Re: PATCH: logical_work_mem and logical streaming of largein-progress transactions

On Thu, Dec 12, 2019 at 3:41 AM Amit Kapila <amit.kapila16@gmail.com> wrote:
> I don't think we have evaluated it yet, but we should do it.  The
> point to note is that it is only for the case when wal_level is
> 'logical' (see IsSubTransactionAssignmentPending) in which case we
> already log more WAL, so this might not impact much.  I guess that it
> might be better to have that check in XLogRecordAssemble for the sake
> of clarity.

I don't think that this is really a valid argument. Just because we
have some overhead now doesn't mean that adding more won't hurt. Even
testing the wal_level costs a little something.

> I think the way invalidations work for logical replication is that
> normally, we always start a new transaction before decoding each
> commit which allows us to accept the invalidations (via
> AtStart_Cache).  However, if there are catalog changes within the
> transaction being decoded, we need to reflect those before trying to
> decode the WAL of operation which happened after that catalog change.
> As we are not logging the WAL for each invalidation, we need to
> execute all the invalidation messages for this transaction at each
> catalog change. We are able to do that now as we decode the entire WAL
> for a transaction only once we get the commit's WAL which contains all
> the invalidation messages.  So, we queue them up and execute them for
> each catalog change which we identify by WAL record
> XLOG_HEAP2_NEW_CID.

Thanks for the explanation. That makes sense. But, it's still true,
AFAICS, that instead of doing this stuff with logging invalidations
you could just InvalidateSystemCaches() in the cases where you are
currently applying all of the transaction's invalidations. That
approach might be worse than changing the way invalidations are
logged, but the two approaches deserve to be compared. One approach
has more CPU overhead and the other has more WAL overhead, so it's a
little hard to compare them, but it seems worth mulling over.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company