Thread: Re: BUG #17386: btree index corruption after reindex concurrently on write heavy table

Re: BUG #17386: btree index corruption after reindex concurrently on write heavy table

From

Maxim Boguk

Date:

28 January 2022, 17:17:28

On Fri, Jan 28, 2022 at 3:42 PM PG Bug reporting form <noreply@postgresql.org> wrote:

The following bug has been logged on the website:

Bug reference: 17386
Logged by: Maxim Boguk
Email address: maxim.boguk@gmail.com
PostgreSQL version: 14.1
Operating system: Ubuntu Linux
Description:

Hi,

I found multiple cases of Btree index corruption after REINDEX someindex
CONCURRENTLY;.

What I found so far:
1)corruption exists in master, on all replicas and test server recovered
from base backup + wal archive (so it isn't some local hardware error)
2)it doesn't happen every time but at least two cases of corrupted indexes
found after database verification with amcheck.
3)it happen on huge and write heavy table

Attempt fix situation with REINDEX INDEX jobs_pkey CONCURRENTLY;

lead to corrupted index again in different tuple:

ERROR:  heap tuple (69306318,15) from table "jobs" lacks matching index tuple within index "jobs_pkey"

SELECT * FROM heap_page_item_attrs(get_raw_page('jobs', 69306318), 'jobs'::regclass) where lp=15;
-[ RECORD 1 ]---
lp          | 15
lp_off      | 24
lp_flags    | 2
lp_len      | 0

select  ctid,id,updated_at,created_at from jobs where ctid='(69306318,24)'::tid;
-[ RECORD 1 ]--------------------------
ctid       | (69306318,24)
id         | 26192320674
updated_at | 2022-01-28 13:08:27.224275
created_at | 2021-11-17 14:35:59.441979

And again updated_at happens during the REINDEX run.

select ctid,id,updated_at,created_at from jobs where id=26192320674;
(0 rows)

SELECT t_ctid, raw_flags, combined_flags FROM heap_page_item_attrs(get_raw_page('jobs', 69306318), 'jobs'::regclass), LATERAL heap_tuple_infomask_flags(t_infomask, t_infomask2)  where lp=24;
-[ RECORD 1 ]--+---------------------------------------------------------------------------------------------------------------------------------------------
t_ctid         | (69306318,24)
raw_flags      | {HEAP_HASNULL,HEAP_HASVARWIDTH,HEAP_XMAX_KEYSHR_LOCK,HEAP_XMAX_LOCK_ONLY,HEAP_XMIN_COMMITTED,HEAP_XMAX_INVALID,HEAP_UPDATED,HEAP_ONLY_TUPLE}
combined_flags | {}

Seems something broken in combination of HOT update and REINDEX CONCURRENTLY.

Maxim Boguk
Senior Postgresql DBA
https://dataegret.com/

Phone RU: +7 985 433 0000
Phone UA: +380 99 143 0000
Phone AU: +61 45 218 5678

LinkedIn: http://www.linkedin.com/pub/maksym-boguk/80/b99/b1b
Skype: maxim.boguk

"Доктор, вы мне советовали так не делать, но почему мне по-прежнему больно когда я так делаю ещё раз?"

Re: BUG #17386: btree index corruption after reindex concurrently on write heavy table

From

Noah Misch

Date:

29 January 2022, 05:34:18

On Fri, Jan 28, 2022 at 04:17:28PM +0200, Maxim Boguk wrote:
> On Fri, Jan 28, 2022 at 3:42 PM PG Bug reporting form <noreply@postgresql.org> wrote:
> > PostgreSQL version: 14.1

> Seems something broken in combination of HOT update and REINDEX
> CONCURRENTLY.

Agreed.  Can you create a self-contained test case, perhaps by adapting
contrib/amcheck/t/002_cic.pl?  If this bug is like the ones fixed between 14.0
and 14.1, the base backup and WAL won't help us, unfortunately.  The transient
states are what matter.  I did try modifying the test to have elements of your
scenario, including HOT updates, KEY SHARE locks, and frozen tuples.  That
didn't reproduce the bug.  I'm attaching what I tried.

