Thread: Using old master as new replica after clean switchover
Currently, the documentation explicitly states, that after failover, the old master must be recreated from scratch, or pg_rewind should be used (requiring wal_log_hints to be on, which is off by default):
> The former standby is now the primary, but the former primary is down and might stay down. To return to normal operation, a standby server must be recreated, either on the former primary system when it comes up, or on a third, possibly new, system. The pg_rewind utility can be used to speed up this process on large clusters.
Regards,
My research shows that some people already rely on the following when planned failover (aka switchover) procedure, doing it in production:
1) shutdown the current master
2) ensure that the "master candidate" replica has received all WAL data including shutdown checkpoint from the old master
3) promote the master candidate to make it new master
4) configure recovery.conf on the old master node, while it's inactive
5) start the old master node as a new replica following the new master.
It looks to me now, that if no steps missed in the procedure, this approach is eligible for Postgres versions 9.3+ (for older versions like 9.3 maybe not really always – people who know details better will correct me here maybe). Am I right? Or I'm missing some risks here?
Two changes were made in 9.3 which allowed this approach in general [1] [2]. Also, I see from the code [3] that during shutdown process, the walsenders are the last who are stopped, so allow replicas to get the shutdown checkpoint information.
Is this approach considered as safe now?
if so, let's add it to the documentation, making it official. The patch is attached.
Links:
[1] Support clean switchover https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=985bd7d49726c9f178558491d31a570d47340459
[2] Allow a streaming replication standby to follow a timeline switch https://git.postgresql.org/gitweb/?p=postgresql.git;a=commit;h=abfd192b1b5ba5216ac4b1f31dcd553106304b19
Regards,
Nik
Attachment
On Thu, 25 Oct 2018 02:57:18 -0400 Nikolay Samokhvalov <samokhvalov@gmail.com> wrote: ... > My research shows that some people already rely on the following when > planned failover (aka switchover) procedure, doing it in production: > > 1) shutdown the current master > 2) ensure that the "master candidate" replica has received all WAL data > including shutdown checkpoint from the old master > 3) promote the master candidate to make it new master > 4) configure recovery.conf on the old master node, while it's inactive > 5) start the old master node as a new replica following the new master. Indeed. > It looks to me now, that if no steps missed in the procedure, this approach > is eligible for Postgres versions 9.3+ (for older versions like 9.3 maybe > not really always – people who know details better will correct me here > maybe). Am I right? Or I'm missing some risks here? As far as I know, this is correct. > Two changes were made in 9.3 which allowed this approach in general [1] > [2]. Also, I see from the code [3] that during shutdown process, the > walsenders are the last who are stopped, so allow replicas to get the > shutdown checkpoint information. I had the same conclusions when I was studying controlled failover some years ago to implement it PAF project (allowing controlled switchover in one command). Here is a discussions around switchover taking place three years ago on Pacemaker mailing list: https://lists.clusterlabs.org/pipermail/users/2016-October/011568.html > Is this approach considered as safe now? Considering above points, I do think so. The only additional nice step would be to be able to run some more safety tests AFTER the switchover process on te old master. The only way I can think of would be to run pg_rewind even if it doesn't do much. > if so, let's add it to the documentation, making it official. The patch is > attached. I suppose we should add the technical steps in a sample procedure?
On Thu, 25 Oct 2018 02:57:18 -0400 Nikolay Samokhvalov <samokhvalov@gmail.com> wrote: ... > My research shows that some people already rely on the following when > planned failover (aka switchover) procedure, doing it in production: > > 1) shutdown the current master > 2) ensure that the "master candidate" replica has received all WAL data > including shutdown checkpoint from the old master > 3) promote the master candidate to make it new master > 4) configure recovery.conf on the old master node, while it's inactive > 5) start the old master node as a new replica following the new master. Indeed. > It looks to me now, that if no steps missed in the procedure, this approach > is eligible for Postgres versions 9.3+ (for older versions like 9.3 maybe > not really always – people who know details better will correct me here > maybe). Am I right? Or I'm missing some risks here? As far as I know, this is correct. > Two changes were made in 9.3 which allowed this approach in general [1] > [2]. Also, I see from the code [3] that during shutdown process, the > walsenders are the last who are stopped, so allow replicas to get the > shutdown checkpoint information. I had the same conclusions when I was studying controlled failover some years ago to implement it PAF project (allowing controlled switchover in one command). Here is a discussions around switchover taking place three years ago on Pacemaker mailing list: https://lists.clusterlabs.org/pipermail/users/2016-October/011568.html > Is this approach considered as safe now? Considering above points, I do think so. The only additional nice step would be to be able to run some more safety tests AFTER the switchover process on te old master. The only way I can think of would be to run pg_rewind even if it doesn't do much. > if so, let's add it to the documentation, making it official. The patch is > attached. I suppose we should add the technical steps in a sample procedure?
