Обсуждение: Replication failed after stalling

Joe Van Dyk <joe@tanga.com> writes:

> I'm running Postgresql 9.3. I have a streaming replication server. Someone was running a long COPY query (8 hours) on
thestandby which halted replication. The 
> replication stopped at 3:30 am. I canceled the long-running query at 9:30 am and replication data started catching
up.
>
> The data up until 10 am got restored fine (took until 10:30 am to restore that much). Then I started getting errors
like"FATAL:  could not receive data from WAL 
> stream: ERROR:  requested WAL segment 00000001000003C300000086 has already been removed".
>
> I'm confused about how pg could restore data from 3:30 am to 10 am, then start complaining about missing WAL files.
>
> What's the best way to avoid this problem? Increase wal_keep_segments?

Yes and/or implement as a hybrid of streaming and WAL shipping.

Quite simply, your wal_keep segments was almost enough to  get you
through that backlog period but as your standby was catching up, it hit
a point  where  there was a gap.

Depending on how much traffic your master sees at various times of the
day, it's unsurprising that during peak loads, your grace-period is a
lot lower than during off-peak times due to variations in how quickly
WAL segments are filled and cycled over.

HTH

>
> Joe
>

--
Jerry Sievers
Postgres DBA/Development Consulting
e: postgres.consulting@comcast.net
p: 312.241.7800

On 12/18/2013 12:15 PM, Joe Van Dyk wrote:
> A possibly related question:
>
> I've set wal_keep_segments to 10,000 and also have archive_command
> running wal-e. I'm seeing my wal files disappear from pg_xlog after 30
> minutes. Is that expected? Is there a way around that?

Well a WAL segment is 16MB in size so that should give you a basis for
determining whether the above is appropriate, my guess it is not. I do
not know enough about Wal-e, but my guess is it is siphoning off WAL
segments before you want it to.




--
Adrian Klaver
adrian.klaver@gmail.com

On Wed, Dec 18, 2013 at 11:26 AM, Joe Van Dyk <joe@tanga.com> wrote:
> I'm running Postgresql 9.3. I have a streaming replication server. Someone
> was running a long COPY query (8 hours) on the standby which halted
> replication. The replication stopped at 3:30 am. I canceled the long-running
> query at 9:30 am and replication data started catching up.

What do you mean by "COPY on the standby halted replication"?

--
Kind regards,
Sergey Konoplev
PostgreSQL Consultant and DBA

http://www.linkedin.com/in/grayhemp
+1 (415) 867-9984, +7 (901) 903-0499, +7 (988) 888-1979
gray.ru@gmail.com

On Sun, Dec 29, 2013 at 9:56 PM, Joe Van Dyk <joe@tanga.com> wrote:
> On Wed, Dec 18, 2013 at 3:39 PM, Sergey Konoplev <gray.ru@gmail.com> wrote:
>>
>> On Wed, Dec 18, 2013 at 11:26 AM, Joe Van Dyk <joe@tanga.com> wrote:
>> > I'm running Postgresql 9.3. I have a streaming replication server.
>> > Someone
>> > was running a long COPY query (8 hours) on the standby which halted
>> > replication. The replication stopped at 3:30 am. I canceled the
>> > long-running
>> > query at 9:30 am and replication data started catching up.
>>
>> What do you mean by "COPY on the standby halted replication"?
>
> If I run "COPY (select * from complicate_view) to stdout" on the standby,
> I've noticed that sometimes halts replication updates to the slave.
>
> For example, that's happening right now and "now() -
> pg_last_xact_replay_timestamp()" is 22 minutes. There's many transactions
> per second being committed on the master. Once that query is canceled, the
> slave catches up immediately.

And what

\x
select * from pg_stat_repication;

shows?

--
Kind regards,
Sergey Konoplev
PostgreSQL Consultant and DBA

http://www.linkedin.com/in/grayhemp
+1 (415) 867-9984, +7 (901) 903-0499, +7 (988) 888-1979
gray.ru@gmail.com

Joe Van Dyk wrote:
> If I run "COPY (select * from complicate_view) to stdout" on the standby, I've noticed that sometimes
> halts replication updates to the slave.
> 
> For example, that's happening right now and "now() - pg_last_xact_replay_timestamp()" is 22 minutes.
> There's many transactions per second being committed on the master. Once that query is canceled, the
> slave catches up immediately.

You have hot_standby_feedback = on, right?

In that case that is expected behaviour.
Some change on the master conflicted with the query on the standby,
perhaps with a tuple cleaned up after a HOT update.  Replication will
stall until the query is done.

If that is unacceptable to you and you would rather have queries
canceled on the standby if they take too long, set hot_standby_feedback
back to off and increase vacuum_defer_cleanup_age.

But it is not possible to have a predictable maximum replication lag
and arbitrarily long running queries on the standby at the same time.

Yours,
Laurenz Albe

On Mon, Dec 30, 2013 at 12:27 AM, Albe Laurenz <laurenz.albe@wien.gv.at> wrote:
> Joe Van Dyk wrote:
>> If I run "COPY (select * from complicate_view) to stdout" on the standby, I've noticed that sometimes
>> halts replication updates to the slave.
>>
>> For example, that's happening right now and "now() - pg_last_xact_replay_timestamp()" is 22 minutes.
>> There's many transactions per second being committed on the master. Once that query is canceled, the
>> slave catches up immediately.
>
> You have hot_standby_feedback = on, right?
>
> In that case that is expected behaviour.
> Some change on the master conflicted with the query on the standby,
> perhaps with a tuple cleaned up after a HOT update.  Replication will
> stall until the query is done.

IIRC, the applying process is paused but the receiving one is going on
in this case, isn't it?

