Reading backup label file for checkpoint and redo location during crash recovery

Greetings Satya,

* Satyanarayana Narlapuram (Satyanarayana.Narlapuram@microsoft.com) wrote:

During crash recovery, last checkpoint record information is obtained from the backup label if present, instead of getting it from the control file. This behavior is causing PostgreSQL database cluster not to come up until the backup label file is deleted (as the error message says).

if (checkPoint.redo < checkPointLoc)
{
if (!ReadRecord(xlogreader, checkPoint.redo, LOG, false))
ereport(FATAL,
(errmsg("could not find redo location referenced by checkpoint record"),
errhint("If you are not restoring from a backup, try removing the file \"%s/backup_label\".", DataDir)));
}

If we are recovering from a dump file, reading from the backup label files makes sense as the control file could be archived after a few checkpoints. But this is not the case for crash recovery, and is always safe to read the checkpoint record information from the control file.
Is this behavior kept this way as there is no clear way to distinguish between the recovery from the dump and the regular crash recovery?

This is why the exclusive backup method has been deprecated in PG10 in
favor of the non-exclusive backup method, which avoids this by not
creating a backup label file (it's up to the backup software to store
the necessary information and create the file for use during recovery).

Please see:

https://www.postgresql.org/docs/10/static/continuous-archiving.html

In particular, section 25.3.3.

Thanks!

Stephen

On 2017-09-25 13:43:32 -0400, Stephen Frost wrote:

Greetings Satya,

* Satyanarayana Narlapuram (Satyanarayana.Narlapuram@microsoft.com) wrote:

During crash recovery, last checkpoint record information is obtained from the backup label if present, instead of getting it from the control file. This behavior is causing PostgreSQL database cluster not to come up until the backup label file is deleted (as the error message says).

if (checkPoint.redo < checkPointLoc)
{
if (!ReadRecord(xlogreader, checkPoint.redo, LOG, false))
ereport(FATAL,
(errmsg("could not find redo location referenced by checkpoint record"),
errhint("If you are not restoring from a backup, try removing the file \"%s/backup_label\".", DataDir)));
}

If we are recovering from a dump file, reading from the backup label files makes sense as the control file could be archived after a few checkpoints. But this is not the case for crash recovery, and is always safe to read the checkpoint record information from the control file.
Is this behavior kept this way as there is no clear way to distinguish between the recovery from the dump and the regular crash recovery?

This is why the exclusive backup method has been deprecated in PG10 in
favor of the non-exclusive backup method, which avoids this by not
creating a backup label file (it's up to the backup software to store
the necessary information and create the file for use during recovery).

Please see:

https://www.postgresql.org/docs/10/static/continuous-archiving.html

In particular, section 25.3.3.

Might not be obvious for the more casual contributor:

And even before then, using pg_basebackup / the streaming replication
protocol version of creating a base backup, it was avoided.

Greetings,

Andres Freund

--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers

On Mon, Sep 25, 2017 at 7:43 PM, Stephen Frost <sfrost@snowman.net> wrote:

Greetings Satya,

* Satyanarayana Narlapuram (Satyanarayana.Narlapuram@microsoft.com) wrote:

During crash recovery, last checkpoint record information is obtained

from the backup label if present, instead of getting it from the control
file. This behavior is causing PostgreSQL database cluster not to come up
until the backup label file is deleted (as the error message says).

if (checkPoint.redo < checkPointLoc)
{
if (!ReadRecord(xlogreader,

checkPoint.redo, LOG, false))

ereport(FATAL,
(errmsg("could not

find redo location referenced by checkpoint record"),

errhint("If you are

not restoring from a backup, try removing the file \"%s/backup_label\".",
DataDir)));

}

If we are recovering from a dump file, reading from the backup label

files makes sense as the control file could be archived after a few
checkpoints. But this is not the case for crash recovery, and is always
safe to read the checkpoint record information from the control file.

Is this behavior kept this way as there is no clear way to distinguish

between the recovery from the dump and the regular crash recovery?

