Thread: BUG #11603: replication, pg_basebackup and high load
The following bug has been logged on the website: Bug reference: 11603 Logged by: Michiel Email address: mdglange@gmail.com PostgreSQL version: 9.4beta2 Operating system: Linux Description: The test I did involved the following: a master database with two slaves. On the master two replication slots have been configured as per the documentation. One slave active before I put some "heavy" load (the environment is scaled such, that inserting a few gigabytes of insert statements is a heavy load. This is on purpose) During this load I started a pg_basebackup from the master, immediately followed (pg_basebackup && mv recovery.conf $PG_DATA/ && pg_ctl start) by placing the prepared recovery.conf and starting this slave. The pg_basebackup took a few hours (as expected) but starting this latter slave would not work because the WALs were no longer available. I'd expect to see that pg_basebackup restores up to the last WAL, so that regardless of the load and changes done on the (master) database replication picks up.
On 10/08/2014 04:19 PM, mdglange@gmail.com wrote: > The following bug has been logged on the website: > > Bug reference: 11603 > Logged by: Michiel > Email address: mdglange@gmail.com > PostgreSQL version: 9.4beta2 > Operating system: Linux > Description: > > The test I did involved the following: a master database with two slaves. On > the master two replication slots have been configured as per the > documentation. One slave active before I put some "heavy" load (the > environment is scaled such, that inserting a few gigabytes of insert > statements is a heavy load. This is on purpose) > > During this load I started a pg_basebackup from the master, immediately > followed (pg_basebackup && mv recovery.conf $PG_DATA/ && pg_ctl start) by > placing the prepared recovery.conf and starting this slave. The > pg_basebackup took a few hours (as expected) but starting this latter slave > would not work because the WALs were no longer available. > > I'd expect to see that pg_basebackup restores up to the last WAL, so that > regardless of the load and changes done on the (master) database replication > picks up. After pg_basebackup, the system needs to have all the WAL available from the time the backup *started*, unfortunately. Try using pg_basebackup's "--xlog-method=stream" option. That way it streams the WAL at the same time the backup is taken, making it less likely that the master will recycle the segments too quickly. Even that is not bullet-proof, though; if there's a network hickup or something that makes the backup process to stall for long enough, the master might still recycle the segments that the backup would need. Or, set up WAL archiving, and use a restore_command in the recovery.conf file to pull the files from the archive. I didn't understand what replication slots have to do with this though; sorry if I misunderstood the whole thing.. - Heikki
On 2014-10-08 13:19:27 +0000, mdglange@gmail.com wrote: > The test I did involved the following: a master database with two slaves. On > the master two replication slots have been configured as per the > documentation. One slave active before I put some "heavy" load (the > environment is scaled such, that inserting a few gigabytes of insert > statements is a heavy load. This is on purpose) Replication slots currently only reserve resources after they've been used the first time. I.e. when you create a physical replication slot it doesn't immediately reserve resources - a client needs to connect to it once, telling it from where on to reserve resources. You can see the slot's reserved resources in the pg_replication_slots view. So, what you could do is to connect to the slots once, for a short time, using pg_receivexlog --slots. Or just use the -X stream method for pg_basebackup. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
On 10/09/2014 07:54 PM, Andres Freund wrote: > On 2014-10-08 13:19:27 +0000, mdglange@gmail.com wrote: >> The test I did involved the following: a master database with two slaves. On >> the master two replication slots have been configured as per the >> documentation. One slave active before I put some "heavy" load (the >> environment is scaled such, that inserting a few gigabytes of insert >> statements is a heavy load. This is on purpose) > > Replication slots currently only reserve resources after they've been > used the first time. I.e. when you create a physical replication slot it > doesn't immediately reserve resources - a client needs to connect to it > once, telling it from where on to reserve resources. Oh, now I understand what Michiel was trying to do with the replication slots. The idea is to prevent the master from recycling the segments while the backup runs, by creating a replication slot before the backup. > You can see the slot's reserved resources in the pg_replication_slots > view. > > So, what you could do is to connect to the slots once, for a short time, > using pg_receivexlog --slots. Or just use the -X stream method for > pg_basebackup. Hmm. Should we have an additional flag to "pg_basebackup -R" to create and "reserve" a replication slot, too, all in one command? That would be handy, although there's some potential of shooting your foot with replication slots; if the backup is aborted for some reason, the slot remains. - Heikki
