a reasonable way, but it died under an avalanche of bikeshedding and "we-must-do-everything-the-way-we-always-have-done". I refer, of course, to the "configuration directory" patch, which wasa fine solution, and would indeed take care of the recovery.conf issues as well had we implemented it. We can *still*implement it, for 9.2. > pg_ctl start -c work_mem=8MB -c recovery_target_time='...' This wouldn't survive a restart, and isn't compatible with init scripts.
On Sat, Sep 24, 2011 at 3:49 PM, Joshua Berkus <josh@agliodbs.com> wrote: >> Since we haven't yet come up with a reasonable way of machine-editing >> postgresql.conf, this seems like a fairly serious objection to >> getting >> rid of recovery.conf. I wonder if there's a way we can work around >> that... > > Well, we *did* actually come up with a reasonable way, but it died under an avalanche of bikeshedding and "we-must-do-everything-the-way-we-always-have-done". I refer, of course, to the "configuration directory" patch, which wasa fine solution, and would indeed take care of the recovery.conf issues as well had we implemented it. We can *still*implement it, for 9.2. Well, I find that a fairly ugly solution to the problem, but I agree that it is solvable, if we could get a critical mass on any some particular solution. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Simon Riggs <simon@2ndQuadrant.com> writes:
> On Sat, Sep 24, 2011 at 6:01 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Okay, so you do agree that eventually we want to be rid of
>> recovery.conf? �I think everyone else agrees on that. �But if we are
>> going to remove recovery.conf eventually, what is the benefit of
>> postponing doing so?
> My joyous rush into agreeing to removal has since been replaced with
> the cold reality that we must support backwards compatibility.
> Emphasise "must".
[ shrug... ] I do not agree with your conclusion. We have to break
some eggs to make this omelet. The reason why we have a mess here is
that the recovery.conf mechanism, which was designed with only the
one-shot archive-recovery case in mind, has been abused beyond its
capacity. If we don't break with past practice we are not going to be
able to fix it. And it's not like we don't break configuration file
contents in most releases anyway, so I really fail to see why this one
has suddenly become sacrosanct.
regards, tom lane
Robert Haas <robertmhaas@gmail.com> writes:
> On Sep 24, 2011, at 1:04 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> I don't exactly buy this argument. If postgresql.conf is hard to
>> machine-edit, why is recovery.conf any easier?
> Because you generally just write a brand-new file, without worrying
> about preserving existing settings. You aren't really editing at all,
> just writing.
If that's all the requirement is, it's trivial to implement.
1. Write your-random-configuration-settings into recovery.conf (or any
other file name you choose ... something named after your tool would be
a better idea).
2. Temporarily append "include recovery.conf" to the end of
postgresql.conf. Restart server.
3. When done, remove "include recovery.conf" from the end of
postgresql.conf.
The hard cases involve merging user-supplied and tool-supplied settings,
but "let's overwrite recovery.conf in toto" never would have been able
to handle such cases either.
regards, tom lane
Folks, What happens currently if we have an \include in postgresql.conf for a file which doesn't exist? Is it ignored, or do weerror out? If it could just be ignored, maybe with a note in the logs, then we could be a lot more flexible. --Josh Berkus
Joshua Berkus <josh@agliodbs.com> writes:
> What happens currently if we have an \include in postgresql.conf for a file which doesn't exist? Is it ignored, or
dowe error out?
It's an error, and I think changing that would be a really bad idea.
> If it could just be ignored, maybe with a note in the logs, then we could be a lot more flexible.
There might be a use case for a separate directive include_if_exists,
or some such name. But I think the user should have to tell us very
clearly that it's okay for the file to not be found.
regards, tom lane
> There might be a use case for a separate directive include_if_exists, > or some such name. But I think the user should have to tell us very > clearly that it's okay for the file to not be found. Better to go back to include_directory, then. --Josh Berkus
> I rather like Tom's suggestion of include_if_exists. include_if_exists certainly solves the recovery.conf/recovery.done problem. We can even phase it out, like this: 9.2: include_if_exists = 'recovery.conf' in the default postgresql.conf file. 9.3: include_if_exists = 'recovery.conf' commented out by default 9.4: renaming recovery.conf to recovery.done by core PG code removed. This gives users/vendors 3 years to update their scripts to remove dependence on recovery.conf. I'm afraid that I agreewith Simon that there's already a whole buncha 3rd-party code out there to support the current system. --Josh Berkus
Joshua Berkus <josh@agliodbs.com> writes:
> include_if_exists certainly solves the recovery.conf/recovery.done problem. We can even phase it out, like this:
> 9.2: include_if_exists = 'recovery.conf' in the default postgresql.conf file.
> 9.3: include_if_exists = 'recovery.conf' commented out by default
> 9.4: renaming recovery.conf to recovery.done by core PG code removed.
> This gives users/vendors 3 years to update their scripts to remove dependence on recovery.conf. I'm afraid that I
agreewith Simon that there's already a whole buncha 3rd-party code out there to support the current system.
If there is indeed code out there that depends on the current system,
why do you think that waiting several releases to change it will make
things better? I think it's more likely that we'd just have *more* code
that needs to be changed, and no reduction in the pain level when the
transition does finally happen.
In any case, I thought we'd agreed that the use of that file as a flag
should go away now, so I quite fail to understand your comment about 9.4.
regards, tom lane
On lör, 2011-09-24 at 13:04 -0400, Tom Lane wrote: > > What if we modified pg_ctl to allow passing configuration parameters > > through to postmaster, > > You mean like pg_ctl -o? Note that pg_ctl -o saves the parameters it uses and reapplies them after a restart. So this is not really the way to apply parameter settings only once for recovery. (Or at least it has the potential to be a very confusing way.)
On lör, 2011-09-24 at 14:02 +0100, Simon Riggs wrote: > The semantics are clear: recovery.conf is read first, then > postgresql.conf. It's easy to implement (1 line of code) and easy to > understand. What's clear about that? My intuition would have been that recovery.conf is read second.
On sön, 2011-09-25 at 12:58 -0400, Tom Lane wrote: > And it's not like we don't break configuration file > contents in most releases anyway, so I really fail to see why this one > has suddenly become sacrosanct. Well, there is a slight difference. Changes in postgresql.conf parameter names and settings are adjusted automatically for me by my package upgrade script. If we, say, changed the names of recovery.conf parameters, I'd have to get a new version of my $SuperReplicationTool. That tool could presumably look at PG_VERSION and put a recovery.conf with the right spellings in the right place. But if we completely change the way the replication configuration interacts, it's not clear that a smooth upgrade is possible without significant effort. That said, I don't see why it wouldn't be possible, but let's design with upgradability in mind, instead of claiming that we have never supported upgrades of this kind anyway.
On 09/25/2011 02:39 PM, Joshua Berkus wrote: >> There might be a use case for a separate directive include_if_exists, >> or some such name. But I think the user should have to tell us very >> clearly that it's okay for the file to not be found. > Better to go back to include_directory, then. > > I rather like Tom's suggestion of include_if_exists. There might be a case for include_directory, but I think that needs to be made separately. cheers andrew
On Tue, Sep 20, 2011 at 3:38 PM, Fujii Masao <masao.fujii@gmail.com> wrote: >>> What about "recovery.trigger"? > > I'm OK with that name. <snip> >>> * is there any security hazard from ordinary users being able to see >>> what settings had been used? >> >> primary_conninfo could be a problem, since it's possible to set a password there. > > True. I agree that primary_conninfo should be restricted to superuser. Though there is still ongonig discussion, since there is no objection about the above two changes, I revised the patch that way. And I fixed the minor bug handling the default value of recovery_target_timeline wrongly. Attached is the revised version of the patch. Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
Attachment
On Mon, Sep 26, 2011 at 7:45 PM, Peter Eisentraut <peter_e@gmx.net> wrote: > On sön, 2011-09-25 at 12:58 -0400, Tom Lane wrote: >> And it's not like we don't break configuration file >> contents in most releases anyway, so I really fail to see why this one >> has suddenly become sacrosanct. > > Well, there is a slight difference. Changes in postgresql.conf > parameter names and settings are adjusted automatically for me by my > package upgrade script. If we, say, changed the names of recovery.conf > parameters, I'd have to get a new version of my $SuperReplicationTool. > That tool could presumably look at PG_VERSION and put a recovery.conf > with the right spellings in the right place. > > But if we completely change the way the replication configuration > interacts, it's not clear that a smooth upgrade is possible without > significant effort. That said, I don't see why it wouldn't be possible, > but let's design with upgradability in mind, instead of claiming that we > have never supported upgrades of this kind anyway. Currently recovery.conf has two roles: #1. recovery.conf is used as a trigger file to enable archive recovery. At the end of recovery, recovery.conf is renamedto recovery.done. #2. recovery.conf is used as a configuration file for recovery parameters. Which role do you think we should support in 9.2 because of the backward compatibility? Both? Unless I misunderstand the discussion so far, Tom and Robert (and I) agree to get rid of both. Simon seems to agree to remove only the former role, but not the latter. How about you? If you agree to remove the former, too, let's focus on the discussion about whether the latter role should be supported in 9.2. Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
On Tue, Sep 27, 2011 at 1:33 AM, Fujii Masao <masao.fujii@gmail.com> wrote: > > Though there is still ongonig discussion, since there is no objection about > the above two changes, I revised the patch that way. And I fixed the minor > bug handling the default value of recovery_target_timeline wrongly. > Attached is the revised version of the patch. This patch no longer applies as it conflicts with the following commit: commit d56b3afc0376afe491065d9eca6440b3cc7b1346 Author: Tom Lane <tgl@sss.pgh.pa.us> Date: Sun Oct 2 16:50:04 2011 -0400 Restructure error handling in reading of postgresql.conf. Cheers, Jeff
On Tue, Sep 27, 2011 at 10:34 AM, Fujii Masao <masao.fujii@gmail.com> wrote: > On Mon, Sep 26, 2011 at 7:45 PM, Peter Eisentraut <peter_e@gmx.net> wrote: >> On sön, 2011-09-25 at 12:58 -0400, Tom Lane wrote: >>> And it's not like we don't break configuration file >>> contents in most releases anyway, so I really fail to see why this one >>> has suddenly become sacrosanct. >> >> Well, there is a slight difference. Changes in postgresql.conf >> parameter names and settings are adjusted automatically for me by my >> package upgrade script. If we, say, changed the names of recovery.conf >> parameters, I'd have to get a new version of my $SuperReplicationTool. >> That tool could presumably look at PG_VERSION and put a recovery.conf >> with the right spellings in the right place. >> >> But if we completely change the way the replication configuration >> interacts, it's not clear that a smooth upgrade is possible without >> significant effort. That said, I don't see why it wouldn't be possible, >> but let's design with upgradability in mind, instead of claiming that we >> have never supported upgrades of this kind anyway. > > Currently recovery.conf has two roles: > > #1. recovery.conf is used as a trigger file to enable archive recovery. > At the end of recovery, recovery.conf is renamed to recovery.done. > > #2. recovery.conf is used as a configuration file for recovery parameters. > > Which role do you think we should support in 9.2 because of the backward > compatibility? Both? Unless I misunderstand the discussion so far, Tom and > Robert (and I) agree to get rid of both. Simon seems to agree to remove > only the former role, but not the latter. How about you? If you agree to > remove the former, too, let's focus on the discussion about whether the > latter role should be supported in 9.2. Tatsuo/Josh/Robert also discussed how recovery.conf can be used to provide parameters solely for recovery. That is difficult to do without causing all downstream tools to make major changes in the ways they supply parameters. Keeping our APIs relatively stable is important to downstream tools. I have no objection to a brave new world, as long as you don't chuck out the one that works right now. Breaking APIs needs a good reason and I've not seen one discussed anywhere. No problem with immediately deprecating the old API and declare is planned to be removed in release 10.0. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Simon, > Tatsuo/Josh/Robert also discussed how recovery.conf can be used to > provide parameters solely for recovery. That is difficult to do > without causing all downstream tools to make major changes in the ways > they supply parameters. Actually, this case is easily solved by an "include recovery.conf" parameter. So it's a non-issue. > Keeping our APIs relatively stable is important to downstream tools. I > have no objection to a brave new world, as long as you don't chuck out > the one that works right now. Breaking APIs needs a good reason and > I've not seen one discussed anywhere. No problem with immediately > deprecating the old API and declare is planned to be removed in > release 10.0. So after debugging some of our failover scripts, here's the real-world problems I'm trying to solve. These design flaws are issues which cause automated failover or failback to abort, leading to unexpected downtime, so they are not just issues of neatness: 1. Recovery.conf being both a configuration file AND a trigger to initiate recovery mode, preventing us from separating configuration management from failover. 2. The inability to read recovery.conf parameters via SQL on a hot standby, forcing us to parse the file to find out its settings, or guess. (1) is a quite serious issue; it effectively makes recovery.conf impossible to integrate with puppet and other configuration management frameworks. I also don't see a way to fix it without breaking backwards compatibility. BTW, I'm not criticizing the original design for this. We simply didn't know better until lots of people were using these tools in production. But it's time to fix them, and the longer we wait, the more painful it will be. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On 10/10/11 10:52 AM, Josh Berkus wrote: > So after debugging some of our failover scripts, here's the real-world > problems I'm trying to solve. These design flaws are issues which cause > automated failover or failback to abort, leading to unexpected downtime, > so they are not just issues of neatness: That's "automated failover or *manual* failback". I never, ever recommend automated failback. Just FYI. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Mon, Oct 10, 2011 at 6:52 PM, Josh Berkus <josh@agliodbs.com> wrote: >> Tatsuo/Josh/Robert also discussed how recovery.conf can be used to >> provide parameters solely for recovery. That is difficult to do >> without causing all downstream tools to make major changes in the ways >> they supply parameters. > > Actually, this case is easily solved by an "include recovery.conf" > parameter. So it's a non-issue. That is what I've suggested and yes, doing that is straightforward. If you mean "do that in a program" if we had a problem with adding parameters, we also have a problem adding an include. We should avoid breaking programs which we have no reason to break. Stability is good, change without purpose is not. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Sun, Oct 9, 2011 at 9:15 AM, Jeff Janes <jeff.janes@gmail.com> wrote: > On Tue, Sep 27, 2011 at 1:33 AM, Fujii Masao <masao.fujii@gmail.com> wrote: >> >> Though there is still ongonig discussion, since there is no objection about >> the above two changes, I revised the patch that way. And I fixed the minor >> bug handling the default value of recovery_target_timeline wrongly. >> Attached is the revised version of the patch. > > This patch no longer applies as it conflicts with the following commit: Thanks! Attached patch is the updated version. Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
Attachment
On Tue, Oct 11, 2011 at 6:37 AM, Simon Riggs <simon@2ndquadrant.com> wrote: > On Mon, Oct 10, 2011 at 6:52 PM, Josh Berkus <josh@agliodbs.com> wrote: > >>> Tatsuo/Josh/Robert also discussed how recovery.conf can be used to >>> provide parameters solely for recovery. That is difficult to do >>> without causing all downstream tools to make major changes in the ways >>> they supply parameters. >> >> Actually, this case is easily solved by an "include recovery.conf" >> parameter. So it's a non-issue. > > That is what I've suggested and yes, doing that is straightforward. Even if we do that, you still need to modify the tool so that it can handle the recovery trigger file. recovery.conf is used as just a configuration file (not recovery trigger file at all). It's not renamed to recovery.done at the end of recovery. If the tool depends on the renaming from recovery.conf to recovery.done, it also would need to be modified. If the tool needs to be changed anyway, why do you hesitate in changing it so that it adds "include recovery.conf" into postgresql.conf automatically? Or you think that, to keep the backward compatibility completely, recovery.conf should be used as not only a configuration file but also a recovery trigger one and it should be renamed to recovery.done at the end of recovery? Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
Fujii Masao wrote: > On Tue, Oct 11, 2011 at 6:37 AM, Simon Riggs <simon@2ndquadrant.com> wrote: > > On Mon, Oct 10, 2011 at 6:52 PM, Josh Berkus <josh@agliodbs.com> wrote: > > > >>> Tatsuo/Josh/Robert also discussed how recovery.conf can be used to > >>> provide parameters solely for recovery. That is difficult to do > >>> without causing all downstream tools to make major changes in the ways > >>> they supply parameters. > >> > >> Actually, this case is easily solved by an "include recovery.conf" > >> parameter. ?So it's a non-issue. > > > > That is what I've suggested and yes, doing that is straightforward. > > Even if we do that, you still need to modify the tool so that it can handle > the recovery trigger file. recovery.conf is used as just a configuration file > (not recovery trigger file at all). It's not renamed to recovery.done at the > end of recovery. If the tool depends on the renaming from recovery.conf > to recovery.done, it also would need to be modified. If the tool needs to > be changed anyway, why do you hesitate in changing it so that it adds > "include recovery.conf" into postgresql.conf automatically? > > Or you think that, to keep the backward compatibility completely, > recovery.conf should be used as not only a configuration file but also a > recovery trigger one and it should be renamed to recovery.done at > the end of recovery? As much as I appreciate Simon's work in this area, I think we are still unclear if keeping backward-compatibility is worth the complexity required for future users. Historically we have been bold in changing postgresql.conf settings to improve clarity, and that approach has served us well. -- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB http://enterprisedb.com + It's impossible for everything to be true. +
On 10/10/11 9:53 PM, Fujii Masao wrote: > Or you think that, to keep the backward compatibility completely, > recovery.conf should be used as not only a configuration file but also a > recovery trigger one and it should be renamed to recovery.done at > the end of recovery? That's precisely my point. The trigger file nature of recovery.conf is a problem in itself, and I don't see any way to support that and fix it at the same time. Maybe Simon can? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Tue, Oct 11, 2011 at 3:29 PM, Bruce Momjian <bruce@momjian.us> wrote: > As much as I appreciate Simon's work in this area, I think we are still > unclear if keeping backward-compatibility is worth the complexity > required for future users. Historically we have been bold in changing > postgresql.conf settings to improve clarity, and that approach has > served us well. You raise a good point. First, thank you for the respectful comment; my viewpoint is not formed from resistance to change per se, even if may appear to be so. Thank you for raising that possibility to allow me to explain and refute that. I am genuinely concerned that we show respect to downstream software that uses our APIs and have no personal or corporate ulterior motive. Most people are used to the 3 year cycle of development on which SQLServer and Oracle have now standardised. Our 1 year cycle provides a considerable benefit in agility, but it also provides for x3 complexity in release management and a continual temptation to change for no good reason. I want to encourage people to adopt our APIs, not give them a headache for attempting to do so. We know that software exists that follows the previous API and we should take steps to deprecate that across multiple releases, with appropriate notice, just as we do in other cases, such as standard conforming strings where our lack of boldness is appropriate. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Simon Riggs wrote: > On Tue, Oct 11, 2011 at 3:29 PM, Bruce Momjian <bruce@momjian.us> wrote: > > > As much as I appreciate Simon's work in this area, I think we are still > > unclear if keeping backward-compatibility is worth the complexity > > required for future users. ?Historically we have been bold in changing > > postgresql.conf settings to improve clarity, and that approach has > > served us well. > > You raise a good point. First, thank you for the respectful comment; > my viewpoint is not formed from resistance to change per se, even if > may appear to be so. Thank you for raising that possibility to allow > me to explain and refute that. > > I am genuinely concerned that we show respect to downstream software > that uses our APIs and have no personal or corporate ulterior motive. > > Most people are used to the 3 year cycle of development on which > SQLServer and Oracle have now standardised. Our 1 year cycle provides > a considerable benefit in agility, but it also provides for x3 > complexity in release management and a continual temptation to change > for no good reason. I want to encourage people to adopt our APIs, not > give them a headache for attempting to do so. We know that software > exists that follows the previous API and we should take steps to > deprecate that across multiple releases, with appropriate notice, just > as we do in other cases, such as standard conforming strings where our > lack of boldness is appropriate. Well, let me be specific. Around 2003 to 2006, we added many new configuration parameters for logging, which required renaming or removing older parameters. There really wasn't a smooth way to allow for this to be done without impacting users, and the current system we have enjoyed since 2006 is logical only because we made the changes necessary. We can look at trying to phase changes in, but often the phasing becomes more complicated that just doing the change. Logging parameter changes were easier because it was assumed logging was an admin-only task, as I assume pitr and replication are as well. Standard conforming strings was tricky because it was more user-facing, or certainly SQL-facing. -- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB http://enterprisedb.com + It's impossible for everything to be true. +
On Tue, Oct 11, 2011 at 9:28 PM, Bruce Momjian <bruce@momjian.us> wrote: > Standard conforming strings > was tricky because it was more user-facing, or certainly SQL-facing. Why is SQL more important than backup? There is no good reason to do this so quickly. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Simon Riggs wrote: > On Tue, Oct 11, 2011 at 9:28 PM, Bruce Momjian <bruce@momjian.us> wrote: > > > Standard conforming strings > > was tricky because it was more user-facing, or certainly SQL-facing. > > Why is SQL more important than backup? Because the percentage of database users it affects is different. Administrators know when they are installing a new version of Postgres and already are probably changing these configuration files. Application binaries and perhaps application developers are not as aware of a change, and there are a far higher percentage of them in an organization than administrators. > There is no good reason to do this so quickly. I just gave you a reason above, and as I said, doing backward compatibility can make the system more complex. -- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB http://enterprisedb.com + It's impossible for everything to be true. +