On Thu, Oct 25, 2018 at 11:15:51AM +0200, Jehan-Guillaume de Rorthais wrote: > On Thu, 25 Oct 2018 02:57:18 -0400 > Nikolay Samokhvalov <samokhvalov@gmail.com> wrote: >> My research shows that some people already rely on the following when >> planned failover (aka switchover) procedure, doing it in production: >> >> 1) shutdown the current master >> 2) ensure that the "master candidate" replica has received all WAL data >> including shutdown checkpoint from the old master >> 3) promote the master candidate to make it new master >> 4) configure recovery.conf on the old master node, while it's inactive >> 5) start the old master node as a new replica following the new master. > > Indeed. The important point here is that the primary will wait for the shutdown checkpoint record to be replayed on the standbys before finishing to shut down. > The only additional nice step would be to be able to run some more safety tests > AFTER the switchover process on te old master. The only way I can think of > would be to run pg_rewind even if it doesn't do much. Do you have something specific in mind here? I am curious if you're thinking about things like page-level checks for LSN matches under some threshold or such, because you should not have pages on the previous primary which have LSNs newer than the point up to which the standby has replayed. >> if so, let's add it to the documentation, making it official. The patch is >> attached. > > I suppose we should add the technical steps in a sample procedure? If an addition to the docs is done, symbolizing the steps in a list would be cleaner, with perhaps something in a dedicated section or a new sub-section. The failover flow you are mentioning is good practice because that's safe, and there is always room for improvements in the docs. -- Michael
On Thu, Oct 25, 2018 at 11:15:51AM +0200, Jehan-Guillaume de Rorthais wrote: > On Thu, 25 Oct 2018 02:57:18 -0400 > Nikolay Samokhvalov <samokhvalov@gmail.com> wrote: >> My research shows that some people already rely on the following when >> planned failover (aka switchover) procedure, doing it in production: >> >> 1) shutdown the current master >> 2) ensure that the "master candidate" replica has received all WAL data >> including shutdown checkpoint from the old master >> 3) promote the master candidate to make it new master >> 4) configure recovery.conf on the old master node, while it's inactive >> 5) start the old master node as a new replica following the new master. > > Indeed. The important point here is that the primary will wait for the shutdown checkpoint record to be replayed on the standbys before finishing to shut down. > The only additional nice step would be to be able to run some more safety tests > AFTER the switchover process on te old master. The only way I can think of > would be to run pg_rewind even if it doesn't do much. Do you have something specific in mind here? I am curious if you're thinking about things like page-level checks for LSN matches under some threshold or such, because you should not have pages on the previous primary which have LSNs newer than the point up to which the standby has replayed. >> if so, let's add it to the documentation, making it official. The patch is >> attached. > > I suppose we should add the technical steps in a sample procedure? If an addition to the docs is done, symbolizing the steps in a list would be cleaner, with perhaps something in a dedicated section or a new sub-section. The failover flow you are mentioning is good practice because that's safe, and there is always room for improvements in the docs. -- Michael
Attachment
On Thu, 25 Oct 2018 20:45:57 +0900 Michael Paquier <michael@paquier.xyz> wrote: > On Thu, Oct 25, 2018 at 11:15:51AM +0200, Jehan-Guillaume de Rorthais wrote: > > On Thu, 25 Oct 2018 02:57:18 -0400 > > Nikolay Samokhvalov <samokhvalov@gmail.com> wrote: > >> My research shows that some people already rely on the following when > >> planned failover (aka switchover) procedure, doing it in production: > >> > >> 1) shutdown the current master > >> 2) ensure that the "master candidate" replica has received all WAL data > >> including shutdown checkpoint from the old master > >> 3) promote the master candidate to make it new master > >> 4) configure recovery.conf on the old master node, while it's inactive > >> 5) start the old master node as a new replica following the new master. > > > > Indeed. > > The important point here is that the primary will wait for the shutdown > checkpoint record to be replayed on the standbys before finishing to > shut down. Yes. However, it gives up if the connection to the standby fails. This is obvious. But that's why we really need to double check on the standby the shutdown checkpoints has been received. Just in case of some network troubles or such. > > The only additional nice step would be to be able to run some more safety > > tests AFTER the switchover process on te old master. The only way I can > > think of would be to run pg_rewind even if it doesn't do much. > > Do you have something specific in mind here? I am curious if you're > thinking about things like page-level checks for LSN matches under some > threshold or such, because you should not have pages on the previous > primary which have LSNs newer than the point up to which the standby has > replayed. This could be a decent check. Heavy and slow, but safe. Other ideas I have (see bellow) are only related to ease the existing procedure. Both are interesting projects I could hopefully work on. > >> if so, let's add it to the documentation, making it official. The patch is > >> attached. > > > > I suppose we should add the technical steps in a sample procedure? > > If an addition to the docs is done, symbolizing the steps in a list > would be cleaner, with perhaps something in a dedicated section or a new > sub-section. The failover flow you are mentioning is good practice > because that's safe, and there is always room for improvements in the > docs. The hardest part to explain here is how to check the shutdown checkpoint hit the standby-to-promote. * in PAF, I'm using pg_waldump to check if the shutdown checkpoint has been received. * in manual operation, I force a checkpoint on the standby and compare "Latest checkpoint's REDO location" from the controldata file with the one on the old master. I'm not sure how to explain clearly one or the other method in the doc. Two ideas come in mind to improve this. What about logging the shutdown checkpoint on the old master? On the standby side, we could cross-check it with a function confirming: 1/ the very last XLogRecord received was the old master shutdown checkpoint 2/ the received shutdown checkpoint has the same LSN Second idea would be that an old master detect it has been started as a new standby and only replay XLogRecord from the new master if its TL fork is following its previous TL and shutdown checkpoint?
On Thu, 25 Oct 2018 20:45:57 +0900 Michael Paquier <michael@paquier.xyz> wrote: > On Thu, Oct 25, 2018 at 11:15:51AM +0200, Jehan-Guillaume de Rorthais wrote: > > On Thu, 25 Oct 2018 02:57:18 -0400 > > Nikolay Samokhvalov <samokhvalov@gmail.com> wrote: > >> My research shows that some people already rely on the following when > >> planned failover (aka switchover) procedure, doing it in production: > >> > >> 1) shutdown the current master > >> 2) ensure that the "master candidate" replica has received all WAL data > >> including shutdown checkpoint from the old master > >> 3) promote the master candidate to make it new master > >> 4) configure recovery.conf on the old master node, while it's inactive > >> 5) start the old master node as a new replica following the new master. > > > > Indeed. > > The important point here is that the primary will wait for the shutdown > checkpoint record to be replayed on the standbys before finishing to > shut down. Yes. However, it gives up if the connection to the standby fails. This is obvious. But that's why we really need to double check on the standby the shutdown checkpoints has been received. Just in case of some network troubles or such. > > The only additional nice step would be to be able to run some more safety > > tests AFTER the switchover process on te old master. The only way I can > > think of would be to run pg_rewind even if it doesn't do much. > > Do you have something specific in mind here? I am curious if you're > thinking about things like page-level checks for LSN matches under some > threshold or such, because you should not have pages on the previous > primary which have LSNs newer than the point up to which the standby has > replayed. This could be a decent check. Heavy and slow, but safe. Other ideas I have (see bellow) are only related to ease the existing procedure. Both are interesting projects I could hopefully work on. > >> if so, let's add it to the documentation, making it official. The patch is > >> attached. > > > > I suppose we should add the technical steps in a sample procedure? > > If an addition to the docs is done, symbolizing the steps in a list > would be cleaner, with perhaps something in a dedicated section or a new > sub-section. The failover flow you are mentioning is good practice > because that's safe, and there is always room for improvements in the > docs. The hardest part to explain here is how to check the shutdown checkpoint hit the standby-to-promote. * in PAF, I'm using pg_waldump to check if the shutdown checkpoint has been received. * in manual operation, I force a checkpoint on the standby and compare "Latest checkpoint's REDO location" from the controldata file with the one on the old master. I'm not sure how to explain clearly one or the other method in the doc. Two ideas come in mind to improve this. What about logging the shutdown checkpoint on the old master? On the standby side, we could cross-check it with a function confirming: 1/ the very last XLogRecord received was the old master shutdown checkpoint 2/ the received shutdown checkpoint has the same LSN Second idea would be that an old master detect it has been started as a new standby and only replay XLogRecord from the new master if its TL fork is following its previous TL and shutdown checkpoint?
On Thu, Oct 25, 2018 at 6:03 AM Jehan-Guillaume de Rorthais <jgdr@dalibo.com> wrote:
What about logging the shutdown checkpoint on the old master?
On the standby side, we could cross-check it with a function confirming:
1/ the very last XLogRecord received was the old master shutdown checkpoint
2/ the received shutdown checkpoint has the same LSN
Additionally, the new instructions in the doc might include recommendation, what to do if we
found that shutdown checkpoint wasn't received and replayed by the replica-to-promote. From my
understanding, before promotion, we could "manually" transfer missing WAL data from the old,
inactive master and replay it on the replica-to-promote (of course, if recovery_command is
properly configured on it). Right?
By the way, if it looks to me that it might be better to write more than just few sentences, what if it
will be a new chapter – say, "Switchover", next to "Failover". It would also give better understanding
to the reading, explicitly distinguishing planned and unplanned processes of master/replica role
changes.
Regards,
Nik
On Thu, Oct 25, 2018 at 6:03 AM Jehan-Guillaume de Rorthais <jgdr@dalibo.com> wrote:
What about logging the shutdown checkpoint on the old master?
On the standby side, we could cross-check it with a function confirming:
1/ the very last XLogRecord received was the old master shutdown checkpoint
2/ the received shutdown checkpoint has the same LSN
Additionally, the new instructions in the doc might include recommendation, what to do if we
found that shutdown checkpoint wasn't received and replayed by the replica-to-promote. From my
understanding, before promotion, we could "manually" transfer missing WAL data from the old,
inactive master and replay it on the replica-to-promote (of course, if recovery_command is
properly configured on it). Right?
By the way, if it looks to me that it might be better to write more than just few sentences, what if it
will be a new chapter – say, "Switchover", next to "Failover". It would also give better understanding
to the reading, explicitly distinguishing planned and unplanned processes of master/replica role
changes.
Regards,
Nik