On Sat, Nov 20, 2021 7:31 PM Amit Kapila <[email protected]> wrote: > On Fri, Nov 19, 2021 at 10:58 AM Amit Kapila <[email protected]> > wrote: > > > > On Fri, Nov 19, 2021 at 7:19 AM Amit Langote <[email protected]> > wrote: > > > > > > The problematic case is attaching the partition *after* the > > > subscriber has already marked the root parent as synced and/or ready > > > for replication. Refreshing the subscription doesn't help it > > > discover the newly attached partition, because a > > > publish_via_partition_root only ever tells about the root parent, > > > which would be already synced, so the subscriber would think there's > > > nothing to copy. > > > > > > > Okay, I see this could be a problem but I haven't tried to reproduce it. > > One more thing you mentioned is that the initial sync won't work after refresh > but later changes will be replicated but I noticed that later changes also > don't > get streamed till we restart the subscriber server. > I am not sure but we might not be invalidating apply workers cache due to > which it didn't notice the same.
I investigated this bug recently, and I think the reason is that when receiving
relcache invalidation message, the callback function[1] in walsender only reset
the schema sent status while it doesn't reset the replicate_valid flag. So, it
won’t rebuild the publication actions of the relation.
[1]
static void
rel_sync_cache_relation_cb(Datum arg, Oid relid)
...
/*
* Reset schema sent status as the relation definition may have changed.
* Also free any objects that depended on the earlier definition.
*/
if (entry != NULL)
{
entry->schema_sent = false;
list_free(entry->streamed_txns);
...
Also, when you DETACH a partition, the publication won’t be rebuilt too because
of the same reason. Which could cause unexpected behavior if we modify the
detached table's data . And the bug happens regardless of whether pubviaroot is
set or not.
For the fix:
I think if we also reset replicate_valid flag in rel_sync_cache_relation_cb,
then the bug can be fixed. I have a bit hesitation about this approach, because
it could increase the frequency of invalidating and rebuilding the publication
action. But I haven't produced some other better approaches.
Another possibility could be that we add a syscache callback function for
pg_inherits table, but we currently don't have syscache for pg_inherits. We
might need to add the cache pg_inherits first which doesn't seems better
than the above approach.
What do you think ?
Attach an initial patch which reset the replicate_valid flag in
rel_sync_cache_relation_cb and add some reproduction tests in 013_partition.pl.
Best regards,
Hou zj
0001-fix-replication-after-ATTACH-or-DETACH-partition.patch
Description: 0001-fix-replication-after-ATTACH-or-DETACH-partition.patch