Attachment

repro-reindex-hot-v0.patch

Re: BUG #17386: btree index corruption after reindex concurrently on write heavy table

From

Peter Geoghegan

Date:

29 January 2022, 06:00:31

On Fri, Jan 28, 2022 at 6:34 PM Noah Misch <noah@leadboat.com> wrote:
> Agreed.  Can you create a self-contained test case, perhaps by adapting
> contrib/amcheck/t/002_cic.pl?  If this bug is like the ones fixed between 14.0
> and 14.1, the base backup and WAL won't help us, unfortunately.  The transient
> states are what matter.  I did try modifying the test to have elements of your
> scenario, including HOT updates, KEY SHARE locks, and frozen tuples.  That
> didn't reproduce the bug.  I'm attaching what I tried.

If I had to guess, then I'd guess that this has something to do with
orphaned HOT chains, like those we saw in the bug report that led to
bugfix commit 18b87b20 (which is in 14.2 but not 14.1). I could easily
be wrong about that, so take it with a grain of salt. I find it a
little suspicious that we're hearing about a REINDEX CONCURRENTLY
problem in Postgres 14, which is much less mature than Postgres 12
(where REINDEX CONCURRENTLY first appeared).

-- 
Peter Geoghegan

Re: BUG #17386: btree index corruption after reindex concurrently on write heavy table

From

Michael Paquier

Date:

29 January 2022, 08:43:08

On Fri, Jan 28, 2022 at 07:00:31PM -0800, Peter Geoghegan wrote:
> If I had to guess, then I'd guess that this has something to do with
> orphaned HOT chains, like those we saw in the bug report that led to
> bugfix commit 18b87b20 (which is in 14.2 but not 14.1). I could easily
> be wrong about that, so take it with a grain of salt. I find it a
> little suspicious that we're hearing about a REINDEX CONCURRENTLY
> problem in Postgres 14, which is much less mature than Postgres 12
> (where REINDEX CONCURRENTLY first appeared).

Possible, but hard to say without an actual proof.  Maxim, if the
problem is reproducible easily on your end, could you give a try to v12
and v13 and see if it happens as well there?
--
Michael

Attachment

signature.asc

Re: BUG #17386: btree index corruption after reindex concurrently on write heavy table

From

Maxim Boguk

Date:

29 January 2022, 18:23:49

On Sat, Jan 29, 2022 at 7:43 AM Michael Paquier <michael@paquier.xyz> wrote:

On Fri, Jan 28, 2022 at 07:00:31PM -0800, Peter Geoghegan wrote:
> If I had to guess, then I'd guess that this has something to do with
> orphaned HOT chains, like those we saw in the bug report that led to
> bugfix commit 18b87b20 (which is in 14.2 but not 14.1). I could easily
> be wrong about that, so take it with a grain of salt. I find it a
> little suspicious that we're hearing about a REINDEX CONCURRENTLY
> problem in Postgres 14, which is much less mature than Postgres 12
> (where REINDEX CONCURRENTLY first appeared).

Possible, but hard to say without an actual proof. Maxim, if the
problem is reproducible easily on your end, could you give a try to v12
and v13 and see if it happens as well there?
--
Michael

I don't remember such problems during the last year on v13 with the same workload and the same periodic reindex.

It isn't easily reproduced, table question is 800GB size (almost 2TB with indexes) and around 1000 rows/s updated (5k in peaks), under such load chance to hit the error seems around 60% (e.g. more than half reindex attempts end with broken index).

I have a suitable powerful server for tests, but there is no good way to simulate production workload (especially reproducible).

I'll see what I can do next week.

