Thread: Postgres crash? could not write to log file: No space left on device
Dear All,
I have the following issue on Sun Solaris 10. PostgreSQL version is
9.2.3. The wall logging is minimal and no archiving. The DB restarted
several time, the box is up for last 23 days. The PostgreSQL
installation and files under /data/postgres that is half empty. Is it
some other destination that might cause the problem? Can I log the space
consumption and directory name where the problem is happening by some
debug level or trace setting? =20
=20
PANIC: could not write to log file 81, segment 125 at offset 13959168,
length 1392640: No space left on device
LOG: process 10203 still waiting for ShareLock on transaction 3010915
after 1004.113 ms
STATEMENT: UPDATE tctuserinfo SET clickmiles =3D clickmiles + $1,
periodicalclickmiles =3D periodicalclickmiles + $2, active =3D $3,
activeupdatetime =3D $4, activationsetby =3D $5, smscid =3D $6, sdrtime =
=3D $7,
simvalue =3D simvalue + $8, totalsimval
ue =3D totalsimvalue + $9, firstclick =3D $10, lastclick =3D $11,
firstactivationtime =3D $12, cbchannel =3D $13, clickmilesupdatetime =3D =
$14,
ci =3D $15, lac =3D $16, bscid =3D $17, lastlocationupdatetime =3D $18,
subscriptiontype =3D $19, contentcategory =3D
$20, livechannels =3D $21, contextclicks =3D $22 WHERE phonenumber =3D =
$23
LOG: WAL writer process (PID 10476) was terminated by signal 6
LOG: terminating any other active server processes
FATAL: the database system is in recovery mode
=20
But it looks OK:
=20
dbnetapp:/vol/postgres
90G 44G 46G 49% /data/postgres
=20
Is it possible that "heavy" queries consuming disk space (as temporary
space) and after the crash and recovery it becoming OK?=20
=20
"Yuri Levinsky" <yuril@celltick.com> writes:
> I have the following issue on Sun Solaris 10. PostgreSQL version is
> 9.2.3. The wall logging is minimal and no archiving. The DB restarted
> several time, the box is up for last 23 days. The PostgreSQL
> installation and files under /data/postgres that is half empty. Is it
> some other destination that might cause the problem? Can I log the space
> consumption and directory name where the problem is happening by some
> debug level or trace setting?
> PANIC: could not write to log file 81, segment 125 at offset 13959168,
> length 1392640: No space left on device
That's definitely telling you it got ENOSPC from a write in
$PGDATA/pg_xlog. Maybe you have a user-specific space quota affecting
the postgres account?
regards, tom lane
Tom,
There are no any space quota on user postgres. It always happens on
different log files. When I am creating huge dummy file in same FS -
nothing wrong.=20
P.S. I am getting same error when I am using pg_dump in tar format. I
don't know if there is some relation between 2 issues. The reason of the
pg_dump space failure is: when using tar format, pg_dump write temporary
data into /var, which might be smaller than biggest table requires. The
fix is simple in this case: to use custom format or increase /var. I
actually don't know what PostgreSQL has to do in /var and if it is same
problem? By the way it's impossible to understand where not enough space
from pg_dump output as well as in my original issue.=20
Sincerely yours,
Yuri Levinsky, DBA
Celltick Technologies Ltd., 32 Maskit St., Herzliya 46733, Israel
Mobile: +972 54 6107703, Office: +972 9 9710239; Fax: +972 9 9710222
-----Original Message-----
From: Tom Lane [mailto:tgl@sss.pgh.pa.us]=20
Sent: Tuesday, June 25, 2013 4:47 PM
To: Yuri Levinsky
Cc: pgsql-bugs@postgresql.org
Subject: Re: [BUGS] Postgres crash? could not write to log file: No
spaceleft on device
"Yuri Levinsky" <yuril@celltick.com> writes:
> I have the following issue on Sun Solaris 10. PostgreSQL version is=20
> 9.2.3. The wall logging is minimal and no archiving. The DB restarted=20
> several time, the box is up for last 23 days. The PostgreSQL=20
> installation and files under /data/postgres that is half empty. Is it=20
> some other destination that might cause the problem? Can I log the=20
> space consumption and directory name where the problem is happening by
> some debug level or trace setting?