This is why the exclusive backup method has been deprecated in PG10 in
favor of the non-exclusive backup method, which avoids this by not
creating a backup label file (it's up to the backup software to store
the necessary information and create the file for use during recovery).

Actally, it was deprecated already in 9.6, not just 10.

--
Magnus Hagander
Me: https://www.hagander.net/ <http://www.hagander.net/&gt;
Work: https://www.redpill-linpro.com/ <http://www.redpill-linpro.com/&gt;

* Magnus Hagander (magnus@hagander.net) wrote:

On Mon, Sep 25, 2017 at 7:43 PM, Stephen Frost <sfrost@snowman.net> wrote:

* Satyanarayana Narlapuram (Satyanarayana.Narlapuram@microsoft.com) wrote:

During crash recovery, last checkpoint record information is obtained

from the backup label if present, instead of getting it from the control
file. This behavior is causing PostgreSQL database cluster not to come up
until the backup label file is deleted (as the error message says).

if (checkPoint.redo < checkPointLoc)
{
if (!ReadRecord(xlogreader,

checkPoint.redo, LOG, false))

ereport(FATAL,
(errmsg("could not

find redo location referenced by checkpoint record"),

errhint("If you are

not restoring from a backup, try removing the file \"%s/backup_label\".",
DataDir)));

}

If we are recovering from a dump file, reading from the backup label

files makes sense as the control file could be archived after a few
checkpoints. But this is not the case for crash recovery, and is always
safe to read the checkpoint record information from the control file.

Is this behavior kept this way as there is no clear way to distinguish

between the recovery from the dump and the regular crash recovery?

This is why the exclusive backup method has been deprecated in PG10 in
favor of the non-exclusive backup method, which avoids this by not
creating a backup label file (it's up to the backup software to store
the necessary information and create the file for use during recovery).

Actally, it was deprecated already in 9.6, not just 10.

Whoops, right. Thanks for the clarification. :)

Stephen

Thank you! Got it.

-----Original Message-----
From: Stephen Frost [mailto:sfrost@snowman.net]
Sent: Monday, September 25, 2017 10:57 AM
To: Magnus Hagander <magnus@hagander.net>
Cc: Satyanarayana Narlapuram <Satyanarayana.Narlapuram@microsoft.com>; PostgreSQL-development <pgsql-hackers@postgresql.org>
Subject: Re: [HACKERS] Reading backup label file for checkpoint and redo location during crash recovery

* Magnus Hagander (magnus@hagander.net) wrote:

On Mon, Sep 25, 2017 at 7:43 PM, Stephen Frost <sfrost@snowman.net> wrote:

* Satyanarayana Narlapuram (Satyanarayana.Narlapuram@microsoft.com) wrote:

During crash recovery, last checkpoint record information is
obtained

from the backup label if present, instead of getting it from the
control file. This behavior is causing PostgreSQL database cluster
not to come up until the backup label file is deleted (as the error message says).

if (checkPoint.redo < checkPointLoc)
{
if (!ReadRecord(xlogreader,

checkPoint.redo, LOG, false))

ereport(FATAL,
(errmsg("could
not

find redo location referenced by checkpoint record"),

errhint("If you
are

not restoring from a backup, try removing the file
\"%s/backup_label\".", DataDir)));

}

If we are recovering from a dump file, reading from the backup
label

files makes sense as the control file could be archived after a few
checkpoints. But this is not the case for crash recovery, and is
always safe to read the checkpoint record information from the control file.

Is this behavior kept this way as there is no clear way to
distinguish

between the recovery from the dump and the regular crash recovery?

This is why the exclusive backup method has been deprecated in PG10
in favor of the non-exclusive backup method, which avoids this by
not creating a backup label file (it's up to the backup software to
store the necessary information and create the file for use during recovery).

Actally, it was deprecated already in 9.6, not just 10.

Whoops, right. Thanks for the clarification. :)

Stephen

--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers