Thread

Re: Bound memory usage during manual slot sync retries

Xuneng Zhou <xunengzhou@gmail.com> — 2026-05-16T11:35:04Z

On Fri, May 15, 2026 at 9:20 PM Xuneng Zhou <xunengzhou@gmail.com> wrote:
>
> Hi Amit,
>
> On Fri, May 15, 2026 at 5:23 PM Amit Kapila <amit.kapila16@gmail.com> wrote:
> >
> > On Fri, May 15, 2026 at 11:02 AM Xuneng Zhou <xunengzhou@gmail.com> wrote:
> > >
> > > pg_sync_replication_slots() now retries inside a single SQL function
> > > call.  Unlike the slotsync worker, it does not get a transaction boundary
> > > between retry cycles, so allocations made while fetching and synchronizing
> > > remote slots can accumulate until the function returns.
> > >
> > > The existing list_free_deep(remote_slots) is not enough to bound this.
> > > It frees the List cells and the RemoteSlot structs stored as list
> > > elements, but it does not recursively free allocations owned by those
> > > structs, such as the copied slot name, plugin name, and database name.
> > >
> >
> > Right.
> >
> > > It also does not release unrelated per-cycle allocations made while
> > > fetching and processing the remote slots.
> > >
> >
> > BTW, did you notice via test or otherwise, what and how much other
> > unrelated memory is being allocated during each sync cycle if we
> > manually free the allocations you observed? This is mainly to learn
> > the impact of not doing all these allocations in some short-duration
> > memory context.
>
> I noticed this by reading the feature code while walking through the
> PG 19 release notes, not by observing actual memory growth in a
> running system. Besides the RemoteSlot field strings, there seems to
> be a few smaller per-cycle allocations that also accumulate:
> quote_literal_cstr() strings used to build the filtered query, a
> temporary TextDatumGetCString() result for invalidation_reason, the
> standalone TupleTableSlot in fetch_remote_slots(), and the list
> container built by get_local_synced_slots(). I chose a per-cycle
> memory context over manual pfrees to make the retry-cycle lifetime
> explicit and avoid maintaining a destructor for every current and
> future allocation in this path. It may be worth measuring how much
> extra memory actually accumulates during an extended wait to confirm
> the practical impact.

I did some measurements for the memory growth in the manual
pg_sync_replication_slots() retry path.

The test used 100 failover logical slots and forced
pg_sync_replication_slots() to keep retrying on the standby. Memory
was sampled with pg_log_backend_memory_contexts() on the backend
running the function.

On HEAD, the short run showed:
timestamp total_bytes delta
2026-05-16T10:47:54Z,63422,1339920,
2026-05-16T10:47:56Z,63422,1405456,65536
2026-05-16T10:47:59Z,63422,1405456,0
2026-05-16T10:48:01Z,63422,1405456,0
2026-05-16T10:48:03Z,63422,1536528,131072
2026-05-16T10:48:05Z,63422,1536528,0

So the total increase was about 192 KiB.

After adding a targeted cleanup for the copied RemoteSlot strings, the
same test showed:

timestamp total_bytes delta
2026-05-16T11:04:58Z,77000,1339920,
2026-05-16T11:05:00Z,77000,1339920,0
2026-05-16T11:05:02Z,77000,1405456,65536
2026-05-16T11:05:04Z,77000,1405456,0

So the increase dropped to about 64 KiB.

With a per-retry memory context around the fetch/synchronize cycle,
the same test stayed flat:

2026-05-16T11:09:10Z,84600,1315344,
2026-05-16T11:09:12Z,84600,1315344,0
2026-05-16T11:09:14Z,84600,1315344,0
2026-05-16T11:09:16Z,84600,1315344,0
2026-05-16T11:09:18Z,84600,1315344,0
2026-05-16T11:09:20Z,84600,1315344,0

Looking at the memory-context dumps, the growth on HEAD was visible
under ExprContext. The grand-total increase matched the ExprContex
increase, which is consistent with retry-cycle allocations surviving
for the duration of the single SQL function call.

That said, the practical severity looks small. This is mainly because
the retry loop is not a tight loop. With no slot activity,
wait_for_slot_activity() doubles the sleep time up to 30 seconds.

So after about 51 seconds it retries only about twice per minute. For
100 slots and no slot activity, a rough 1-hour test from the short
runs is on the order of a few MB on HEAD, and around 1 MB with the
manual RemoteSlot cleanup. For installations with fewer slots, it
should be smaller.

So my read is:
the accumulation is real;
it is modest because of the retry backoff;
manually freeing RemoteSlot’s copied strings removes most of the
observed growth;
a per-retry memory context fully bounds the retry-cycle lifetime


-- 
Regards,
Xuneng Zhou
HighGo Software Co., Ltd.