Simon, I haven't see a response from you on a proposed way to keep backwards compatibility with recovery.conf as a trigger file, while also eliminating its trigger status as an unmanagable misfeature. As far as I can tell, that's the one area where we *cannot* maintain backwards compatibility. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Fri, Sep 9, 2011 at 10:56 AM, Fujii Masao <masao.fujii@gmail.com> wrote: > In previous discussion, we've reached the consensus that we should unite > recovery.conf and postgresql.conf. The attached patch does that. The > patch is WIP, I'll have to update the document, but if you notice something, > please feel free to comment. My short summary of the thread is Fujii proposes we allow parameters currently in recovery.conf to be specified in postgresql.conf. All agree to that. Fujii suggests that if we have both postgresql.conf and recovery.conf then recovery.conf should contain overrides. Fujii then suggests that if such an override exists, then SHOW would not work properly. Magnus is rightly horrified and many speak against allowing recovery.conf to continue to exist for this reason. I note that if recovery.conf is an include file of postgresql.conf then the overrides would work correctly, just as if postgresql.conf had multiple settings for that parameter. So the premise is incorrect, so the conclusion is not relevant. Simon, JD, Greg Stark speak in favour of the usefulness of having a recovery.conf separate from postgresql.conf. Tatsuo confirms pgpool uses this. Simon, Fujii, Peter agree an automatic include of recovery.conf would be useful Robert points out that pg_ctl promote was a good feature Simon, JD say that backwards compatibility is important Everybody agrees a neater way of invoking standby mode would be good. Peter points out that including recovery target parameters in postgresql.conf would be difficult and require manual editing, and also that pg_ctl -o is not a suitable interface. The thread also includes a variety of other alternate ideas, misunderstandings and other commentary. - - - My thoughts on how to resolve this are... Everybody agrees these two points: * Allow recovery parameters to be handled same way as other GUCs, and specified in postgresql.conf if desired. * Allow parameters to be reloaded at SIGHUP and visible using SHOW. Those two things do not themselves force us to break backwards compatibility. We also agree that we want a neater way to startup in standby mode. In 9.1 we added "pg_ctl promote" as a better way of indicating failover/switchover. When we did that we kept the trigger_file parameter added in 9.0, which shows it is possible to add a new API without breaking backwards compatibility. We should add a "pg_ctl standby" command as a better way of indicating starting up (also described as triggering) standby mode. We keep standby_mode parameter. There is no difference here between file based and stream based replication: you can have file, stream or both file and stream (as intended). In this mode the recovery target parameters are *ignored* even if specified (explained below). http://developer.postgresql.org/pgdocs/postgres/recovery-target-settings.html In 9.2 the presence of recovery.conf can and therefore should continue to act as it does in 9.1. This should be automatically included at the end of postgresql.conf, which naturally and with no additional code allows us to override settings, with overrides visible by SHOW. We don't make any specific checks to see if someone has added a postgresql.conf parameter in there. If there is a recovery target parameter in recovery.conf we enter recovery, otherwise we operate as a standby. recovery.conf is no longer *required* for standby modes. These things are announced as deprecated and will be removed when we go to release 10.0 * trigger_file * standby_mode * recovery.conf indicates standby recovery.conf should continue to be required to perform a PITR. If we place the recovery_target parameters into postgresql.conf we will have no way to differentiate between (1) a recovery that has successfully completed then crashed and (2) a user-specified recovery, which was the original rationale for its use. This is OK, since we now encourage people to enter a recovery by creating recovery.conf and for entering a standby to use a new cleaner API without the confusing use of the word "recovery". I think that meets all requirements, as far as technically possible. Best Regards -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Sat, Oct 29, 2011 at 7:54 PM, Simon Riggs <simon@2ndquadrant.com> wrote: > On Fri, Sep 9, 2011 at 10:56 AM, Fujii Masao <masao.fujii@gmail.com> wrote: > >> In previous discussion, we've reached the consensus that we should unite >> recovery.conf and postgresql.conf. The attached patch does that. The >> patch is WIP, I'll have to update the document, but if you notice something, >> please feel free to comment. > > My short summary of the thread is Thanks! > In 9.1 we added "pg_ctl promote" as a better way of indicating > failover/switchover. When we did that we kept the trigger_file > parameter added in 9.0, which shows it is possible to add a new API > without breaking backwards compatibility. > > We should add a "pg_ctl standby" command as a better way of indicating > starting up (also described as triggering) standby mode. We keep > standby_mode parameter. There is no difference here between file > based and stream based replication: you can have file, stream or both > file and stream (as intended). > In this mode the recovery target parameters are *ignored* even if > specified (explained below). > http://developer.postgresql.org/pgdocs/postgres/recovery-target-settings.html Agreed to add "pg_ctl standby". I think that this can be committed separately from the change of recovery.conf. > In 9.2 the presence of recovery.conf can and therefore should continue > to act as it does in 9.1. This means that recovery.conf is renamed to recovery.done at the end of recovery. IOW, all settings in recovery.conf are reset when recovery ends. Then if you run "pg_ctl reload" after recovery, you'll get something like the following error message and the reload will always fail; LOG: parameter "standby_mode" cannot be changed without restarting the server To resolve this issue, I think that we need to introduce new GUC flag indicating parameters are used only during recovery, and need to define recovery parameters with that flag. Whenever "pg_ctl reload" is executed, all the processes check whether recovery is in progress, and ignore silently the parameters with that flag if not. I'm not sure how easy we can implement this. In addition to introducing that flag, we might need to change some processes which cannot access to the shared memory so that they can. Otherwise, they cannot know whether recovery is in progress. Or we might need to change them so that they always ignore recovery parameters. Another simple but somewhat restricted approach is to read and set all parameters specified in recovery.conf by using PGC_S_OVERRIDE. If we do this, those parameters cannot be changed after startup even if recovery.conf is renamed. But the problem is that a user also cannot change their settings by reloading the configuration files. This is obviously a restriction. But it doesn't break any backward compatibility, I believe. No? If a user prefers new functionality (i.e., reload recovery parameters) rather than the backward compatibility, he/she can specify parameters in postgresql.conf. Thought? Regards, -- Fujii Masao NIPPON TELEGRAPH AND TELEPHONE CORPORATION NTT Open Source Software Center
On Mon, Oct 31, 2011 at 7:38 AM, Fujii Masao <masao.fujii@gmail.com> wrote: >> In 9.2 the presence of recovery.conf can and therefore should continue >> to act as it does in 9.1. > > This means that recovery.conf is renamed to recovery.done at the end of > recovery. IOW, all settings in recovery.conf are reset when recovery ends. > Then if you run "pg_ctl reload" after recovery, you'll get something like > the following error message and the reload will always fail; > > LOG: parameter "standby_mode" cannot be changed without restarting > the server > To resolve this issue, This issue exists whether or not we have recovery.conf etc., so yes, we must solve the problem. > I think that we need to introduce new GUC flag > indicating parameters are used only during recovery, and need to define > recovery parameters with that flag. Whenever "pg_ctl reload" is executed, > all the processes check whether recovery is in progress, and ignore > silently the parameters with that flag if not. I'm not sure how easy we > can implement this. In addition to introducing that flag, we might need to > change some processes which cannot access to the shared memory so that > they can. Otherwise, they cannot know whether recovery is in progress. > Or we might need to change them so that they always ignore recovery > parameters. The postmaster knows whether its in recovery or not without checking shared memory. Various postmaster states describe this. If not postmaster, other backends can run recoveryinprogress() as normal. It makes sense to have a new flag and that is easily created and used. > Another simple but somewhat restricted approach is to read and set > all parameters specified in recovery.conf by using PGC_S_OVERRIDE. > If we do this, those parameters cannot be changed after startup > even if recovery.conf is renamed. But the problem is that a user also > cannot change their settings by reloading the configuration files. This is > obviously a restriction. But it doesn't break any backward compatibility, > I believe. No? If a user prefers new functionality (i.e., reload recovery > parameters) rather than the backward compatibility, he/she can specify > parameters in postgresql.conf. Thought? No need to create problems. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Simon, > Everybody agrees a neater way of invoking standby mode would be good. I don't think this goes far enough. The whole recovery.conf/recovery.done thing is a serious problem for automated management of servers and automated failover. So it's not just "a neater way would be good" but "using recovery.conf as a trigger file is a broken idea and needs to be changed." > These things are announced as deprecated and will be removed when we > go to release 10.0 > * trigger_file > * standby_mode > * recovery.conf indicates standby So you're idea is that people who don't want recovery.conf to be used as a trigger file would not have the file at all, but would have something like "replication.conf" instead? If it's possible to run a replica without having a recovery.conf file, then I'm fine with your solution. If it's not, then I find your solution not to be a solution at all. > recovery.conf should continue to be required to perform a PITR. If we > place the recovery_target parameters into postgresql.conf we will have > no way to differentiate between (1) a recovery that has successfully > completed then crashed and (2) a user-specified recovery, which was > the original rationale for its use. This is OK, since we now encourage > people to enter a recovery by creating recovery.conf and for entering > a standby to use a new cleaner API without the confusing use of the > word "recovery". Sure. recovery.conf worked fine for PITR. We've just overextended it for other purposes. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Mon, Oct 31, 2011 at 7:05 PM, Josh Berkus <josh@agliodbs.com> wrote: > If it's possible to run a replica without having a recovery.conf file, > then I'm fine with your solution. If it's not, then I find your > solution not to be a solution at all. Then you are fine with the solution - not mine alone, just the sum of everybody's inputs. So we can teach the new way, while supporting the old way a while longer. > Sure. recovery.conf worked fine for PITR. We've just overextended it > for other purposes. Agreed. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Mon, Oct 31, 2011 at 3:19 PM, Simon Riggs <simon@2ndquadrant.com> wrote: > On Mon, Oct 31, 2011 at 7:05 PM, Josh Berkus <josh@agliodbs.com> wrote: > >> If it's possible to run a replica without having a recovery.conf file, >> then I'm fine with your solution. If it's not, then I find your >> solution not to be a solution at all. > > Then you are fine with the solution - not mine alone, just the sum of > everybody's inputs. > > So we can teach the new way, while supporting the old way a while longer. > In most cases we either break backwards compatibility or require some type of switch to turn on backwards compatibility for those who want it. While the above plan tries to do one better, it leaves me feeling that the thing I don't like about this is that it sounds like you are forcing backwards compatibility on people who would much rather just do things the new way. Given that, I foresee a whole new generation of confused users who end up setting their configs one way only to have someone else set the same config in the other file, or some tool dump out some config file, overriding what was really intended. This will also make things *harder* for those tool providers you are trying to help, as they will be forced to support the behavior *both ways*. I'd much rather see some type of switch which turns on the old behavior for those who really want it, because while you can teach the new behavior, if you can't prevent the old behavior, you're creating operational headaches for yourself. Robert Treat conjecture: xzilla.net consulting: omniti.com
On Mon, Oct 31, 2011 at 5:23 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
> On Mon, Oct 31, 2011 at 7:38 AM, Fujii Masao <masao.fujii@gmail.com> wrote:
>
>>> In 9.2 the presence of recovery.conf can and therefore should continue
>>> to act as it does in 9.1.
>>
>> This means that recovery.conf is renamed to recovery.done at the end of
>> recovery. IOW, all settings in recovery.conf are reset when recovery ends.
>> Then if you run "pg_ctl reload" after recovery, you'll get something like
>> the following error message and the reload will always fail;
>>
>> LOG: parameter "standby_mode" cannot be changed without restarting
>> the server
>
>> To resolve this issue,
>
> This issue exists whether or not we have recovery.conf etc., so yes,
> we must solve the problem.
No. If we don't have recovery.conf, all parameters exist in postgresql.conf.
The above issue would not occur unless a user makes a mistake, e.g.,
change the value of the parameter which cannot be changed without
the server restart.
>> I think that we need to introduce new GUC flag
>> indicating parameters are used only during recovery, and need to define
>> recovery parameters with that flag. Whenever "pg_ctl reload" is executed,
>> all the processes check whether recovery is in progress, and ignore
>> silently the parameters with that flag if not. I'm not sure how easy we
>> can implement this. In addition to introducing that flag, we might need to
>> change some processes which cannot access to the shared memory so that
>> they can. Otherwise, they cannot know whether recovery is in progress.
>> Or we might need to change them so that they always ignore recovery
>> parameters.
>
> The postmaster knows whether its in recovery or not without checking
> shared memory. Various postmaster states describe this. If not
> postmaster, other backends can run recoveryinprogress() as normal.
AFAIR archiver and syslogger cannot access to the shared memory, i.e.,
they cannot run RecoveryInProgress(). They don't use any recovery
parameters for now, so we can change them so that they always ignore
those parameters. Though I'm not inclined to add the process-specific code
like the following into the GUC mechanism as much as possible.
if (I am postmaster) { if (recovery is NOT in progress) reset the recovery parameters; } else
if(I am archiver or syslogger) /* always ignore */ else { if (recovery is NOT in progress)
resetthe recovery parameters; }
Regards,
--
Fujii Masao
NIPPON TELEGRAPH AND TELEPHONE CORPORATION
NTT Open Source Software Center
On Tue, Nov 1, 2011 at 5:59 AM, Fujii Masao <masao.fujii@gmail.com> wrote:
> On Mon, Oct 31, 2011 at 5:23 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
>> On Mon, Oct 31, 2011 at 7:38 AM, Fujii Masao <masao.fujii@gmail.com> wrote:
>>
>>>> In 9.2 the presence of recovery.conf can and therefore should continue
>>>> to act as it does in 9.1.
>>>
>>> This means that recovery.conf is renamed to recovery.done at the end of
>>> recovery. IOW, all settings in recovery.conf are reset when recovery ends.
>>> Then if you run "pg_ctl reload" after recovery, you'll get something like
>>> the following error message and the reload will always fail;
>>>
>>> LOG: parameter "standby_mode" cannot be changed without restarting
>>> the server
>>
>>> To resolve this issue,
>>
>> This issue exists whether or not we have recovery.conf etc., so yes,
>> we must solve the problem.
>
> No. If we don't have recovery.conf, all parameters exist in postgresql.conf.
> The above issue would not occur unless a user makes a mistake, e.g.,
> change the value of the parameter which cannot be changed without
> the server restart.
I don't understand what you are saying. You seem to be suggesting that
it would be OK for someone to set standby_mode = on and reload the
config and for that to go completely unremarked. If we are adding new
parameters to postgresql.conf then its clear that some people will
misconfigure them and we need to have error messages for that.
If you change a parameter that only has effect during recovery then
must get an error if it is changed during normal running.
That is the reason we need to mark the recovery parameters with a new
flag, so that we can ignore any errors caused by changing them.
That has nothing at all to do with recovery.conf - you need it even if
you put everything in postgresql.conf.
>>> I think that we need to introduce new GUC flag
>>> indicating parameters are used only during recovery, and need to define
>>> recovery parameters with that flag. Whenever "pg_ctl reload" is executed,
>>> all the processes check whether recovery is in progress, and ignore
>>> silently the parameters with that flag if not. I'm not sure how easy we
>>> can implement this. In addition to introducing that flag, we might need to
>>> change some processes which cannot access to the shared memory so that
>>> they can. Otherwise, they cannot know whether recovery is in progress.
>>> Or we might need to change them so that they always ignore recovery
>>> parameters.
>>
>> The postmaster knows whether its in recovery or not without checking
>> shared memory. Various postmaster states describe this. If not
>> postmaster, other backends can run recoveryinprogress() as normal.
>
> AFAIR archiver and syslogger cannot access to the shared memory, i.e.,
> they cannot run RecoveryInProgress(). They don't use any recovery
> parameters for now, so we can change them so that they always ignore
> those parameters. Though I'm not inclined to add the process-specific code
> like the following into the GUC mechanism as much as possible.
>
> if (I am postmaster)
> {
> if (recovery is NOT in progress)
> reset the recovery parameters;
> }
> else if (I am archiver or syslogger)
> /* always ignore */
> else
> {
> if (recovery is NOT in progress)
> reset the recovery parameters;
> }
>
I don't see the problem. The code above could easily be simplified and
only needs to exist in one place, not copied around for each GUC.
When we had recovery.conf and postgresql.conf you knew which
parameters took effects at specific times. That metadata needs to be
added back into the system so we can report errors properly.
Considering the amount of code we will be removing, a couple of extra
lines seems trivial.
AFAICS we need to reset all recovery parameters at the end of recovery
anyway. Having SHOW recovery_target_xid return a value other than 0 is
not appropriate during normal running, same for standby_mode and the
other parameters.
--
Simon Riggs http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services
On Tue, Nov 1, 2011 at 7:46 AM, Simon Riggs <simon@2ndquadrant.com> wrote: > If you change a parameter that only has effect during recovery then > must get an error if it is changed during normal running. I don't see why. If you're in normal running and someone changes a parameter that is irrelevant during normal running, that should be a no-op, not an error. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Tue, Nov 1, 2011 at 12:06 PM, Robert Haas <robertmhaas@gmail.com> wrote: > On Tue, Nov 1, 2011 at 7:46 AM, Simon Riggs <simon@2ndquadrant.com> wrote: >> If you change a parameter that only has effect during recovery then >> must get an error if it is changed during normal running. > > I don't see why. If you're in normal running and someone changes a > parameter that is irrelevant during normal running, that should be a > no-op, not an error. How will it be made into a no-op, except by having a specific flag to show that it is irrelevant during normal running? Fujii is saying we only need to mark GUCs if we keep recovery.conf. I am saying we need to mark them whatever we do elsewhere. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Tue, Nov 1, 2011 at 8:14 AM, Simon Riggs <simon@2ndquadrant.com> wrote: > On Tue, Nov 1, 2011 at 12:06 PM, Robert Haas <robertmhaas@gmail.com> wrote: >> On Tue, Nov 1, 2011 at 7:46 AM, Simon Riggs <simon@2ndquadrant.com> wrote: >>> If you change a parameter that only has effect during recovery then >>> must get an error if it is changed during normal running. >> >> I don't see why. If you're in normal running and someone changes a >> parameter that is irrelevant during normal running, that should be a >> no-op, not an error. > > How will it be made into a no-op, except by having a specific flag to > show that it is irrelevant during normal running? By default, changing a GUC just updates the value of some global variable inside every backend. But unless there's some code that makes use of that global variable for some purpose, it doesn't have any practical effect. Apart from whatever complexities may be imposed by our choice of implementation, I don't see how this would be any different from setting maintenance_work_mem in a particular session and then not running any CREATE INDEX or VACUUM commands in that session. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Tue, Nov 1, 2011 at 9:45 AM, Simon Riggs <simon@2ndquadrant.com> wrote: > Why do we have this log message then, if it is OK to ignore changes > that have no effect? > > LOG: parameter "shared_buffers" cannot be changed without restarting the server I believe we're logging the fact that we were unable to make the change, not the fact that it didn't have any effect. Certainly, it *would* have an effect, if we were able to make it. But we can't, without a restart, so we tell that to the user. But, for example, the hot_standby GUC - which already exists - does not do anything in normal running. We don't need to (and don't) complain if the user tries to change the value in normal running, though: they're presumably just hoping it will take effect the next time they start up, which it will. And the autovacuum parameter does not do anything during hot standby, but we don't need to (and don't) complain if the user changes it them; it just takes effect when we enter normal running. The only time I think we need to complain about an effort to change a GUC is when the GUC won't take effect as soon as the user might normally expect. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Tue, Nov 1, 2011 at 1:23 PM, Robert Haas <robertmhaas@gmail.com> wrote: > On Tue, Nov 1, 2011 at 8:14 AM, Simon Riggs <simon@2ndquadrant.com> wrote: >> On Tue, Nov 1, 2011 at 12:06 PM, Robert Haas <robertmhaas@gmail.com> wrote: >>> On Tue, Nov 1, 2011 at 7:46 AM, Simon Riggs <simon@2ndquadrant.com> wrote: >>>> If you change a parameter that only has effect during recovery then >>>> must get an error if it is changed during normal running. >>> >>> I don't see why. If you're in normal running and someone changes a >>> parameter that is irrelevant during normal running, that should be a >>> no-op, not an error. >> >> How will it be made into a no-op, except by having a specific flag to >> show that it is irrelevant during normal running? > > By default, changing a GUC just updates the value of some global > variable inside every backend. But unless there's some code that > makes use of that global variable for some purpose, it doesn't have > any practical effect. Apart from whatever complexities may be imposed > by our choice of implementation, I don't see how this would be any > different from setting maintenance_work_mem in a particular session > and then not running any CREATE INDEX or VACUUM commands in that > session. Why do we have this log message then, if it is OK to ignore changes that have no effect? LOG: parameter "shared_buffers" cannot be changed without restarting the server -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Tue, Nov 1, 2011 at 2:04 PM, Robert Haas <robertmhaas@gmail.com> wrote: > On Tue, Nov 1, 2011 at 9:45 AM, Simon Riggs <simon@2ndquadrant.com> wrote: >> Why do we have this log message then, if it is OK to ignore changes >> that have no effect? >> >> LOG: parameter "shared_buffers" cannot be changed without restarting the server > > I believe we're logging the fact that we were unable to make the > change, not the fact that it didn't have any effect. Certainly, it > *would* have an effect, if we were able to make it. But we can't, > without a restart, so we tell that to the user. > > But, for example, the hot_standby GUC - which already exists - does > not do anything in normal running. We don't need to (and don't) > complain if the user tries to change the value in normal running, > though: they're presumably just hoping it will take effect the next > time they start up, which it will. If there is precedence then we should follow it. So no error messages. The only reason this point has any importance is that Fujii is suggesting the code will be much more complicated if we retain recovery.conf compatibility, since we need to mark GUCs with a flag as to whether they can take effect during recovery. I don't think that's a lot of work to add, and would make the code clearer than it is now. > And the autovacuum parameter does > not do anything during hot standby, but we don't need to (and don't) > complain if the user changes it them; it just takes effect when we > enter normal running. That one needs to be set on a standby, in case it becomes a primary. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Tue, Nov 1, 2011 at 6:36 PM, Josh Berkus <josh@agliodbs.com> wrote: > On 11/1/11 10:34 AM, Simon Riggs wrote: >> On Tue, Nov 1, 2011 at 5:11 PM, Joshua Berkus <josh@agliodbs.com> wrote: >> >>> So, we have four potential paths regarding recovery.conf: >>> >>> 1) Break backwards compatibility entirely, and stop supporting recovery.conf as a trigger file at all. >> >> Note that is exactly what I have suggested when using "standby" mode >> from pg_ctl. > > I wasn't clear on that from the description of your proposal. So are > you suggesting that, if we start postgresql with "pg_ctl standby" then > recovery.conf would not behave as a trigger file? Yes -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Tue, Nov 1, 2011 at 6:12 PM, Robert Treat <rob@xzilla.net> wrote: > On Tue, Nov 1, 2011 at 1:34 PM, Simon Riggs <simon@2ndquadrant.com> wrote: >> On Tue, Nov 1, 2011 at 5:11 PM, Joshua Berkus <josh@agliodbs.com> wrote: >> >>> So, we have four potential paths regarding recovery.conf: >>> >>> 1) Break backwards compatibility entirely, and stop supporting recovery.conf as a trigger file at all. >> >> Note that is exactly what I have suggested when using "standby" mode >> from pg_ctl. >> >> But you already know that, since you said "If it's possible to run a >> replica without having a recovery.conf file, >> then I'm fine with your solution", and I already confirmed back to you >> that would be possible. >> > > "It's possible to run a replica without having a recovery.conf file" > is not the same thing as "If someone makes a recovery.conf file, it > won't break my operations". AIUI, you are not supporting the latter. Yes, that is part of the "combined proposal", which allows both old and new APIs. New API pg_ctl standby will startup a server in standby mode, do not implicitly include recovery.conf and disallow recovery_target parameters in postgresql.conf (you may, if you wish, explicitly have "include recovery.conf" in your postgresql.conf, should you desire that) Old API pg_ctl start and a recovery.conf has been created will startup a server in PITR and/or replication, recovery.conf will be read automatically (as now) so the presence of the recovery.conf acts as a trigger, only if we issue "start" So the existing syntax works exactly as now, but a new API has been created to simplify things in exactly the way requested. The old and the new API don't mix, so there is no confusion between them. You must still use the old API when you wish to perform a PITR, as explained by me, following comments by Peter. There is no significant additional code or complexity required to do this, but it adds considerable usefulness. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Robert, > In most cases we either break backwards compatibility or require some > type of switch to turn on backwards compatibility for those who want > it. While the above plan tries to do one better, it leaves me feeling > that the thing I don't like about this is that it sounds like you are > forcing backwards compatibility on people who would much rather just > do things the new way. Given that, I foresee a whole new generation > of > confused users who end up setting their configs one way only to have > someone else set the same config in the other file, or some tool dump > out some config file, overriding what was really intended. This will > also make things *harder* for those tool providers you are trying to > help, as they will be forced to support the behavior *both ways*. I'd > much rather see some type of switch which turns on the old behavior > for those who really want it, because while you can teach the new > behavior, if you can't prevent the old behavior, you're creating > operational headaches for yourself. This is a good point. There's also the second drawback, which is complexity of code, which I believe that Tom Lane has broughtup before; having two separate-but-equal paths for configuration is liable to lead to a lot of bugs. So, we have four potential paths regarding recovery.conf: 1) Break backwards compatibility entirely, and stop supporting recovery.conf as a trigger file at all. 2) Offer backwards compatibility only if "recovery_conf='filename'" is set in postgresql.conf, then behave like Simon's compromise. 3) Simon's compromise. 4) Don't ever change how recovery.conf works. The only two of the above I see as being real options are (1) and (2). (3) would, as Robert points out, cause DBAs to haveunpleasant surprises when some third-party tool creates a recovery.conf they weren't expecting. So: (1) pros: * new, clean API * makes everyone update their tools * no confusion on "how to do failover" * code simplicitycons: * breaks a bunch of 3rd-party tools * or forces them to maintain separate 9.1 and 9.2 branches (2) pros: * allows people to use only new API if they want * allows gradual update of tools * can also lump in relocatablerecovery.conf as feature cons: * puts off the day when vendors pay attention to the new API (and even morekicking & screaming when that day comes) * confusion about "how to do failover" * code complexity
On 11/1/11 10:34 AM, Simon Riggs wrote: > On Tue, Nov 1, 2011 at 5:11 PM, Joshua Berkus <josh@agliodbs.com> wrote: > >> So, we have four potential paths regarding recovery.conf: >> >> 1) Break backwards compatibility entirely, and stop supporting recovery.conf as a trigger file at all. > > Note that is exactly what I have suggested when using "standby" mode > from pg_ctl. I wasn't clear on that from the description of your proposal. So are you suggesting that, if we start postgresql with "pg_ctl standby" then recovery.conf would not behave as a trigger file? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Tue, Nov 1, 2011 at 2:34 PM, Simon Riggs <simon@2ndquadrant.com> wrote: > On Tue, Nov 1, 2011 at 6:12 PM, Robert Treat <rob@xzilla.net> wrote: >> On Tue, Nov 1, 2011 at 1:34 PM, Simon Riggs <simon@2ndquadrant.com> wrote: >>> On Tue, Nov 1, 2011 at 5:11 PM, Joshua Berkus <josh@agliodbs.com> wrote: >>> >>>> So, we have four potential paths regarding recovery.conf: >>>> >>>> 1) Break backwards compatibility entirely, and stop supporting recovery.conf as a trigger file at all. >>> >>> Note that is exactly what I have suggested when using "standby" mode >>> from pg_ctl. >>> >>> But you already know that, since you said "If it's possible to run a >>> replica without having a recovery.conf file, >>> then I'm fine with your solution", and I already confirmed back to you >>> that would be possible. >>> >> >> "It's possible to run a replica without having a recovery.conf file" >> is not the same thing as "If someone makes a recovery.conf file, it >> won't break my operations". AIUI, you are not supporting the latter. > > Yes, that is part of the "combined proposal", which allows both old > and new APIs. > > New API > > pg_ctl standby > will startup a server in standby mode, do not implicitly include > recovery.conf and disallow recovery_target parameters in > postgresql.conf > (you may, if you wish, explicitly have "include recovery.conf" in > your postgresql.conf, should you desire that) > > Old API > > pg_ctl start > and a recovery.conf has been created > will startup a server in PITR and/or replication, recovery.conf > will be read automatically (as now) > so the presence of the recovery.conf acts as a trigger, only if we > issue "start" > > So the existing syntax works exactly as now, but a new API has been > created to simplify things in exactly the way requested. The old and > the new API don't mix, so there is no confusion between them. > > You must still use the old API when you wish to perform a PITR, as > explained by me, following comments by Peter. > Ah, thanks for clarifying, your earlier proposal didn't read that way to me. It still doesn't solve the problem for tool makers of needing to be able to deal with two possible implementation methods, but it should be easier for them to make a choice. One thing though, I think it would be better to have this work the other way around. ISTM we're going to end up telling people to avoid using pg_ctl start and instead use pg_ctl standby, which doesn't feel like the right answer. Ie. "Starting in 9.2, you should use pg_ctl standby to launch your database for normal operations and/or in cases where you are writing init scripts to control your production databases. For backwards compatibility, if you require the old behavior of using a recovery.conf, we would recommend you use pg_ctl start instead". Robert Treat conjecture: xzilla.net consulting: omniti.com
On Tue, Nov 1, 2011 at 1:34 PM, Simon Riggs <simon@2ndquadrant.com> wrote: > On Tue, Nov 1, 2011 at 5:11 PM, Joshua Berkus <josh@agliodbs.com> wrote: > >> So, we have four potential paths regarding recovery.conf: >> >> 1) Break backwards compatibility entirely, and stop supporting recovery.conf as a trigger file at all. > > Note that is exactly what I have suggested when using "standby" mode > from pg_ctl. > > But you already know that, since you said "If it's possible to run a > replica without having a recovery.conf file, > then I'm fine with your solution", and I already confirmed back to you > that would be possible. > "It's possible to run a replica without having a recovery.conf file" is not the same thing as "If someone makes a recovery.conf file, it won't break my operations". AIUI, you are not supporting the latter. Robert Treat conjecture: xzilla.net consulting: omniti.com
On Tue, Nov 1, 2011 at 5:11 PM, Joshua Berkus <josh@agliodbs.com> wrote: > So, we have four potential paths regarding recovery.conf: > > 1) Break backwards compatibility entirely, and stop supporting recovery.conf as a trigger file at all. Note that is exactly what I have suggested when using "standby" mode from pg_ctl. But you already know that, since you said "If it's possible to run a replica without having a recovery.conf file, then I'm fine with your solution", and I already confirmed back to you that would be possible. > 2) Offer backwards compatibility only if "recovery_conf='filename'" is set in postgresql.conf, then behave like Simon'scompromise. > > 3) Simon's compromise. See above. Calling it a compromise in this way implies nobody has been given exactly what they ask for, but that is not the case. > 4) Don't ever change how recovery.conf works. > > The only two of the above I see as being real options are (1) and (2). (3) would, as Robert points out, cause DBAs tohave unpleasant surprises when some third-party tool creates a recovery.conf they weren't expecting. So: Please read my proposal again. I'll be happy to answer questions if you have any. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On 11/1/11 12:29 PM, Robert Treat wrote: > "Starting in 9.2, you should use pg_ctl standby to launch your > database for normal operations and/or in cases where you are writing > init scripts to control your production databases. For backwards > compatibility, if you require the old behavior of using a > recovery.conf, we would recommend you use pg_ctl start instead". Gah. There is no way we're getting distro packagers to switch from pg_ctl start. Also, a lot of distros use the "postgres" command rather than pg_ctl anything. Messing with pg_ctl is not really a solution for this, since few people in production environments call it directly. Nobody I know, anyway. So Simon's suggested compromise still puts backwards compatibility ahead of promoting the new API. This would result in nobody supporting the new API until the day we remove the old one from the code. I think adding "recovery_conf_location = ''" to postgresql.conf is a much better compromise. Assuming we can stand the code complexity ... -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Tue, Nov 1, 2011 at 9:11 PM, Josh Berkus <josh@agliodbs.com> wrote: > On 11/1/11 12:29 PM, Robert Treat wrote: >> "Starting in 9.2, you should use pg_ctl standby to launch your >> database for normal operations and/or in cases where you are writing >> init scripts to control your production databases. For backwards >> compatibility, if you require the old behavior of using a >> recovery.conf, we would recommend you use pg_ctl start instead". > > Gah. > > There is no way we're getting distro packagers to switch from pg_ctl > start. Also, a lot of distros use the "postgres" command rather than > pg_ctl anything. > > Messing with pg_ctl is not really a solution for this, since few people > in production environments call it directly. Nobody I know, anyway. > > So Simon's suggested compromise still puts backwards compatibility ahead > of promoting the new API. This would result in nobody supporting the > new API until the day we remove the old one from the code. Which will never happen, since part of the proposal is that PITR will only be supported using the old method. > I think adding "recovery_conf_location = ''" to postgresql.conf is a > much better compromise. Assuming we can stand the code complexity ... I think that might have some possibilities. But how does that work in detail? If you set it to empty, then the recovery_* parameters are just GUCs, I suppose: which seems fine. But if you set it to a non-empty value then what happens, exactly? The recovery.conf settings clobber anything in postgresql.conf, and when we exit recovery we reload the config, discarding any settings we got from recovery.conf? That might not be too bad. I think we need to back up and figure out what problem we're trying to solve here. IMV, the answer is "setting up a standby is too complicated and requiring yet another configuration file to do it makes it even more complicated". If the mechanism we introduce to "solve" that problem is more complicated than what we have now, it might end up being a net regression in terms of usability. I feel like changing everything that's currently in recovery.conf to GUCs and implementing SET PERSISTENT would give everyone what they need, admittedly without perfect backward compatibility, but perhaps close enough for government work, and a step forward overall. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Wed, Nov 2, 2011 at 1:45 AM, Robert Haas <robertmhaas@gmail.com> wrote: > I think that might have some possibilities. But how does that work in > detail? My thoughts also. I want to see the detail on an alternate proposal so we can decide things sensibly. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Wed, Nov 2, 2011 at 1:11 AM, Josh Berkus <josh@agliodbs.com> wrote: > There is no way we're getting distro packagers to switch from pg_ctl > start. Also, a lot of distros use the "postgres" command rather than > pg_ctl anything. So backwards compatibility is important for downstream software. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On Wed, Nov 2, 2011 at 1:11 AM, Josh Berkus <josh@agliodbs.com> wrote: > There is no way we're getting distro packagers to switch from pg_ctl > start. Also, a lot of distros use the "postgres" command rather than > pg_ctl anything. So backwards compatibility is important for downstream software. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
