On 22 November 2016 at 17:49, Kyotaro HORIGUCHI
<horiguchi.kyotaro@lab.ntt.co.jp> wrote:
>> > Yeah, I definitely don't think it's as simple as just using
>> > procsignal_sigusr1_handler as-is. I expect we'd likely create a new
>> > global IsWalSender and ignore some RecoveryConflictInterrupt cases
>> > when in a walsender, at least PROCSIG_RECOVERY_CONFLICT_SNAPSHOT, and
>> > probably add a new case for catalog_xmin conflicts that's only acted
>> > on when IsWalSender.
>>
>> The global is unncecessary if walsender have a different handler
>> from normal backends. If there's at least one or few additional
>> reasons for signal, sharing SendProcSignal and having dedicate
>> handler might be better.
>
> If no behavior is shared among normal backend and walsender, it
> would be a good reason not to share the handler function. What
> you are willing to do seems so.
I've explored this some more, and it looks like using
procsignal_sigusr1_handler for handling recovery conflicts in the
walsender during logical decoding actually makes a lot of sense.
Almost all behaviour is shared, and so far I haven't needed any
special cases at all. I needed to add a new recovery signal for
conflict with catalog_xmin advance on upstream, but that was it.
Many of the cases make no sense for physical walsenders, so it
probably makes sense to bail out early if it's a physical walsender,
but for a walsender doing logical replication the only one that I
don't think makes sense is conflict with snapshot, which won't get
sent and is harmless if received.
(The comment on it is slightly wrong anyway; it claims it's only used
by normal user backends in transactions, but database conflicts are
fired even when not in an xact.)
-- Craig Ringer http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services