Re: Issue with WAL logs temporary not replaying

On 02/01/2018 08:17 AM, Eugene Pirogov wrote:
> Tomas,
> 
> Indeed, I haven't put enough detail in the original post. That is a fair
> point, thank you. Let me follow up on each of your comments.
> 
>     unfortunately your message is missing some pretty important
>     information, like for example which OS you're running, what
>     PostgreSQL version are you using (and if you built it on your own or
>     installed from some package repository).
> 
> 
> We are using Docker container based on Alpine 3.5. It has a custom-built
> PostgreSQL 9.6.3 installation. The customization is negligible, e.g. we
> did not alter configure and make process in any way. We only included
> pglogical extension. Here are the very sources for our image
> – https://github.com/edenlabllc/alpine-postgre/blob/master/pglogical/Dockerfile
> <https://github.com/edenlabllc/alpine-postgre/blob/master/pglogical/Dockerfile>
> . The Dockerfile is based on the "official" postgres image 9.6.3 from
> docker library,
> e.g. https://github.com/docker-library/postgres/tree/a554d043a3b77937120d325b30fefaad2e3be12d/9.6/alpine
> <https://github.com/docker-library/postgres/tree/a554d043a3b77937120d325b30fefaad2e3be12d/9.6/alpine>
> .
> 
>     Also, you mention replicas. What type of replication are you using?
>     The built-in streaming replication, or something else?
> 
> 
> We use logical replication. We are interested in replicating selected
> columns from tables across several databases in our kubernetes cluster.
> For this we are using pglogical REL2.0.1, built with patch that allows
> it to compile under Alpine. Here's the
> patch: https://github.com/alpinelinux/aports/commit/f4f856564dfb2589c1800bd51322f4eadfc67cf1
> <https://github.com/alpinelinux/aports/commit/f4f856564dfb2589c1800bd51322f4eadfc67cf1>
> 
>     > Firstly, I very much doubt using pg_resetxlog is a good idea, especially when it's rather unclear what was the
actualroot cause.
 
> 
> 
> Noted. In hindsight I think our best course of action would have been to
> make a copy of PG_DATA table.

Yeah, that would be helpful, but it's too late to do that.

> 
>     Secondly, I don't quite understand what you mean by "recovery" here.
>     The regular recovery process (say, after a crash) makes it
>     impossible to connect to the database, so I'm not sure what you mean
>     by running vacuum analyze (not to mention how it could have improved
>     the recovery speed).
> 
> 
> By recovery I mean a sort of self-healing of a running database. Let me
> make a little step back and explain this.
> 
> After several restart attempts, the DB would still not get up. Instead
> it was complaining about inability to "locate a valid checkpoint
> record". After we ran pg_resetxlog and tried to start the database
> again, the "locate a valid checkpoint record" error was gone, and we
> ended up with a 100% working database.
> 

OK. So what pg_resetxlog command have you used exactly? There are
various options.

> From this point onwards we began issuing SELECT statements, once in
> several minutes, to see if the database sees the missing records. The
> records indeed began to slowly appear in our tables. I _think_ this was
> happening because the database began reading them from WAL, in the
> background. Again – at this point the database was (or seemed to be)
> fully operational. It was just catching up on data (this is my impression).
> 

That is not how WAL works. It's only read when the database is starting
after an unclean shutdown/crash. But it's not possible to connect to a
database in this recovery mode. And after the database gets into a
consistent state, it does not read WAL anymore.

Moreover, pg_resetxlog actually removes all the WAL segments, so there's
nothing to read ...

So the records were appearing for some other reasons. Either there were
being inserted by some other process, or becoming visible for some other
reason (e.g. because there were XIDs initially from the "future" and the
database eventually generated them again).

> It seems as if pg_resetxlog removed some WAL segment (hence the
> mentioned "PANIC:  could not locate a valid checkpoint record" log),
> which initially caused the database to stuck and not read further WAL
> segments.
> 
>     Well, considering you did pg_resetxlog ... I'm surprised you only
>     lost two records, but it's a question what state are the indexes in etc.
> 
> 
> I never considered indexes to play part in the issue. Noted on that,
> thank you.
> 

Well, the pg_resetxlog docs actually say this:

    After running this command, it should be possible to start the
    server, but bear in mind that the database might contain
    inconsistent data due to partially-committed transactions. You
    should immediately dump your data, run initdb, and reload. After
    reload, check for inconsistencies and repair as needed.

and I strongly recommend doing exactly that - dump/restore. The sooner
the better.

>     Man, that's an extremely noisy log file. There are errors caused by
>     broken pg_hba.conf entries and so on, which makes it extremely
>     difficult to look for possible other errors ... I'm not going to
>     waste time by doing that.
> 
> 
> Ugh, sorry about that. I wanted to provide a vanilla version of the log,
> e.g. "as is". Below is an updated variant. I removed duplicate lines.
> The log file went down from 14794 to 329 lines:
>
https://gist.githubusercontent.com/gmile/6df9ce393b11fb369856b84c4e9f755a/raw/7779c8341482d8df39257ab033e809b187227389/postgresql-short.log
>
<https://gist.githubusercontent.com/gmile/6df9ce393b11fb369856b84c4e9f755a/raw/7779c8341482d8df39257ab033e809b187227389/postgresql-short.log>
> 
> The most interesting stuff, in my opinion, is around PANIC lines. Also,
> there are a lot of the following lines:
> 
>     ERROR:  requested WAL segment 000000010000000300000089 has already
>     been removed
>
> 
> I'm yet to find out why these happened and what they mean.
> 

Well, I suppose those started after the pg_resetxlog, which removes all
the WAL segments. So if you have a replica (say, logical replica which
connects through a replication slot) that needs old WAL, you'll get
exactly this error. I.e. your pglogical replication was likely broken.

I find this much more interesting:

2018-01-30T14:49:31Z FATAL:  terminating connection due to unexpected
postmaster exit

There's a log message about smart shutdown request, which should wait
for all clients to regularly disconnect. So likely something crashed.

Can you provide some timeline of the events? I mean, at which time you
discovered the issues, at which time you did the pg_resetxlog etc?

> For the record, these are the lines that I removed from the log:
> 
>     sed -i '' '/LOG:  incomplete startup packet/d' postgres.log
>     sed -i '' '/FATAL:  password authentication failed for user "db"/d
>     postgres.log
>     sed -i '' '/DETAIL:  Password does not match for user "db"./d
>     postgres.log
>     sed -i '' '/Connection matched pg_hba.conf line 95: "host all all
>     all md5"/d' postgres.log
>     sed -i '' '/DETAIL:  Password does not match for user "pghoard"/d'
>     postgres.log
>     sed -i '' '/FATAL:  password authentication failed for user
>     "pghoard"/d' postgres.log
>     sed -i '' '/WARNING:  concurrent insert in progress within table
>     "declarations_status_hstr"/d' postgres.log
>     sed -i '' '/WARNING:  concurrent insert in progress within table
>     "audit_log"/d' postgres.log
> 
> 
>     I don't know. You have not provided any information about the
>     storage system, and we don't have access to system logs that might
>     show if something happened at this level. But in general yes, a
>     storage issue might cause this.
> 
> 
> We use volumes provided by Kubernetes which, in turn, are disks from
> Google Compute Engine. I don't know the exact characteristics of SSDs.
> Would it make sense to try and get to the bottom of the SSDs stats?
> 
> The file system used by PG_DATA folder is ext4.
> 

OK. I have little experience with GCE, but my guess there was some sort
of I/O issue and the control file (which tracks checkpoint records etc.)
got corrupted. And then it went downhill. Unfortunately, we'll likely
never know without a backup of the original DATA directory.

>     It wouldn't. But it's unclear how you came to that conclusion.
> 
> 
> My logic is the following:
> 
>  1. db is up and responding to write requests,
>  2. the data is not actually seen in the tables,
>  3. the data is actually seen in a logical replica,
>  4. we restart db, it refuses to start and raises an error,
>  5. we mitigate the error,
>  6. db restarts successfully,
>  7. data slowly restores "by itself".
> 
> Given replica had good data and since by design replica takes its data
> from original DB's WAL... we concluded that all inserted data made it to
> the WAL but not further, e.g. not in the actual tables. Again: since the
> records were missing from tables in the original database, somehow
> PostgreSQL was not applying data from WAL to tables.
> > It all looked as if some code in postgresql that does "now, move data
> from WAL to table" action would get stuck or something. While at the
> same time worker processes that deal with DB connections and queries
> would continue to operate normally. Ultimately pg_resetxlog would
> unstuck the "now, move data from WAL to table" code.
> > I'm totally may be making things up here. This is only my impression
> based on observation. I'd very much like to be corrected.
> 


Again, that's not how WAL is used - your mental model of the inner
workings is incorrect. We do generate the WAL and then modify the table
data. We don't "apply" WAL unless a recovery is necessary, which only
happens after a crash etc.

So if you haven't seen the data in the tables, it was for some other
unknown reason.

>     There's only one log for PostgreSQL, so it would have to be in that
>     file depending on the log level.
> 
> 
> Got it.
> 
>     Please inspect the core dumps using gdb and share the backtraces.
>     There are instructions how to do that on the wiki
> 
> 
> I'm going to try and extract backtraces from the core dumps. I don't
> usually do this on a daily basis, so it will take some time. I will come
> back with results when I have something.
> 

OK. That would be useful.

regards

-- 
Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services