--
Kind regards,
Sergey Konoplev
PostgreSQL Consultant and DBA

http://www.linkedin.com/in/grayhemp
+1 (415) 867-9984, +7 (901) 903-0499, +7 (988) 888-1979
gray.ru@gmail.com

On Mon, Dec 30, 2013 at 8:56 PM, Joe Van Dyk <joe@tanga.com> wrote:
> On Mon, Dec 30, 2013 at 10:27 AM, Sergey Konoplev <gray.ru@gmail.com> wrote:
>> On Mon, Dec 30, 2013 at 12:02 AM, Joe Van Dyk <joe@tanga.com> wrote:
>> > On Sun, Dec 29, 2013 at 10:52 PM, Sergey Konoplev <gray.ru@gmail.com>
>> > wrote:
>> >> On Sun, Dec 29, 2013 at 9:56 PM, Joe Van Dyk <joe@tanga.com> wrote:
>> >> > On Wed, Dec 18, 2013 at 3:39 PM, Sergey Konoplev <gray.ru@gmail.com>
>> >> > wrote:
>> >> > If I run "COPY (select * from complicate_view) to stdout" on the
>> >> > standby,
>> >> > I've noticed that sometimes halts replication updates to the slave.
>> >>
>> >> \x
>> >> select * from pg_stat_repication;
>>
>> And it would be very useful to take a look at your checkpoints and
>> replication configuration parameters on both master and replica.
>
> master and replica have same settings.
>
> checkpoint_completion_target: 0.9
> checkpoint_segments: 16
> checkpoint_timeout: 5m
> checkpoint_warning: 30s
> hot_standby: on
> hot_standby_feedback: on

I meant all the replication settings, see [1]. And pg_stat_statements
when there is a problem, preferable the error, because when everything
is okay it is not very useful actually.

[1] http://www.postgresql.org/docs/9.3/static/runtime-config-replication.html

--
Kind regards,
Sergey Konoplev
PostgreSQL Consultant and DBA

http://www.linkedin.com/in/grayhemp
+1 (415) 867-9984, +7 (901) 903-0499, +7 (988) 888-1979
gray.ru@gmail.com

On Wed, Dec 18, 2013 at 1:51 PM, Adrian Klaver <adrian.klaver@gmail.com> wrote:
> On 12/18/2013 12:15 PM, Joe Van Dyk wrote:
>>
>> A possibly related question:
>>
>> I've set wal_keep_segments to 10,000 and also have archive_command
>> running wal-e. I'm seeing my wal files disappear from pg_xlog after 30
>> minutes. Is that expected? Is there a way around that?
>
>
> Well a WAL segment is 16MB in size so that should give you a basis for
> determining whether the above is appropriate, my guess it is not. I do not
> know enough about Wal-e, but my guess is it is siphoning off WAL segments
> before you want it to.

Don't some operations like forced checkpoints etc skip to the next WAL
resulting in them not necessarily being "full"?

On Mon, Dec 30, 2013 at 10:05 PM, Joe Van Dyk <joe@tanga.com> wrote:
>> I meant all the replication settings, see [1]. And pg_stat_statements
>> when there is a problem, preferable the error, because when everything
>> is okay it is not very useful actually.
>
> I don't understand, how is pg_stat_statements helpful here, and what error?

The error you showed in the initial email.

My guess is that the master might stop sending WAL records to the
replica, that is why I wanted to check the stat_replication query. Oh,
yes, and I forget to put current_xlog_location in the query. So, the
correct one is below.

\x
select pg_current_xlog_location(), * from pg_stat_replication;

> checkpoint_completion_target: 0.9
> checkpoint_segments: 16
> checkpoint_timeout: 5m
> checkpoint_warning: 30s
[...]
> max_wal_senders: 5
> wal_keep_segments: 10000
> vacuum_defer_cleanup_age: 0
> max_standby_archive_delay: 30s
> max_standby_streaming_delay: -1
> wal_receiver_status_interval: 10s
> hot_standby_feedback: on
[...]

That 10000 looks weird and I would increase checkpoint_segments and
checkpoint_timeout, but first let us check how often checkpoints and
checkpoint warnings happen on master. You can see it in logs. Turn
log_checkpoints on if it is off.

And also how many WAL your system generates and for what period.

ls -lt /path/to/pg_xlog/ | wc -l
ls -lt /path/to/pg_xlog/ | head
ls -lt /path/to/pg_xlog/ | tail

--
Kind regards,
Sergey Konoplev
PostgreSQL Consultant and DBA

http://www.linkedin.com/in/grayhemp
+1 (415) 867-9984, +7 (901) 903-0499, +7 (988) 888-1979
gray.ru@gmail.com

Sergey Konoplev wrote:
> On Mon, Dec 30, 2013 at 12:27 AM, Albe Laurenz <laurenz.albe@wien.gv.at> wrote:
>> Joe Van Dyk wrote:
>>> If I run "COPY (select * from complicate_view) to stdout" on the standby, I've noticed that sometimes
>>> halts replication updates to the slave.
>>>
>>> For example, that's happening right now and "now() - pg_last_xact_replay_timestamp()" is 22 minutes.
>>> There's many transactions per second being committed on the master. Once that query is canceled, the
>>> slave catches up immediately.
>>
>> You have hot_standby_feedback = on, right?
>>
>> In that case that is expected behaviour.
>> Some change on the master conflicted with the query on the standby,
>> perhaps with a tuple cleaned up after a HOT update.  Replication will
>> stall until the query is done.
> 
> IIRC, the applying process is paused but the receiving one is going on
> in this case, isn't it?

Frankly, I don't know.

Yours,
Laurenz Albe