Re: Minimal logical decoding on standbys

On Mon, Apr 3, 2023 at 4:26 AM Jeff Davis <pgsql(at)j-davis(dot)com> wrote:
>
> On Fri, 2023-03-31 at 02:50 -0700, Jeff Davis wrote:
> > But if the ConditionVariableEventSleep() API is added, then I think
> > we
> > should change the non-recovery case to use a CV as well for
> > consistency, and it would avoid the need for WalSndWakeup().
>
> It seems like what we ultimately want is for WalSndWakeup() to
> selectively wake up physical and/or logical walsenders depending on the
> caller. For instance:
>
> WalSndWakeup(bool physical, bool logical)
>
> The callers:
>
> * On promotion, StartupXLog would call:
> - WalSndWakeup(true, true)
> * XLogFlush/XLogBackgroundFlush/XLogWalRcvFlush would call:
> - WalSndWakeup(true, !RecoveryInProgress())
> * ApplyWalRecord would call:
> - WalSndWakeup(switchedTLI, switchedTLI || RecoveryInProgress())
>
> There seem to be two approaches to making that work:
>
> 1. Use two ConditionVariables, and WalSndWakeup would broadcast to one
> or both depending on its arguments.
>
> 2. Have a "replicaiton_kind" variable in WalSnd (either set based on
> MyDatabaseId==InvalidOid, or set at START_REPLICATION time) to indicate
> whether it's a physical or logical walsender. WalSndWakeup would wake
> up the right walsenders based on its arguments.
>
> #2 seems simpler at least for now. Would that work?
>

Agreed, even Bertrand and myself discussed the same approach few
emails above. BTW, if we have this selective logic to wake
physical/logical walsenders and for standby's, we only wake logical
walsenders at the time of ApplyWalRecord() then do we need the new
conditional variable enhancement being discussed, and if so, why?

--
With Regards,
Amit Kapila.