Обсуждение: BUG #16094: Database entering recovery mode

The following bug has been logged on the website:

Bug reference:      16094
Logged by:          Mircea Pirv
Email address:      mircea@reva.tech
PostgreSQL version: 12.0
Operating system:   OSX
Description:

Hello, 

We've updated our database from 10.6 to 12.0 recently and we keep
encountering an error which says that the database is entering recovery
mode.
Our setup has replication enabled, with a replication slot, and using a
replication subscription. On our development machines replication is setup
to use the same machine.
The problem is, that with replication enabled, changing a column value using
an update statement, causes the database to enter recovery mode.
This happens only when replication is enabled.
We could not find any pattern, like what kind of columns might be the
culprit. It happens with all kinds of columns, regardless of the fact that
we have constraints, triggers, defaults on the columns. With some columns it
happens randomly, with some it happens all the time.

In the logs the only error we see, is a segmentation fault, when replication
tries to run.
We tried with both a local server, and a dockerized version (12.0).

It only reproduces in 12.0. 10.6 and 11.5 do not have this problem, we could
not reproduce it.

Any ideas on what may have changed in 12.0 to cause this behavior ?

Thanks,
Mircea

PG Bug reporting form <noreply@postgresql.org> writes:
> We've updated our database from 10.6 to 12.0 recently and we keep
> encountering an error which says that the database is entering recovery
> mode.
> ...
> In the logs the only error we see, is a segmentation fault, when replication
> tries to run.

It should be possible to collect a stack trace for the segfault,
which would greatly assist debugging this.  Please see

https://wiki.postgresql.org/wiki/Generating_a_stack_trace_of_a_PostgreSQL_backend

            regards, tom lane

On Wed, Nov 06, 2019 at 03:37:31PM +0200, Mircea Pirv wrote:
> #2  WaitEventSetWait (set=0x55bc137a5948, timeout=<error reading variable:
> Cannot access memory at address 0x7ffee2d44ae0>, occurred_events=<error
> reading variable: Cannot access memory at address 0x7ffee2d44ad8>,
> nevents=<error reading variable: Cannot access memory at address
> 0x7ffee2d44afc>, wait_event_info=<optimized out>) at
> ./build/../src/backend/storage/ipc/latch.c:1032
>         rc = <optimized out>
>         returned_events = 0
>         start_time = <error reading variable start_time (Cannot access
> memory at address 0x7ffee2d44b00)>
>         cur_time = <error reading variable cur_time (Cannot access memory
> at address 0x7ffee2d44b10)>
>         cur_timeout = <error reading variable cur_timeout (Cannot access
> memory at address 0x7ffee2d44ae8)>
> Backtrace stopped: Cannot access memory at address 0x7ffee2d44b78

Is this the only part of the stack you can get?  WaitEventSetWait()
gets called in three places, which are be-secure.c, syslogger.c and
condition_variable.c.  Could it be possible to see more of the actual
callers here?
--
Michael