On Tue, Sep 09, 2025 at 11:58:51AM +0800, 邱宇航 wrote:
>> Oops. When redo XLOG_CHECKPOINT_SHUTDOWN, smgrdestroyall should also be
>> called, since the startup may not exit on standby.
>>
>> The patch is updated.
True that the situation sucks for the startup process, bloating its
memory. That's hard to reach, still for long-running startup
processes, which is a common thing, that's rather bad.
> LGTM.
Hmm. I was playing a bit with the startup process and, after planting
a few calls to hash_get_num_entries(SMgrRelationHash) the bloat is
measurable. On wraparound, it would mean that the hash table could
point to past entries in this context.
I can get behind the patch and the proposal of forcing a cleanup each
time a checkpoint record is replayed, outside of
RecoveryRestartPoint(), so I'll see about applying and backpatching
that. Thanks for the report.
--
Michael