> PANIC: could not write to log file 81, segment 125 at offset=20
> 13959168, length 1392640: No space left on device
That's definitely telling you it got ENOSPC from a write in
$PGDATA/pg_xlog. Maybe you have a user-specific space quota affecting
the postgres account?
regards, tom lane
Yuri, You're sure the pg xlogs are going where you expect them to? Have you= fine-tooth-combed your conf file for log file-related settings? Log files = may well be directed to fs other than /data/postgres (as is common in our e= nvironments, e.g.) Do a $ df -h on the various FSes involved... Are you using Solaris 10 ACLs? Dig deeper on Tom's point on user-specific q= uotas. ZFS in use? Various quota settings under Solaris can get you really = unexpected mileage. Lou Picciano ----- Original Message ----- From: Yuri Levinsky=20 To: pgsql-bugs@postgresql.org Sent: Tue, 25 Jun 2013 12:23:00 -0000 (UTC) Subject: [BUGS] Postgres crash? could not write to log file: No space left = on device Dear All, I have the following issue on Sun Solaris 10. PostgreSQL version= is 9.2.3. The wall logging is minimal and no archiving. The DB restarted s= everal time, the box is up for last 23 days. The PostgreSQL installation an= d files under /data/postgres that is half empty. Is it some other destinati= on that might cause the problem? Can I log the space consumption and direct= ory name where the problem is happening by some debug level or trace settin= g? PANIC: could not write to log file 81, segment 125 at offset 13959168= , length 1392640: No space left on deviceLOG: process 10203 still waiting = for ShareLock on transaction 3010915 after 1004.113 msSTATEMENT: UPDATE tc= tuserinfo SET clickmiles =3D clickmiles + $1, periodicalclickmiles =3D peri= odicalclickmiles + $2, active =3D $3, activeupdatetime =3D $4, activationse= tby =3D $5, smscid =3D $6, sdrtime =3D $7, simvalue =3D simvalue + $8, tota= lsimvalue =3D totalsimvalue + $9, firstclick =3D $10, lastclick =3D $11, fi= rstactivationtime =3D $12, cbchannel =3D $13, clickmilesupdatetime =3D $14,= ci =3D $15, lac =3D $16, bscid =3D $17, lastlocationupdatetime =3D $18, su= bscriptiontype =3D $19, contentcategory =3D$20, livechannels =3D $21, conte= xtclicks =3D $22 WHERE phonenumber =3D $23LOG: WAL writer process (PID 104= 76) was terminated by signal 6LOG: terminating any other active server pro= cessesFATAL: the database system is in recovery mode But it looks OK: dbne= tapp:/vol/postgres 90G 44G 46G 49% /data= /postgres Is it possible that =E2=80=9Cheavy=E2=80=9D queries consuming dis= k space (as temporary space) and after the crash and recovery it becoming O= K? =20
On Tue, Jun 25, 2013 at 9:43 AM, Yuri Levinsky <yuril@celltick.com> wrote: > I inspected my config file and didn't see any destination that isn't > /data/postgres. Have I perform any specific setting to limit it into > /data/postgres? Execute from psql: show stats_temp_directory stats_temp_directory ---------------------- pg_stat_tmp Assuming your stats_temp_directory is pg_stat_tmp, where is that directory located? Under $PGDATA? What is the path to $PGDATA? If your stats_temp_directory is located on a small partition, that could be a problem Where is your pg_xlog dir located?
bricklen <bricklen@gmail.com> writes:
> Assuming your stats_temp_directory is pg_stat_tmp, where is that directory
> located? Under $PGDATA?
> What is the path to $PGDATA?
> If your stats_temp_directory is located on a small partition, that could be
> a problem
> Where is your pg_xlog dir located?
The given error message is definitely complaining about being unable to
write a pg_xlog file --- stats or other temp files are not relevant.
regards, tom lane
On Tue, 2013-06-25 at 09:46 -0400, Tom Lane wrote:
> "Yuri Levinsky" <yuril@celltick.com> writes:
> > PANIC: could not write to log file 81, segment 125 at offset 13959168,
> > length 1392640: No space left on device
>
> That's definitely telling you it got ENOSPC from a write in
> $PGDATA/pg_xlog.
Either that, or write() wrote less than expected but did not set errno.
It looks like we assume ENOSPC when errno is not set.
Regards,
Jeff Davis
Jeff Davis <pgsql@j-davis.com> writes:
> On Tue, 2013-06-25 at 09:46 -0400, Tom Lane wrote:
>> That's definitely telling you it got ENOSPC from a write in
>> $PGDATA/pg_xlog.