RH, Simon, > I think that might have some possibilities. But how does that work in > detail? If you set it to empty, then the recovery_* parameters are > just GUCs, I suppose: which seems fine. But if you set it to a > non-empty value then what happens, exactly? The recovery.conf > settings clobber anything in postgresql.conf, and when we exit > recovery we reload the config, discarding any settings we got from > recovery.conf? That might not be too bad. Yeah, that's what I was picturing. By tying backwards-compatibility to a setting in pg.conf, you eliminate a lot of changes for a DBA accidentally enabling it. This also then supports re-locating the recovery.conf file, which has been an issue for a long time. > I think we need to back up and figure out what problem we're trying to > solve here. IMV, the answer is "setting up a standby is too > complicated and requiring yet another configuration file to do it > makes it even more complicated". If the mechanism we introduce to > "solve" that problem is more complicated than what we have now, it > might end up being a net regression in terms of usability. Well, as someone who sets up and admins replication for a bunch of clients, here's what I'd like to see: 1. no more using a configuration file as a trigger 2. ability to put replication configuration in postgresql.conf or in a manually designated include file 3. having replication configuration show up in pg_settings The three settings above would make my life as a contract DBA much easier ... and I presume help a lot of our users like me. Among other things, fixing the 3 things above would make replication integrate a lot better with configuration management systems and monitoring. > I feel like changing everything that's currently in recovery.conf to > GUCs and implementing SET PERSISTENT would give everyone what they > need, admittedly without perfect backward compatibility, but perhaps > close enough for government work, and a step forward overall. Is anyone working on SET PERSISTENT? I thought that got bike-shedded to death. > So backwards compatibility is important for downstream software. > If it wasn't, we wouldn't be having this discussion. However, backwards compatibility is not necessarily the *most* important consideration. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Wed, Nov 2, 2011 at 2:48 PM, Josh Berkus <josh@agliodbs.com> wrote: > Is anyone working on SET PERSISTENT? I thought that got bike-shedded to > death. I think we had a fairly good sketch of how it could work mapped out, mostly based around adding a postgresql.auto file. I could dig up the old discussions, if need be. Basically, I'd be willing to get that implemented and bull it through to completion in time for 9.2 if having that would make it easier for everyone to accept the idea of turning the recovery.conf parameters into GUCs. Otherwise I probably will not get to it this cycle. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Wed, Nov 2, 2011 at 6:48 PM, Josh Berkus <josh@agliodbs.com> wrote: > Well, as someone who sets up and admins replication for a bunch of > clients, here's what I'd like to see: Everyone has their own set of requirements. I've tried hard to fuse those together into a useful proposal, listening to all. Please bear in mind that I make my living in exactly the same way you do, so you must surely be aware I do this solely in the common interest. I don't force you to accept that proposal, but challenging it does require somebody to step up to the plate and work out a better detailed proposal rather than just restate what they personally want. I don't rule out out that a better proposal exists and would be incredibly happy if someone worked out the design, wrote it up and desnaggled it. We can always do nothing, which is a safe and viable option. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Simon, > Everyone has their own set of requirements. I've tried hard to fuse > those together into a useful proposal, listening to all. Please bear > in mind that I make my living in exactly the same way you do, so you > must surely be aware I do this solely in the common interest. Thank you for giving us a place to start. I have seen and commented on your compromise. Both Robert Treat and I poked holes in it. It was a good first attempt, but not a final attempt -- your compromise proposal was heavily skewed towards maintaining the status quo at the expense of improving functionality. As a compromise, it was "70% Simon, 30% everyone else". > I don't force you to accept that proposal, but challenging it does > require somebody to step up to the plate and work out a better > detailed proposal rather than just restate what they personally want. I made a proposal, which was based on modifying (and I believe, improving) your proposal. > We can always do nothing, which is a safe and viable option. Not really, no. The whole recovery.conf thing is very broken and inhibits adoption of PostgreSQL because our users can't figure it out. You've made it pretty clear that you're personally comfortable with how replication configuration works now, and aren't really interested in any changes. That's certainly a valid viewpoint, but the users and contributors who find the API horribly unusable also have a valid viewpoint. You don't automatically win arguments because you're on the side of backwards compatibility. When we released binary replication in 9.0, I thought everyone knew that it was a first cut and that we'd be making some dramatic changes -- including ones which broke things -- over the next few versions. There was simply no way for us to know real user requirements until the feature was in the field and being deployed in production. We would discover some things which really didn't work and that we had to break and remake. And we have. Now you are arguing for premature senescence, where our first API becomes our only API now and forever. That's a road to project death. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On 09/24/2011 04:49 PM, Joshua Berkus wrote: > Well, we *did* actually come up with a reasonable way, but it died > under an avalanche of bikeshedding and > "we-must-do-everything-the-way-we-always-have-done". I refer, of > course, to the "configuration directory" patch, which was a fine > solution, and would indeed take care of the recovery.conf issues as > well had we implemented it. We can *still* implement it, for 9.2. That actually died from a lack of round-tuits, the consensus at the end of the bike-sheeding was pretty clear. Last night I finally got motivated to fix the bit rot and feature set on that patch, to match what seemed to be the easiest path toward community approval. One known bug left to resolve and I think it's ready to submit for the next CF. I think includeifexists is also a good improvement, too, on a related arc to the main topic here. If I can finish off the directory one (or get someone else to fix my bug) I should be able to follow up with that one. The patches won't be that different. -- Greg Smith 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.us
Josh Berkus wrote: > > We can always do nothing, which is a safe and viable option. > > Not really, no. The whole recovery.conf thing is very broken and > inhibits adoption of PostgreSQL because our users can't figure it out. > > You've made it pretty clear that you're personally comfortable with how > replication configuration works now, and aren't really interested in any > changes. That's certainly a valid viewpoint, but the users and > contributors who find the API horribly unusable also have a valid > viewpoint. You don't automatically win arguments because you're on the > side of backwards compatibility. > > When we released binary replication in 9.0, I thought everyone knew that > it was a first cut and that we'd be making some dramatic changes -- > including ones which broke things -- over the next few versions. There > was simply no way for us to know real user requirements until the > feature was in the field and being deployed in production. We would > discover some things which really didn't work and that we had to break > and remake. And we have. > > Now you are arguing for premature senescence, where our first API > becomes our only API now and forever. That's a road to project death. Agreed. This thread has already expended too much of our valuable time, in my opinion. I think we have enough agreement that we need a new API, so let's design one. If we can add some backward-compatibility here, great, but let's not have that driving the discussion. Replication is already complex enough that having two ways to set this up just adds confusion. Replication/PITR does not affect SQL or applications --- it affects admin scripts and tools, so they are just going to have to adjust. We are not going to make everyone happy, so let's just move forward --- if people want to pout in the corner, I really don't care. -- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB http://enterprisedb.com + It's impossible for everything to be true. +
> Agreed. This thread has already expended too much of our valuable time, > in my opinion. As I said elsewhere, I think that a modified version of Simon's proposal gets us all what we want except code cleanliness. I'd like to hear from Tom on that issue. The proposal is: 1. No changes are expected to PITR mode, unless required by the other changes below. 2. standby_mode becomes an ENUM: "off,standby,pitr". It can be reset on server reload or through "pg_ctl promote" 3. we add a "recovery_file" filepath parameter to postgresql.conf. This points to the location of recovery.conf, if any, but does not error fatally if the file doesn't exist, allowing recovery.conf/.done if we want to support it.3.a. we continue to support recovery.conf/.done trigger file behavior if recovery_file is set. In this case, the recovery.conf file would contain the standby_mode GUC. 3.b. should pitr mode require recovery_file? Doesn't seem like it. 3.c. we begin arguments on whether or not recovery_file should be set by default 4. Haas implements SET PERSISTENT for postgresql.conf 5. pg_ctl promote uses SET PERSISTENT to change standby_mode so that former standbys can be permanently promoted. 6. (optional) we add include_if_exists specifier for postgresql.conf, allowing scripters to handle having a separate recovery.conf file in other ways. 7. (optional) we add a "pg_ctl standby" command which starts up postgres and sets standby_mode to "standby". I'm not sure I see the value in this though. The above allows DBAs to operate in "backwards compatibility mode" without forcing authors of new tools and scripts to abide by it. Question: if both standby_mode and recovery_file are set, what should happen if the recovery_file is not present? For backwards compatibility, we would use SET PERSISTENT to set standby_mode after the server comes up. Arguments? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
Josh Berkus <josh@agliodbs.com> writes:
> As I said elsewhere, I think that a modified version of Simon's proposal
> gets us all what we want except code cleanliness. I'd like to hear from
> Tom on that issue.
Well, code complexity is hard to gauge without coding a draft
implementation, but I think this largely fails on UI complexity.
It's still paying too high a price for backwards compatibility IMO.
The more I read about this, the more doubtful I am that unifying PITR
recovery with standby mode is a good idea from the UI perspective. They
may share a lot of common infrastructure but they need to be triggered
in fundamentally different ways.
In particular, one of the reasons that recovery.conf/.done was set up
the way that it was was to have a simple way of letting the system get
out of PITR mode; without that, a crash during PITR recovery is going
to lead to restarting from the PITR start point (because when we
restart, we find configuration settings telling us to do that).
We could possibly move the necessary state into pg_control, but keeping
it as a GUC is going to be a mess. On the whole I still think a trigger
file is a sane design for that. It may make sense to move the
configuration data somewhere else, but we really need to be able to tell
the difference between "starting PITR", "continuing PITR after a
mid-recovery crash", and "finished PITR, up and running normally".
A GUC is not a good way to do that.
The angst around this issue seems to me to largely stem from trying to
use a configuration setup designed for one-shot PITR scenarios for
hot standby scenarios, which are really pretty different. We have to
divorce those two cases before we're going to have something that's
sane and usable ... and AFAICS that means giving up backwards
compatibility to some degree.
We got this wrong in 9.0, which everyone understood at the time was an
unpolished prototype implementation of replication. I don't think it's
sensible to now move the goalposts and decree that we've got to be
fully backward compatible with our first-cut mistakes.
