Re: backend stuck in DataFileExtend

Hi,

On 2024-05-06 09:05:38 -0500, Justin Pryzby wrote:
> In March, I noticed that a backend got stuck overnight doing:
> 
> backend_start    | 2024-03-27 22:34:12.774195-07
> xact_start       | 2024-03-27 22:34:39.741933-07
> query_start      | 2024-03-27 22:34:41.997276-07
> state_change     | 2024-03-27 22:34:41.997307-07
> wait_event_type  | IO
> wait_event       | DataFileExtend
> state            | active
> backend_xid      | 330896991
> backend_xmin     | 330896991
> query_id         | -3255287420265634540
> query            | PREPARE mml_1 AS INSERT INTO child.un...
> 
> The backend was spinning at 100% CPU:
> 
> [pryzbyj@telsa2021 ~]$ ps -O wchan,pcpu 7881
>    PID WCHAN  %CPU S TTY          TIME COMMAND
>   7881 ?      99.4 R ?        08:14:55 postgres: telsasoft ts [local] INSERT
> 
> This was postgres 16 STABLE compiled at 14e991db8.
> 
> It's possible that this is a rare issue that we haven't hit before.
> It's also possible this this is a recent regression.  We previously
> compiled at b2c9936a7 without hitting any issue (possibly by chance).
> 
> I could neither strace the process nor attach a debugger.  They got
> stuck attaching.  Maybe it's possible there's a kernel issue.  This is a
> VM running centos7 / 3.10.0-1160.49.1.el7.x86_64.

> $ awk '{print $14, $15}' /proc/7881/stat # usr / sys
> 229 3088448
> 
> When I tried to shut down postgres (hoping to finally be able to attach
> a debugger), instead it got stuck:

That strongly indicates either a kernel bug or storage having an issue. It
can't be postgres' fault if an IO never completes.

What do /proc/$pid/stack say?


> In both cases, the backend got stuck after 10pm, which is when a backup
> job kicks off, followed by other DB maintenance.  Our backup job uses
> pg_export_snapshot() + pg_dump --snapshot.  In today's case, the pg_dump
> would've finished and snapshot closed at 2023-05-05 22:15.  The backup
> job did some more pg_dumps involving historic tables without snapshots
> and finished at 01:11:40, at which point a reindex job started, but it
> looks like the DB was already stuck for the purpose of reindex, and so
> the script ended after a handful of commands were "[canceled] due to
> statement timeout".

Is it possible that you're "just" waiting for very slow IO? Is there a lot of
dirty memory? Particularly on these old kernels that can lead to very extreme
delays.

grep -Ei 'dirty|writeback' /proc/meminfo


> [...]
> Full disclosure: the VM that hit this issue today has had storage-level
> errors (reported here at ZZqr_GTaHyuW7fLp@pryzbyj2023), as recently as 3
> days ago.

So indeed, my suspicion from above is confirmed.

Greetings,

Andres Freund