> Either that, or write() wrote less than expected but did not set errno.
Good point. I wonder if he's using a filesystem that is capable of
reporting partial writes for other reasons, eg maybe it allows signals
to end writes early. (Though if it is, it's not apparent why such
failures would only be manifesting on the pg_xlog files and not for
anything else.)
regards, tom lane
On Wed, Jun 26, 2013 at 12:57 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > (Though if it is, it's not apparent why such > failures would only be manifesting on the pg_xlog files and not for > anything else.) Well data files are only ever written to in 8k chunks. Maybe these errors are only occuring on >8k xlog records such as records with multiple full page images. I'm not sure how much we write for other types of files but they won't be written to as frequently as xlog or data files and might not cause errors that are as noticeable. -- greg
On 2013-06-26 13:14:37 +0100, Greg Stark wrote: > On Wed, Jun 26, 2013 at 12:57 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > > (Though if it is, it's not apparent why such > > failures would only be manifesting on the pg_xlog files and not for > > anything else.) > > Well data files are only ever written to in 8k chunks. Maybe these > errors are only occuring on >8k xlog records such as records with > multiple full page images. I'm not sure how much we write for other > types of files but they won't be written to as frequently as xlog or > data files and might not cause errors that are as noticeable. We only write xlog in XLOG_BLCKSZ units - which is 8kb by default as well... Yuri, have you compiled postgres with nonstandard configure or pg_config_manual.h settings? Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Re: Postgres crash? could not write to log file: No space left on device
From
Heikki Linnakangas
Date:
On 26.06.2013 15:21, Andres Freund wrote: > On 2013-06-26 13:14:37 +0100, Greg Stark wrote: >> On Wed, Jun 26, 2013 at 12:57 AM, Tom Lane<tgl@sss.pgh.pa.us> wrote: >>> (Though if it is, it's not apparent why such >>> failures would only be manifesting on the pg_xlog files and not for >>> anything else.) >> >> Well data files are only ever written to in 8k chunks. Maybe these >> errors are only occuring on>8k xlog records such as records with >> multiple full page images. I'm not sure how much we write for other >> types of files but they won't be written to as frequently as xlog or >> data files and might not cause errors that are as noticeable. > > We only write xlog in XLOG_BLCKSZ units - which is 8kb by default as > well... Actually, XLogWrite() writes multiple pages at once. If all wal_buffers are dirty, it can try to write them all in one write() call. We've discussed retrying short writes before, and IIRC Tom has argued that it shouldn't be necessary when writing to disk. Nevertheless, I think we should retry in XLogWrite(). It can write much bigger chunks than most write() calls, so there's more room for a short write to happen there if it can happen at all. Secondly, it PANICs on failure, so it would be nice to try a bit harder to avoid that. - Heikki
On 2013-06-26 15:40:08 +0300, Heikki Linnakangas wrote: > On 26.06.2013 15:21, Andres Freund wrote: > >On 2013-06-26 13:14:37 +0100, Greg Stark wrote: > >>On Wed, Jun 26, 2013 at 12:57 AM, Tom Lane<tgl@sss.pgh.pa.us> wrote: > >>> (Though if it is, it's not apparent why such > >>>failures would only be manifesting on the pg_xlog files and not for > >>>anything else.) > >> > >>Well data files are only ever written to in 8k chunks. Maybe these > >>errors are only occuring on>8k xlog records such as records with > >>multiple full page images. I'm not sure how much we write for other > >>types of files but they won't be written to as frequently as xlog or > >>data files and might not cause errors that are as noticeable. > > > >We only write xlog in XLOG_BLCKSZ units - which is 8kb by default as > >well... > > Actually, XLogWrite() writes multiple pages at once. If all wal_buffers are > dirty, it can try to write them all in one write() call. Oh. Misremembered that. > We've discussed retrying short writes before, and IIRC Tom has argued that > it shouldn't be necessary when writing to disk. Nevertheless, I think we > should retry in XLogWrite(). It can write much bigger chunks than most > write() calls, so there's more room for a short write to happen t$here if it > can happen at all. Secondly, it PANICs on failure, so it would be nice to > try a bit harder to avoid that. At the very least we should log the amount of bytes actually writen if it was a short write to make it possible to discern that case from the direct ENOSPC response. This might also be caused by the fact that until recently the SIGALRM handler didn't set SA_RESTART... If a backend decided to write out the xlog directly it very well might have an active alarm... Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Sorry for delay, the email somehow went into junk mails folder. I didn't compiled Postgres, we downloaded the already compiled version for SUN Solaris 10 64 bit. The same version is working fine on other installations. On this specific installation we got "buggy kernel" error during massive data load (copy) from Informix. As result we added "noac" flag to NFS that looking on NetApp storage. The "noac" means disable NFS caching, the problem disappeared despite multiple and instant reproduces before. Unfortunately after site went up we started to see DB reboots for 2 days. I disabled pg_dump with tar option, that somehow filled up by 100% the /var directory and freed it after failure. I limited the temp storage per session as well. Unfortunately I lost the connection to the site and unable to check today if it helped or not.=20 Sincerely yours, Yuri Levinsky, DBA Celltick Technologies Ltd., 32 Maskit St., Herzliya 46733, Israel Mobile: +972 54 6107703, Office: +972 9 9710239; Fax: +972 9 9710222 -----Original Message----- From: Andres Freund [mailto:andres@2ndquadrant.com]=20 Sent: Wednesday, June 26, 2013 3:21 PM To: Greg Stark Cc: Tom Lane; Jeff Davis; Yuri Levinsky; pgsql-bugs@postgresql.org Subject: Re: [BUGS] Postgres crash? could not write to log file: No spaceleft on device On 2013-06-26 13:14:37 +0100, Greg Stark wrote: > On Wed, Jun 26, 2013 at 12:57 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > > (Though if it is, it's not apparent why such failures would only be > > manifesting on the pg_xlog files and not for anything else.) >=20 > Well data files are only ever written to in 8k chunks. Maybe these=20 > errors are only occuring on >8k xlog records such as records with=20 > multiple full page images. I'm not sure how much we write for other=20 > types of files but they won't be written to as frequently as xlog or=20 > data files and might not cause errors that are as noticeable. We only write xlog in XLOG_BLCKSZ units - which is 8kb by default as well... Yuri, have you compiled postgres with nonstandard configure or pg_config_manual.h settings? Greetings, Andres Freund --=20 Andres Freund http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services This mail was received via Mail-SeCure System.
Heikki Linnakangas <hlinnakangas@vmware.com> writes:
> We've discussed retrying short writes before, and IIRC Tom has argued
> that it shouldn't be necessary when writing to disk. Nevertheless, I
> think we should retry in XLogWrite(). It can write much bigger chunks
> than most write() calls, so there's more room for a short write to
> happen there if it can happen at all. Secondly, it PANICs on failure, so
> it would be nice to try a bit harder to avoid that.
Seems reasonable. My concern about the idea in general was the
impossibility of being sure we'd protected every single write() call.
But if we can identify specific call sites that seem at more risk than
most, I'm okay with adding extra logic there.
regards, tom lane
Dear All, I succeed to find some Postgres note where similar issue is related to NFS mount option 'intr'. It was recommended to use 'hard,nointr' on Solaris NFS or do not use NFS at all. The default on SUN is 'intr' and on Linux is 'nointr', even not specified. Surprisingly the issue doesn't happens on any my Linux installations. According to the very long note the issue happens during heavy load, when PostgreSQL doesn't have answer during some time. By the way adding 'noac' mount option prevent successfully reproduced by me the "buggy kernel" error that also followed by PostgreSQL crash on this my particular system. All this leads me to conclusion that something basically wrong: these mount options have to be documented somewhere as recommended for NFS or log/bg writers have to take it somehow into account. Can this be confirmed as a bug? Can it be fixed in some nearest future? 