On 12/15/21 14:51, Tomas Vondra wrote:
On 12/15/21 14:20, Amit Kapila wrote:
On Tue, Dec 14, 2021 at 7:02 AM Tomas Vondra
<tomas.von...@enterprisedb.com> wrote:
Hi,
here's an updated version of the patches, dealing with almost all of the
issues (at least in the 0001 and 0002 parts). The main changes:
1) I've removed the 'created' flag from fill_seq_with_data, as
discussed. I don't think it's needed by any of the parts (not even 0003,
AFAICS). We still need it in xl_seq_rec, though.
2) GetCurrentTransactionId() added to sequence.c are called only with
wal_level=logical, to minimize the overhead.
There's still one remaining problem, that I already explained in [1].
The problem is that with this:
BEGIN;
SELECT nextval('s') FROM generate_series(1,100);
ROLLBACK;
The root cause is that pg_current_wal_lsn() uses the LogwrtResult.Write,
which is updated by XLogFlush() - but only in RecordTransactionCommit.
Which makes sense, because only the committed stuff is "visible".
But the non-transactional behavior of sequence decoding disagrees with
this, because now some of the changes from aborted transactions may be
replicated. Which means the wait_for_catchup() ends up not waiting for
the sequence change to be replicated. This is an issue for tests in
patch 0003, at least.
My concern is this actually affects other places waiting for things
getting replicated :-/
By any chance, will this impact synchronous replication as well which
waits for commits to be replicated?
Physical or logical replication? Physical is certainly not replicated.
Actually, I take that back. It does affect physical (sync) replication
just as well, and I think it might be considered a data loss issue. I
started a new thread to discuss that, so that it's not buried in this
thread about logical replication.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
