Unexplained rapid growth in memory usage of idle backends

Hi,

We have an application that has a preforked worker pool architecture and opens a relatively large (70+) number of persistent connections to PostgreSQL 14, and have `max_connections` set to 200 accordingly.

This has worked well enough in many deployments, and the backends corresponding to these connections mostly sit idle. Occasionally they are hit with a query, and those workloads are almost entirely PL/PgSQL stored functions. These functions mostly just aggregate output from various tables, and occasionally apply some business logic; no temporary tables, no runaway recursive CTEs, nothing baroque.

Occasionally, about once every 2 months, we'll get a slow-motion implosion over a period of about 24 hours, where the resident memory size of some of the backends shoots up from a reasonable few hundred MB to several gigs -- and will keep growing, e.g.

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
3932733 postgres 20 0 637740 327124 323840 R 10.7 0.5 0:00.11 postgres
782930 postgres 20 0 3151360 2.8g 551704 S 9.7 4.5 221:43.17 postgres
782839 postgres 20 0 2999896 2.7g 551832 S 8.7 4.3 220:54.55 postgres
3932734 postgres 20 0 636128 338168 336096 R 8.7 0.5 0:00.09 postgres
3932735 postgres 20 0 636128 319824 317768 R 8.7 0.5 0:00.09 postgres
782851 postgres 20 0 3142152 2.8g 551872 S 7.8 4.5 221:22.62 postgres
782855 postgres 20 0 3155144 2.8g 551828 S 7.8 4.5 221:38.59 postgres
782897 postgres 20 0 3148808 2.8g 551724 S 7.8 4.5 220:05.94 postgres
783106 postgres 20 0 3152416 2.8g 551912 S 7.8 4.5 221:24.40 postgres
783168 postgres 20 0 2992592 2.7g 551968 S 7.8 4.2 220:36.06 postgres
782843 postgres 20 0 3146772 2.8g 551944 R 4.9 4.5 221:38.46 postgres
782778 postgres 20 0 3150956 2.8g 551852 R 3.9 4.5 220:49.88 postgres

It would be only natural to expect a cornucopia of heavyweight queries there, but there aren't any. `pg_stat_activity` just reveals an active WalSender (for streaming replication) and the `pg_stat_activity` query (WHERE state <> 'idle') itself. Once in a while, I'll catch a shortlived query in _one_ of these, if I run it often enough.

The state of affairs deteriorates until either

1) the grim OOM reaper comes:

[19063737.017400] Out of memory: Killed process 536356 (postgres) total-vm:21703068kB, anon-rss:20804636kB, file-rss:4kB, shmem-rss:534896kB, UID:107 pgtables:42384kB oom_score_adj:0
[19063739.149563] oom_reaper: reaped process 536356 (postgres), now anon-rss:0kB, file-rss:0kB, shmem-rss:534896kB,

or

2) the client application is restarted, closing the persistent connections and terminating their corresponding backends. This releases the memory back to the OS and all is well again.

For reference:

----- postgresql.auto.conf -----
listen_addresses = '*'
timezone = 'UTC'
max_connections = '200'
effective_cache_size = '2GB'
work_mem = '128MB'
wal_level = 'replica'
fsync = 'on'
synchronous_commit = 'off'
checkpoint_completion_target = '0.75'
checkpoint_warning = '15min'
autovacuum = 'on'
autovacuum_freeze_max_age = '200000000'
vacuum_freeze_min_age = '10000'
vacuum_freeze_table_age = '1000000000'
primary_conninfo = 'user=replication passfile=''/root/.pgpass'' channel_binding=prefer host=xxx.xxx.xxx.xxx port=5432'
promote_trigger_file = '/var/lib/postgresql/14/main/failover_trig'
primary_slot_name = 'replication_db5'
log_min_duration_statement = '-1'
max_worker_processes = '4'
max_parallel_workers_per_gather = '4'
shared_buffers = '512MB'
--------------------------------

And:

----- free -----
total used free shared buff/cache available
Mem: 65837856 2623820 47176960 567988 16037076 61989160
Swap: 1000444 0 1000444
----------------

Lastly, I will reiterate that there is no evidence of a slow-brewing memory leak. Memory usage seems more or less steady-state in general, rising and falling in expected ways with peak and off-peak workloads. Then, suddenly, some of the backends go into into the aforementioned nosedive.

I have considered using a connection pooler to limit the overall memory footprint and blast radius, and while this might address the problem from an operational point of view, it does not really resolve the essential question: why is this happening in the first place, seemingly out of nowhere?

I will also say that while the client application does call a lot of stored functions, they are all rather circumscribed in scope, in support of real-time routing decisions. These are not expensive reporting queries as might be issued from an API or a user interface of some kind, for example. The client IPs on the problematic backends above correspond to the client application, not to any other connected clients.

I'm at an utter loss as to how to troubleshoot or prevent this. Any insight would be deeply appreciated!

-- Alex

--
Alex Balashov
Principal Consultant
Evariste Systems LLC
Web: https://evaristesys.com
Tel: +1-706-510-6800