2013-07-02 01:37:58 GMTXX000PANIC: could not write to file "pg_xlog/xlogtemp.24205": Interrupted system call 2013-07-02 01:40:52 GMT00000LOG: WAL writer process (PID 24205) was terminated by signal 6 2013-07-02 01:40:52 GMT00000LOG: terminating any other active server processes 2013-07-02 01:40:52 GMT57P03FATAL: the database system is in recovery mode 2013-07-02 02:17:41 GMT57P03FATAL: the database system is in recovery mode 2013-07-02 02:17:41 GMT00000LOG: autovacuum launcher started 2013-07-02 02:17:41 GMT00000LOG: database system is ready to accept connections 2013-07-02 02:44:50 GMTXX000PANIC: could not write to file "pg_xlog/xlogtemp.14855": Interrupted system call 2013-07-02 02:48:02 GMT00000LOG: WAL writer process (PID 14855) was terminated by signal 6 2013-07-02 02:48:02 GMT00000LOG: terminating any other active server processes 2013-07-02 02:48:02 GMT57P03FATAL: the database system is in recovery mode 2013-07-02 04:15:49 GMTXX000PANIC: could not open file "pg_xlog/00000001000000B9000000C9" (log file 185, segment 201): Interrupted system call 2013-07-02 04:18:55 GMT00000LOG: WAL writer process (PID 2296) was terminated by signal 6 2013-07-02 04:18:55 GMT00000LOG: terminating any other active server processes 2013-07-02 04:18:55 GMT57P03FATAL: the database system is in recovery mode Sincerely yours, Yuri Levinsky, DBA Celltick Technologies Ltd., 32 Maskit St., Herzliya 46733, Israel Mobile: +972 54 6107703, Office: +972 9 9710239; Fax: +972 9 9710222 -----Original Message----- From: Andres Freund [mailto:andres@2ndquadrant.com]=20 Sent: Wednesday, June 26, 2013 4:04 PM To: Heikki Linnakangas Cc: Greg Stark; Tom Lane; Jeff Davis; Yuri Levinsky; pgsql-bugs@postgresql.org Subject: Re: [BUGS] Postgres crash? could not write to log file: No spaceleft on device On 2013-06-26 15:40:08 +0300, Heikki Linnakangas wrote: > On 26.06.2013 15:21, Andres Freund wrote: > >On 2013-06-26 13:14:37 +0100, Greg Stark wrote: > >>On Wed, Jun 26, 2013 at 12:57 AM, Tom Lane<tgl@sss.pgh.pa.us> wrote: > >>> (Though if it is, it's not apparent why such failures would only=20 > >>>be manifesting on the pg_xlog files and not for anything else.) > >> > >>Well data files are only ever written to in 8k chunks. Maybe these=20 > >>errors are only occuring on>8k xlog records such as records with=20 > >>multiple full page images. I'm not sure how much we write for other=20 > >>types of files but they won't be written to as frequently as xlog or > >>data files and might not cause errors that are as noticeable. > > > >We only write xlog in XLOG_BLCKSZ units - which is 8kb by default as=20 > >well... >=20 > Actually, XLogWrite() writes multiple pages at once. If all=20 > wal_buffers are dirty, it can try to write them all in one write() call. Oh. Misremembered that. > We've discussed retrying short writes before, and IIRC Tom has argued=20 > that it shouldn't be necessary when writing to disk. Nevertheless, I=20 > think we should retry in XLogWrite(). It can write much bigger chunks=20 > than most > write() calls, so there's more room for a short write to happen t$here > if it can happen at all. Secondly, it PANICs on failure, so it would=20 > be nice to try a bit harder to avoid that. At the very least we should log the amount of bytes actually writen if it was a short write to make it possible to discern that case from the direct ENOSPC response. This might also be caused by the fact that until recently the SIGALRM handler didn't set SA_RESTART... If a backend decided to write out the xlog directly it very well might have an active alarm... Greetings, Andres Freund --=20 Andres Freund http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services This mail was received via Mail-SeCure System.
Re: Postgres crash? could not write to log file: No space left on device
From
Heikki Linnakangas
Date:
On 26.06.2013 17:15, Tom Lane wrote: > Heikki Linnakangas<hlinnakangas@vmware.com> writes: >> We've discussed retrying short writes before, and IIRC Tom has argued >> that it shouldn't be necessary when writing to disk. Nevertheless, I >> think we should retry in XLogWrite(). It can write much bigger chunks >> than most write() calls, so there's more room for a short write to >> happen there if it can happen at all. Secondly, it PANICs on failure, so >> it would be nice to try a bit harder to avoid that. > > Seems reasonable. My concern about the idea in general was the > impossibility of being sure we'd protected every single write() call. > But if we can identify specific call sites that seem at more risk than > most, I'm okay with adding extra logic there. Committed a patch to add retry loop to XLogWrite(). I noticed that FileWrite() has some additional Windows-specific code to also retry on an ERROR_NO_SYSTEM_RESOURCES error. That's a bit scary, because we don't check for that in any other write() calls in the backend. If we really need to be prepared for that on Windows, I think that would need to be in a wrapper function in src/port or src/backend/port. Would a Windows-person like to comment on that? - Heikki