Re: logical decoding of two-phase transactions - Mailing list pgsql-hackers

> On 28 Mar 2017, at 18:08, Andres Freund <andres@anarazel.de> wrote:
>
> On 2017-03-28 15:55:15 +0100, Simon Riggs wrote:
>>
>>
>> That assertion is obviously false... the plugin can resolve this in
>> various ways, if we allow it.
>
> Handling it by breaking replication isn't handling it (e.g. timeouts in
> decoding etc).  Handling it by rolling back *prepared* transactions
> (which are supposed to be guaranteed to succeed!), isn't either.
>
>
>> You can say that in your opinion you prefer to see this handled in
>> some higher level way, though it would be good to hear why and how.
>
> It's pretty obvious why: A bit of DDL by the user shouldn't lead to the
> issues mentioned above.
>
>
>> Bottom line here is we shouldn't reject this patch on this point,
>
> I think it definitely has to be rejected because of that.  And I didn't
> bring this up at the last minute, I repeatedly brought it up before.
> Both to Craig and Stas.

Okay. In order to find more realistic cases that blocks replication
i’ve created following setup:

* in backend: tests_decoding plugins hooks on xact events and utility
statement hooks and transform each commit into prepare, then sleeps
on latch. If transaction contains DDL that whole statement pushed in
wal as transactional message. If DDL can not be prepared or disallows
execution in transaction block than it goes as nontransactional logical
message without prepare/decode injection. If transaction didn’t issued any
DDL and didn’t write anything to wal, then it skips 2pc too.

* after prepare is decoded, output plugin in walsender unlocks backend
allowing to proceed with commit prepared. So in case when decoding
tries to access blocked catalog everything should stop.

* small python script that consumes decoded wal from walsender (thanks
Craig and Petr)

After small acrobatics with that hooks I’ve managed to run whole
regression suite in parallel mode through such setup of test_decoding
without any deadlocks. I’ve added two xact_events to postgres and
allowedn prepare of transactions that touched temp tables since
they are heavily used in tests and creates a lot of noise in diffs.

So it boils down to 3 failed regression tests out of 177, namely:

* transactions.sql — here commit of tx stucks with obtaining SafeSnapshot().
I didn’t look what is happening there specifically, but just checked that
walsender isn’t blocked. I’m going to look more closely at this.

* prepared_xacts.sql — here select prepared_xacts() sees our prepared
tx. It is possible to filter them out, but obviously it works as expected.

* guc.sql — here pendingActions arrives on 'DISCARD ALL’ preventing tx
from being prepared. I didn’t found the way to check presence of
pendingActions outside of async.c so decided to leave it as is.

It seems that at least in regression tests nothing can block twophase
logical decoding. Is that strong enough argument to hypothesis that current
approach doesn’t creates deadlock except locks on catalog which should be
disallowed anyway?

Patches attached. logical_twophase_v5 is slightly modified version of previous
patch merged with Craig’s changes. Second file is set of patches over previous
one, that implements logic i’ve just described. There is runtest.sh script that
setups postgres, runs python logical consumer in background and starts
regression test.


Stas Kelvich
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company