Hi,
On 2019-05-23 17:39:21 +0530, Amit Khandekar wrote:
> On Tue, 21 May 2019 at 21:49, Andres Freund <andres@anarazel.de> wrote:
> Yeah, I agree we should add such checks to minimize the possibility of
> reading logical records from a master that has insufficient wal_level.
> So to summarize :
> a. CheckLogicalDecodingRequirements() : Add Controlfile wal_level checks
> b. Call this function call in CreateInitDecodingContext() as well.
> c. While decoding XLOG_PARAMETER_CHANGE record, emit recovery conflict
> error if there is an existing logical slot.
>
> This made me think more of the race conditions. For instance, in
> pg_create_logical_replication_slot(), just after
> CheckLogicalDecodingRequirements and before actually creating the
> slot, suppose concurrently Controlfile->wal_level is changed from
> logical to replica. So suppose a new slot does get created. Later the
> slot is read, so in pg_logical_slot_get_changes_guts(),
> CheckLogicalDecodingRequirements() is called where it checks
> ControlFile->wal_level value. But just before it does that,
> ControlFile->wal_level concurrently changes back to logical, because
> of replay of another param-change record. So this logical reader will
> think that the wal_level is sufficient, and will proceed to read the
> records, but those records are *before* the wal_level change, so these
> records don't have logical data.
I don't think that's an actual problem, because there's no decoding
before the slot exists and CreateInitDecodingContext() has determined
the start LSN. And by that point the slot exists, slo
XLOG_PARAMETER_CHANGE replay can error out.
Greetings,
Andres Freund