regards, tom lane
> configuration data somewhere else, but we really need to be able to tell > the difference between "starting PITR", "continuing PITR after a > mid-recovery crash", and "finished PITR, up and running normally". > A GUC is not a good way to do that. Does a GUC make sense to you for how to handle standby/master for replication? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Mon, Nov 7, 2011 at 2:14 PM, Josh Berkus <josh@agliodbs.com> wrote: > 2. standby_mode becomes an ENUM: "off,standby,pitr". It can be reset on > server reload or through "pg_ctl promote" I'm a little bit confused by the way we're dragging standby_mode into this conversation. If you're using pg_standby, you can set standby_mode=off and still have a standby. If you're using a simple recovery command that just copies files, you need to set standby_mode=on if you don't want the standby to exit recovery and promote. But I think of standby_mode as meaning "should we use the internal standby loop rather than depending on an external tool?" rather than "should we become a standby?". Maybe I'm confused. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
All: *ping* Trying to restart this discussion, since the feature bogged down on spec. We have consensus that we need to change how replication mode is mangaged; surely we can reach consensus on how to change it? On 11/8/11 11:39 AM, Josh Berkus wrote: > >> configuration data somewhere else, but we really need to be able to tell >> the difference between "starting PITR", "continuing PITR after a >> mid-recovery crash", and "finished PITR, up and running normally". >> A GUC is not a good way to do that. > > Does a GUC make sense to you for how to handle standby/master for > replication? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
I've submitted two patches for adding new include features to the postgresql.conf file. While not quite commit quality yet, I hope everyone will agree their reviews this week suggest both are close enough that any number of people could finish them off. Before re-opening this can of worms, I wanted people to be comfortable that we can expect them to be available as building blocks before 9.2 development is finished. Both of those came out of requests from this unification thread, and they're a helpful part of what I'd like to propose. I don't see any path forward here that still expects the recovery.conf file to function as it used to. The "make replication easy" crowd will eventually be larger than the pre-9.0 user base, if it isn't already. And they clearly want no parts of that thing. There's been over a year of arguing around how to cope with it that will satisfy everyone, so many messages I couldn't even read them all usefully in our archives and had to go here: http://postgresql.1045698.n5.nabble.com/recovery-conf-location-td2854644.html http://postgresql.1045698.n5.nabble.com/unite-recovery-conf-and-postgresql-conf-td4785717.html I don't think it's possible. What I would propose is a specification based on forced failure if there's any sign of recovery.conf, combined with the simplest migration path we can design to ease upgrades from older versions. I think we can make the transition easy enough. Users and tool vendors can make relatively simple changes to support 9.2 without changing everything they're used to just yet--while still being very clear deprecation has arrived and they should reconsider their approach. Only bug-compatible levels of backwards compatibility would make this transparent to them, and there's way too many issues to allow moving forward that way--a forward path that as far as I can see is desired by the majority of users, and just as importantly for all of the potential new users we're impacting with the current mess. There's another, perhaps under considered, concern I want to highlight as well. Tom has repeatedly noted that one of the worst problems here would go away if the "existence means do recovery" nature of recovery.conf went elsewhere. And we know some packagers want to separate the necessary to manipulate configuration files from the database directory, for permissions and management reasons. As Heikki nicely put it (over a year ago), "You don't want to give monitoring tools that decide on failover write access to the data directory". Any information that the user is supplying for the purpose of specifying things needs to be easy to relocate to a separate config file area, instead of treating it more like a control file in $PGDATA. Some chatting this morning with Simon pointed out a second related concern there, which makes ideas like "specify the path to the recovery.conf file" infeasible. The data_directory is itself a parameter, so anything tied to that or a new GUC means that config files specified there we would need two passes. First identify the data directory, then go back again to read recovery.conf from somewhere else. And nobody wants to wander that way. If it doesn't fit cleanly into the existing postgresql.conf parsing, it's gotta go. Here's the rough outline of what I think would work here: -All settings move into the postgresql.conf. -As of 9.2, relocating the postgresql.conf means there are no user writable files needed in the data directory. -Creating a standby.enabled file in the directory that houses the postgresql.conf (same logic as "include" uses to locate things) puts the system into recovery mode. That feature needs to save some state, and making those decisions based on existence of a file is already a thing we do. Rather than emulating the rename to recovery.done that happens now, the server can just delete it, to keep from incorrectly returning to a state it's exited. A UI along the lines of the promote one, allowing "pg_ctl standby", should fall out of here. I think this is enough that people who just want to use replication features need never hear about this file at all, at least during the important to simplify first run through. -If startup finds a recovery.conf file where it used to live at, abort--someone is expecting the old behavior. Hint to RTFM or include a short migration guide right on the spot. That can have a nice section about how you might use the various postgresql.conf include* features if they want to continue managing those files separately. Example: rename it as replication.conf and use include_if_exists if you want to be able to rename it to recovery.done like before. Or drop it into a conf.d directory where the rename will make it then skipped. -Tools such as pgpool that want to write a simple configuration file, only touching the things that used to go into recovery.conf, can tell people to do the same trick. End their postgresql.conf with a call to \include_if_exists replication.conf as part of setup. While I don't like pushing problems toward tool vendors, as one I think validating if this has been done doesn't require the sort of fully GUC compatible parser people (rightly) want to avoid. A simple scan of the postgresql.conf looking for the recommended text at the front of a line could confirm whether that bit is there. And adding a single "include_if_exists" line to the end of the postgresql.conf is not a terrible edit job to consider pushing toward tools. None of this is any more complicated than the little search and replace job that initdb does right now. -[Optional: for 9.2 but not forever, report incompatibilities, or potentially stop working altogether, if the things that used to go into recovery.conf show up when they aren't going to do anything. That would match the 'stop if the old way is used' theme, but may introduce its own new operational issues.] That's basically it for new ideas. The mechanics of making everything in recovery.conf turn into a GUC seemed pretty uncontroversial when I scanned the archives here. Another idea not to forget is getting the read permissions on the new GUCs correct, like making primary_conninfo superuser visible only. Seems like some of the restart error handling might still have a tricky part or two left in it. But if you've accepted that a clear break from existing behavior is happening, some of those simplify I think. If we're moving forward this way for 9.2, we do need to be careful that a spec is nailed down and people take responsibility for their respective parts of it. There's a few interlocking parts that won't work well if partially completed. I marked myself as reviewer for this patch hoping I could crack the deadlock around the specification. One open item I'm not going to do is a more serious review of the already submitted code. To step back toward the thinking that brought me to here (which has helped pull Simon a bit closer to the middle of the opinion pack too), an application facing change to PostgreSQL has serious backward compatibility requirements. It's really bad if the server runs but will quietly malfunction, such that you may not notice the change until much later when a rare code path is hit. But when we're talking about something that impacts whether the server will start or not, that can be different. You have a new option. Make sure the break is obvious and total when it happens: do not function at all if someone tries to operate the old way. The migration path toward the new one should be as painless as possible. If people only have to tweak one or two simple things to get an approximation of the old behavior, provide that, give migration instructions. But this is a new major version; if people expected all their old configurations to just move over with zero changes, that's full of fail in all directions. The best we can do is try to improve the odds their test migrations will fail early, and make the path to upgrade as simple and well defined as possible. (It can be argued that people will not actually run into the proposed "server blocks if recovery.conf exists" fail-safe during migration testing. It might hit them only when the first fail-over happens in the field. Then they've got a production server down and are staring at the new error message, at best; if they don't see the logs even that may not be facing them. I would say that someone who migrates a high-availability system over to a new major version, and doesn't read the release notes nor do the simplest of fail-over testing, is doomed regardless of what we do to try and help them.) -- Greg Smith 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.us
On Wed, Dec 14, 2011 at 4:16 PM, Greg Smith <greg@2ndquadrant.com> wrote: > [ plan for deprecating recovery.conf ] +1. I'd be very happy with this plan. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Wed, Dec 14, 2011 at 22:16, Greg Smith <greg@2ndquadrant.com> wrote: > I've submitted two patches for adding new include features to the > postgresql.conf file. While not quite commit quality yet, I hope everyone > will agree their reviews this week suggest both are close enough that any > number of people could finish them off. Before re-opening this can of > worms, I wanted people to be comfortable that we can expect them to be > available as building blocks before 9.2 development is finished. Both of > those came out of requests from this unification thread, and they're a > helpful part of what I'd like to propose. > > I don't see any path forward here that still expects the recovery.conf file > to function as it used to. The "make replication easy" crowd will > eventually be larger than the pre-9.0 user base, if it isn't already. And > they clearly want no parts of that thing. There's been over a year of > arguing around how to cope with it that will satisfy everyone, so many > messages I couldn't even read them all usefully in our archives and had to > go here: > > http://postgresql.1045698.n5.nabble.com/recovery-conf-location-td2854644.html > http://postgresql.1045698.n5.nabble.com/unite-recovery-conf-and-postgresql-conf-td4785717.html > > I don't think it's possible. What I would propose is a specification based > on forced failure if there's any sign of recovery.conf, combined with the > simplest migration path we can design to ease upgrades from older versions. > I think we can make the transition easy enough. Users and tool vendors can > make relatively simple changes to support 9.2 without changing everything > they're used to just yet--while still being very clear deprecation has > arrived and they should reconsider their approach. Only bug-compatible > levels of backwards compatibility would make this transparent to them, and > there's way too many issues to allow moving forward that way--a forward path > that as far as I can see is desired by the majority of users, and just as > importantly for all of the potential new users we're impacting with the > current mess. > > There's another, perhaps under considered, concern I want to highlight as > well. Tom has repeatedly noted that one of the worst problems here would go > away if the "existence means do recovery" nature of recovery.conf went > elsewhere. And we know some packagers want to separate the necessary to > manipulate configuration files from the database directory, for permissions > and management reasons. As Heikki nicely put it (over a year ago), "You > don't want to give monitoring tools that decide on failover write access to > the data directory". Any information that the user is supplying for the > purpose of specifying things needs to be easy to relocate to a separate > config file area, instead of treating it more like a control file in > $PGDATA. Some chatting this morning with Simon pointed out a second related > concern there, which makes ideas like "specify the path to the recovery.conf > file" infeasible. The data_directory is itself a parameter, so anything > tied to that or a new GUC means that config files specified there we would > need two passes. First identify the data directory, then go back again to > read recovery.conf from somewhere else. And nobody wants to wander that > way. If it doesn't fit cleanly into the existing postgresql.conf parsing, > it's gotta go. > > Here's the rough outline of what I think would work here: > > -All settings move into the postgresql.conf. > > -As of 9.2, relocating the postgresql.conf means there are no user writable > files needed in the data directory. > > -Creating a standby.enabled file in the directory that houses the > postgresql.conf (same logic as "include" uses to locate things) puts the > system into recovery mode. That feature needs to save some state, and > making those decisions based on existence of a file is already a thing we > do. Rather than emulating the rename to recovery.done that happens now, the > server can just delete it, to keep from incorrectly returning to a state > it's exited. A UI along the lines of the promote one, allowing "pg_ctl > standby", should fall out of here. I think this is enough that people who > just want to use replication features need never hear about this file at > all, at least during the important to simplify first run through. You say "the server can just delete it". But how will this work if the file is *not* in the data directory? If you are on a Debian style system for example, where all these files go in /etc/postgresql - wouldn't that suddenly require the postgres user to have write access in this directory? If it actually has to be the server that modifies the file, I think it *does* make sense for this file to live in the data directory... [cutting lots of good explanations] Other than that consideration, +1 for this proposal. -- Magnus Hagander Me: http://www.hagander.net/ Work: http://www.redpill-linpro.com/