Re: BUG #17386: btree index corruption after reindex concurrently on write heavy table

From

Noah Misch

Date:

30 January 2022, 06:43:27

On Sat, Jan 29, 2022 at 05:23:49PM +0200, Maxim Boguk wrote:
> On Sat, Jan 29, 2022 at 7:43 AM Michael Paquier <michael@paquier.xyz> wrote:
> > On Fri, Jan 28, 2022 at 07:00:31PM -0800, Peter Geoghegan wrote:
> > > If I had to guess, then I'd guess that this has something to do with
> > > orphaned HOT chains, like those we saw in the bug report that led to
> > > bugfix commit 18b87b20 (which is in 14.2 but not 14.1). I could easily
> > > be wrong about that, so take it with a grain of salt. I find it a
> > > little suspicious that we're hearing about a REINDEX CONCURRENTLY
> > > problem in Postgres 14, which is much less mature than Postgres 12
> > > (where REINDEX CONCURRENTLY first appeared).
> >
> > Possible, but hard to say without an actual proof.  Maxim, if the
> > problem is reproducible easily on your end, could you give a try to v12
> > and v13 and see if it happens as well there?
> 
> I don't remember such problems during the last year on v13 with the same
> workload and the same periodic reindex.

Got it.  A low-effort strategy would be to wait for v14.2 to release on
2022-02-10.  If the problem started with a v13.X -> v14.1 upgrade and ceases
in v14.2, then commit 18b87b20 likely was the fix.

Re: BUG #17386: btree index corruption after reindex concurrently on write heavy table

From

Maxim Boguk

Date:

02 February 2022, 12:24:42

On Sun, Jan 30, 2022 at 5:43 AM Noah Misch <noah@leadboat.com> wrote:

On Sat, Jan 29, 2022 at 05:23:49PM +0200, Maxim Boguk wrote:
> On Sat, Jan 29, 2022 at 7:43 AM Michael Paquier <michael@paquier.xyz> wrote:
> > On Fri, Jan 28, 2022 at 07:00:31PM -0800, Peter Geoghegan wrote:
> > > If I had to guess, then I'd guess that this has something to do with
> > > orphaned HOT chains, like those we saw in the bug report that led to
> > > bugfix commit 18b87b20 (which is in 14.2 but not 14.1). I could easily
> > > be wrong about that, so take it with a grain of salt. I find it a
> > > little suspicious that we're hearing about a REINDEX CONCURRENTLY
> > > problem in Postgres 14, which is much less mature than Postgres 12
> > > (where REINDEX CONCURRENTLY first appeared).
> >
> > Possible, but hard to say without an actual proof. Maxim, if the
> > problem is reproducible easily on your end, could you give a try to v12
> > and v13 and see if it happens as well there?
>
> I don't remember such problems during the last year on v13 with the same
> workload and the same periodic reindex.

Got it. A low-effort strategy would be to wait for v14.2 to release on
2022-02-10. If the problem started with a v13.X -> v14.1 upgrade and ceases
in v14.2, then commit 18b87b20 likely was the fix.

Ok, I'll wait for the 14.2 release and provide an update about this issue.

Maxim Boguk
Senior Postgresql DBA
https://dataegret.com/

Re: BUG #17386: btree index corruption after reindex concurrently on write heavy table

From

Andres Freund

Date:

03 February 2022, 00:01:49

Hi,

On 2022-01-28 18:34:18 -0800, Noah Misch wrote:
> If this bug is like the ones fixed between 14.0 and 14.1, the base backup
> and WAL won't help us, unfortunately.

FWIW: I've found that once one has identified a specific page / tuple that is
damaged, searching the WAL for the record causing the problem is really
helpful. Once that record is identified, the records from the transaction
logging the problematic WAL record(s) can often identify the concrete activity
at the time. And looking at the WAL records of concurrently running
transactions can often identify the concurrency problem.

It might not be worth it in this case (given we have a suspicion it's
18b87b20), but when you have a base backup from before the problem, and all
incremental WAL, just replaying forward incrementally and testing for the
problem can help narrow down problems a lot.

Greetings,

Andres Freund