Re: Logical decoding on standby

On 20 March 2017 at 14:57, Simon Riggs <simon(dot)riggs(at)2ndquadrant(dot)com> wrote:

> 2.1 Why does call to ReplicationSlotAcquire() move earlier in
> pg_logical_slot_get_changes_guts()?

That appears to be an oversight from an earlier version where it
looped over timelines in pg_logical_slot_get_changes_guts . Reverted.

> 2.2 sendTimeLineIsHistoric looks incorrect, and at least isn't really
> documented well.
> The setting
> sendTimeLineIsHistoric = state->currTLI == ThisTimeLineID;
> should be
> sendTimeLineIsHistoric = (state->currTLI != ThisTimeLineID);

Definitely wrong. Fixed.

> but that doesn't cause failure because in read_local_xlog_page() we
> say that we are reading from history when
> state->currTLI != ThisTimeLineID explicitly rather than use
> sendTimeLineIsHistoric

XLogRead(...), as called by logical_read_xlog_page, does test it. It's
part of the walsender-local log read callback. We don't hit
read_local_xlog_page at all when we're doing walsender based logical
decoding.

We have two parallel code paths for reading xlogs, one for walsender,
one for normal backends. The walsender one is glued together with a
bunch of globals that pass state "around" the xlogreader - we set it
up before calling into xlogreader, and then examine it when xlogreader
calls back into walsender.c with logical_read_xlog_page.

I really want to refactor that at some stage, getting rid of the use
of walsender globals for timeline state tracking and sharing more of
the xlog reading logic between walsender and normal backends. But
-ENOTIME, especially to do it as carefully as it must be done.

There are comments on read_local_xlog_page, logical_read_xlog_page
that mention this. Also XLogRead in
src/backend/access/transam/xlogutils.c (which has the same name as
XLogRead in src/backend/replication/walsender.c). I have a draft for a
timeline following readme that would address some of this but don't
expect to be able to finish it off for this release cycle, and I'd
really rather clean it up instead.

--
Craig Ringer http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services