> On Tue, Feb 5, 2019 at 12:54 AM Alvaro Herrera <alvherre@2ndquadrant.com> wrote:
>
> Actually, index-based replica identities failed in pg_dump: we first
> create the index ONLY on the partitioned table (marking it as invalid),
> so when we immediately do the ALTER TABLE/REPLICA IDENTITY, it rejects
> the invalid index. If we change the code to allow invalid indexes on
> partitioned tables to become replica identities, we hit the problem that
> the index didn't exist when processing the partition list. In order to
> fix that, I added a flag so that partitions are allowed not to have the
> index, in hopes that the missing index are created in subsequent
> commands; those indexes should become valid & identity afterwards.
>
> There's a small emerging problem, which is that if you have an invalid
> index marked as replica identity, you cannot create any more partitions;
> the reason is that we want to propagate the replica identity to the
> partition, but the index is not there (since invalid indexes are not
> propagated). I don't think this case is worth supporting; it can be
> fixed but requires some work[1].
>
> New patch attached.
Could there be a race condition somewhere? The idea and the code looks
reasonable, but when I'm testing ALTER TABLE ... REPLICA IDENTITY with this
patch and concurrent partition creation, I've got the following error on ATTACH
PARTITION:
ERROR: 42P17: replica index does not exist in partition "test373"
LOCATION: MatchReplicaIdentity, tablecmds.c:15018
This error seems unstable, other time I've got a deadlock. And I don't observe
this behaviour on the master.