On 12/14/11 1:16 PM, Greg Smith wrote: > > -Creating a standby.enabled file in the directory that houses the > postgresql.conf (same logic as "include" uses to locate things) puts the > system into recovery mode. That feature needs to save some state, and > making those decisions based on existence of a file is already a thing > we do. Rather than emulating the rename to recovery.done that happens > now, the server can just delete it, to keep from incorrectly returning > to a state it's exited. A UI along the lines of the promote one, > allowing "pg_ctl standby", should fall out of here. I think this is > enough that people who just want to use replication features need never > hear about this file at all, at least during the important to simplify > first run through. How will things work for PITR? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On 12/14/2011 04:47 PM, Magnus Hagander wrote: > You say "the server can just delete it". But how will this work if the > file is *not* in the data directory? If you are on a Debian style > system for example, where all these files go in /etc/postgresql - > wouldn't that suddenly require the postgres user to have write access > in this directory? If it actually has to be the server that modifies > the file, I think it *does* make sense for this file to live in the > data directory... > Perhaps I should have softened the suggestion to "relocating the postgresql.conf makes it *possible* to have no user writable files in the data directory". That was one of the later additions I made, it didn't bake overnight before sending like the bulk did. A Debian system might want it to stay in the data directory. If we consider this not normally touched by the user state information--they can poke it by hand, but the preferred way is to use pg_ctl--perhaps it could live in /var/run/postgresql instead. [Surely I'll find out momentarily, now that I've trolled everyone here who is more familiar than me with the rules around what goes into /var] I think the bigger idea I was trying to support in this part is just how many benefits there are from breaking this role into one decoupled from the main server configuration. It's not a new suggestion, but I think it was cut down by backward compatibility concerns before being fully explored. It seems all of the good ways to provide cleaner UIs need that, and it surely gives better flexibility to packagers for it to float free from the config. Who can predict what people will end up doing in their packages. (And the Gentoo changes have proven it's not just Debian) If we drag this conversation back toward the best way to provide that cleaner UI, but can pick up enough agreement that backward compatibility limited to the sort of migration ideas I outlined is acceptable, I'd be happy with that progress. Hopes of reaching that point is the reason I dumped time into those alternative include options. -- Greg Smith 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.us
On 12/14/2011 04:57 PM, Josh Berkus wrote: > How will things work for PITR? Left that out mainly because I was already running too long there, but I do think there's a reasonable path. There is one additional wrinkle in there to consider that I've thought of so far, falling into the "what is the best UI for this?" category I think. Put the stuff you used to insert into recovery.conf into postgresql.conf instead. If you don't like that, use another file and include it with one of the multiple options for that--same migration option I already suggested. Run "pg_ctl recovery"; under the hood that's actually creating standby.enabled instead of recovery.conf, but you don't have to know that. You'd suggested renaming it to reflect its most common usage now, and I thought that was quite sensible. It helps with the "things have changed, please drive carefully" feel too. It seems possible to have two files for state kickoff/tracking here instead, maybe have recovery.enabled and standby.enabled. Is that extra complexity a useful thing? I haven't dug into that new topic much yet. (Look at that! I think I just found a *new* topic here!) There are some questions around what to do when it's done. The new proposed behavior is to delete the standby.enabled file. But that doesn't remove the changes made for recovery like the old recovery.done rename did. This is why I commented that some more thinking is likely needed about how to handle seeing those only-makes-sense-in-recovery options when not being started for recovery/standby; it's not obvious that any approach will make everyone happy. If you want to do something special yourself to clean that up, there's already recovery_end_command available for that. Let's say you wanted to force the old name and did "include_if_exists conf.d/recovery.conf", to trick it even if the patrolling for the name idea goes in. Now you could do: recovery_end_command = 'rm -f /tmp/pgsql.trigger.5432 && mv conf.d/recovery.conf conf.d/recovery.done' Like some people are used to and might still prefer for some reason. There'd be time left over to head out to the lawn and yell at the kids there. Actually, this might be the right approach for tools that are trying to change as little as possible but add quick 9.2 compatibility. I think there's enough pluggable bits in every direction here that people can assemble the setup they'd like out of the available parts, Maybe these slightly different semantics between archive recovery and standby mode are exactly why they should be kicked off by differently named files? -- Greg Smith 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.us
Greg, > Put the stuff you used to insert into recovery.conf into postgresql.conf > instead. If you don't like that, use another file and include it with > one of the multiple options for that--same migration option I already > suggested. Run "pg_ctl recovery"; under the hood that's actually > creating standby.enabled instead of recovery.conf, but you don't have to > know that. You'd suggested renaming it to reflect its most common usage > now, and I thought that was quite sensible. It helps with the "things > have changed, please drive carefully" feel too. So for streaming replication, will I need to have a standby.enabled file, or will there be a parameter in postgresql.conf (or included files) which controls whether or not that server is a standby, available? In the best of all possible worlds, I'd really like to have a GUC which 100% controls whether or not the current server is a standby. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On 12/14/2011 08:56 PM, Josh Berkus wrote: > So for streaming replication, will I need to have a standby.enabled > file, or will there be a parameter in postgresql.conf (or included > files) which controls whether or not that server is a standby, available? > > In the best of all possible worlds, I'd really like to have a GUC which > 100% controls whether or not the current server is a standby You keep asking the hard questions. Right now, I kind of like that it's possible to copy a postgresql.conf file from master to standby and just use it. That should still be possible with the realignment into GUCs: -"standby_mode = on": Ignored unless you've started the server with standby.enabled, won't bother the master if you include it. -"primary_conninfo": This will look funny on the master showing it connecting to itself, but it will get ignored there too. I was hoping to just copy over a base backup, "pg_ctl standby" creates the needed file and starts the server, and I'm done. Isn't that the easiest replication "hello, world" possible here? If you think there's an easier way here, please describe it more; I'm missing it so far. Some settings will look a bit weird in the identical postgresql.conf in this case, but it think it can be made to work. Now, eventually you will have to sort this out, but my normal principle here is that any issue deferred until after people have a working system is automatically easier for them to stomach. Yes there's complexity, but people are facing it after the happy dance when the standby works for the first time. The unavoidable bad situation happens if you promote a standby made this way. Replicating more standbys from it won't work; you have to fix primary_conninfo at some point. But once you're the master, you should be able to change primary_conninfo anytime--even if you SIGHUP to reload, it will now be ignored--so sorting that out doesn't even require a server restart. [Problem of how exactly to define a GUC with those properties while also doing the right thing when you are a standby was noted then snuck by quietly] If that is replaced with an edit to the postgresql.conf, that makes the bar for setting up a standby higher in my mind. Now we have every clusterware product forced into the position pgpool already finds itself, where it needs to cope with making at least one change to that file. I can see a middle ground position where you can have the standby.enabled file, but you can also set something in the postgresql.conf, but now we're back to conflict and order resolution madness. [See: "which of postgresql.conf and recovery.conf should be read first?"] [Crazy talk begins here, but without further abuse of parenthetical brackets] There is a route this way I wouldn't mind wandering down, but it circles back to one of the even bigger debates. I would be perfectly happy fully embracing multiple configuration files for the server by default on every install. Then the things that vary depending on current role can all be put into one place, with some expected splits along this line. Put all the stuff related to standby configuration in one file; then tools can know "I can overwrite this whole file" and that will be true. There's an obvious objection here that "having this crap in two files is the problem we're trying to eliminate!" I would still see this as forward, because at the very minimum that split should refactor the replication and recovery target pieces into different files. Different tools will want to feel they "own" them and can rewrite them, and making that easy should be a major goal. Also, it will be possible to rearrange them if you'd like in whatever order makes sense, which you can't do now for the recovery.conf part. You'd just be breaking tools that might expect the default split doing that; if you don't care, have at it. Wandering any distance down that whole road surely stretches the premise of "simple migration procedure using include" too far to be true anymore. I was thinking that for 9.2, it seems feasible to get much of this legacy stuff sorted better (from the perspective of the person focused on simple replication), and add some enabling features. No recovery.conf, everything is a GUC, migration path isn't so bad, people get exposed to new concepts for include file organization. I'd like to do some bigger reorganization too, but that seems too much change to shove all into one release. There's a simple path from there that leads to both easier tools all around and SET PERSISTENT, and it comes with a pile of disruption so big I could throw in "standby controls are now 100% GUC" for you plus a unicorn and it would slip right by unnoticed. That's a tough roadmap to sell unless those promised benefits are proven first though. And I'm thinking a release doing all that is going to want to be named 10.0--and what I could really use is a nice, solid 9.2 that doesn't scare enterprises with too much change next. -- Greg Smith 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.us
On Wed, Dec 14, 2011 at 8:56 PM, Josh Berkus <josh@agliodbs.com> wrote: > So for streaming replication, will I need to have a standby.enabled > file, or will there be a parameter in postgresql.conf (or included > files) which controls whether or not that server is a standby, available? > > In the best of all possible worlds, I'd really like to have a GUC which > 100% controls whether or not the current server is a standby. I think that would be a bad idea, because then how would pg_ctl promote work? It'd have to permanently change the value of that GUC, which means rewriting a postgresql.conf file with an arbitrary forest of comments and include files, and we can't do that now or, probably, ever. At least not reliably enough for this kind of use case. I think what Greg is going for is this: 1. Get rid of recovery.conf - error out if it is seen 2. For each parameter that was previously a recovery.conf parameter, make it a GUC 3. For the parameter that was "does recovery.conf exist?", replace it with "does standby.enabled exist?". IMHO, as Greg says, that's as much change as we can cope with in one release. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
Greg, > You keep asking the hard questions. I practice. ;-) > Right now, I kind of like that it's > possible to copy a postgresql.conf file from master to standby and just > use it. That should still be possible with the realignment into GUCs: ... long discussion omitted here. I agree that GUC vs. standby.enabled is a trade-off. I further agree that where we're going with this eventually is SET PERSISTENT. I feel that Greg's proposal is a substantial improvement on the current arrangement and eliminates *my* #1 source of replication-configuration pain, and we can keep improving it later. I think we need to give some thought as to how this will play out for PITR, since there is far less reason to change the operation of PITR, and much older backup tools which rely on its current operation. Otherwise, +1. > shove all into one release. There's a simple path from there that leads > to both easier tools all around and SET PERSISTENT, and it comes with a > pile of disruption so big I could throw in "standby controls are now > 100% GUC" for you plus a unicorn and it would slip right by unnoticed. > That's a tough roadmap to sell unless those promised benefits are proven > first though. And I'm thinking a release doing all that is going to > want to be named 10.0--and what I could really use is a nice, solid 9.2 > that doesn't scare enterprises with too much change next. I would love to see a writeup on this someday. Blog? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
All, I'll point out that this patch got sandbagged to death, and never made it into 9.2. So, for 9.2 replication is just as hard to configure and manage as it was in 9.1. Are we going to fix it in 9.3, or not? -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On Wed, May 09, 2012 at 08:07:52PM -0700, Josh Berkus wrote: > All, > > I'll point out that this patch got sandbagged to death, and never made > it into 9.2. So, for 9.2 replication is just as hard to configure and > manage as it was in 9.1. Are we going to fix it in 9.3, or not? Greg Smith was going to allow for files in configuration directories, and that was somehow going to make it easier to manage the configuration files and remove recovery.conf. I don't think Greg did it; I didn't see it in the release notes I just wrote. -- Bruce Momjian <bruce@momjian.us> http://momjian.us EnterpriseDB http://enterprisedb.com + It's impossible for everything to be true. +
On 05/09/2012 11:15 PM, Bruce Momjian wrote: > On Wed, May 09, 2012 at 08:07:52PM -0700, Josh Berkus wrote: >> All, >> >> I'll point out that this patch got sandbagged to death, and never made >> it into 9.2. So, for 9.2 replication is just as hard to configure and >> manage as it was in 9.1. Are we going to fix it in 9.3, or not? > > Greg Smith was going to allow for files in configuration directories, > and that was somehow going to make it easier to manage the configuration > files and remove recovery.conf. I don't think Greg did it; I didn't > see it in the release notes I just wrote. That was actually submitted back in November, and rightly kicked back as needing more work. I would have updated it and resubmitted if that was the only blocker. But by the time January rolled around, it was already obvious that the last CommitFest was going into overtime. Didn't seem like a great time to add a disruptive change like this one into the mix. I expect to revisit config directories before the first 9.3 CF, it will help multiple things I'd like to see happen. Then we can circle back to the main unification job with a fairly clear path forward from there. -- Greg Smith 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.com
> I expect to revisit config directories before the first 9.3 CF, it will > help multiple things I'd like to see happen. Then we can circle back to > the main unification job with a fairly clear path forward from there. Yeah, let's discuss this next week. "Easier configuration" is one demand I'm hearing from developers in general, and I don't think that's nearly as hard a feature as, say, parallel query. We can do it. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com
On 10 May 2012 05:44, Josh Berkus <josh@agliodbs.com> wrote: > >> I expect to revisit config directories before the first 9.3 CF, it will >> help multiple things I'd like to see happen. Then we can circle back to >> the main unification job with a fairly clear path forward from there. > > Yeah, let's discuss this next week. "Easier configuration" is one > demand I'm hearing from developers in general, and I don't think that's > nearly as hard a feature as, say, parallel query. We can do it. A key requirement is to be able to drop in new config files without needing to $EDIT anything. OK, its possible to put in lots of includeifexists for non-existent files just in case you need one, but that sucks. -- Simon Riggs http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
> A key requirement is to be able to drop in new config files without > needing to $EDIT anything. Yes, absolutely. I want to move towards the idea that the majority of our users never edit postgresql.conf by hand. > OK, its possible to put in lots of includeifexists for non-existent > files just in case you need one, but that sucks. Yeah, seems like we need something more elegant. -- Josh Berkus PostgreSQL Experts Inc. http://pgexperts.com