On 2014-10-09 20:00:39 +0300, Heikki Linnakangas wrote: > On 10/09/2014 07:54 PM, Andres Freund wrote: > >On 2014-10-08 13:19:27 +0000, mdglange@gmail.com wrote: > >>The test I did involved the following: a master database with two slaves. On > >>the master two replication slots have been configured as per the > >>documentation. One slave active before I put some "heavy" load (the > >>environment is scaled such, that inserting a few gigabytes of insert > >>statements is a heavy load. This is on purpose) > > > >Replication slots currently only reserve resources after they've been > >used the first time. I.e. when you create a physical replication slot it > >doesn't immediately reserve resources - a client needs to connect to it > >once, telling it from where on to reserve resources. > > Oh, now I understand what Michiel was trying to do with the replication > slots. The idea is to prevent the master from recycling the segments while > the backup runs, by creating a replication slot before the backup. Yes, that's how I understand it. > >You can see the slot's reserved resources in the pg_replication_slots > >view. > > > >So, what you could do is to connect to the slots once, for a short time, > >using pg_receivexlog --slots. Or just use the -X stream method for > >pg_basebackup. > > Hmm. Should we have an additional flag to "pg_basebackup -R" to create and > "reserve" a replication slot, too, all in one command? That would be handy, > although there's some potential of shooting your foot with replication > slots; if the backup is aborted for some reason, the slot remains. I'd very much like to do that, but I think we need to build in some defenses against the footgun. I unfortunately ran out of time to implement it. There now exists the concept of a 'ephemeral' replication slot. Such slots are dropped upon release/error. I'm not entirely sure yet how to string that together with pg_basebackup, but I think it should be possible. The easiest way would probably to add a SLOT parameter to BASE_BACKUP that creates the slot, marks it as ephemeral for the duration, and only persists it at the end. The problem is that that still leaves a small window at the end when the server has finished the base backup, but the client fails before finishing. There's also the question how we want to make it cooperate with the forked process that receives WAL. It's a reasonable wish to want it to use the slot so resources are only reserved as much as necessary... Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Indeed, the idea of using slots is to prevent the master from recycling the WAL too early. The ephemeral slots might be the/a solution; the small window should be enough (depending on the definition of 'small', but that would depend on the load, I guess) I guess the use of replication slots is a great way to work with this, and I'm really happy to have seen it appear. It provides the solution to many of the issues I was still trying to work around. This is just 'it'. However, I think creating replication slots should also allow ways to delete those slots. I've been looking at that feature "select * from delete_replication_slot('slot_name'); :-) Or resetting, to tell the master that, based on this replication slot's informatin the master could cycle it's wal segments. I've not come around to search for solutions in that direction yet. What I do, is create the slot up front; I didn't think about connecting to that slot before the backup starts, just because there WAL segment to receive would be way ahead of where the new slave database actually is :-) How about that solution provided earlier: when pg_basebackup starts, it marks itself on the master to keep the wal segments and uses the slot provided as parameter on the replication slot parameter. Something like this: pg_basebackup <all those other params> --use-slot='slot_name'; Once the slave starts it will use the given slot, and the wal segments are still there. Still I think this new feature is a huge step forward, and I follow it with great interest. On Thu, Oct 9, 2014 at 7:10 PM, Andres Freund <andres@2ndquadrant.com> wrote: > On 2014-10-09 20:00:39 +0300, Heikki Linnakangas wrote: > > On 10/09/2014 07:54 PM, Andres Freund wrote: > > >On 2014-10-08 13:19:27 +0000, mdglange@gmail.com wrote: > > >>The test I did involved the following: a master database with two > slaves. On > > >>the master two replication slots have been configured as per the > > >>documentation. One slave active before I put some "heavy" load (the > > >>environment is scaled such, that inserting a few gigabytes of insert > > >>statements is a heavy load. This is on purpose) > > > > > >Replication slots currently only reserve resources after they've been > > >used the first time. I.e. when you create a physical replication slot it > > >doesn't immediately reserve resources - a client needs to connect to it > > >once, telling it from where on to reserve resources. > > > > Oh, now I understand what Michiel was trying to do with the replication > > slots. The idea is to prevent the master from recycling the segments > while > > the backup runs, by creating a replication slot before the backup. > > Yes, that's how I understand it. > > > >You can see the slot's reserved resources in the pg_replication_slots > > >view. > > > > > >So, what you could do is to connect to the slots once, for a short time, > > >using pg_receivexlog --slots. Or just use the -X stream method for > > >pg_basebackup. > > > > Hmm. Should we have an additional flag to "pg_basebackup -R" to create > and > > "reserve" a replication slot, too, all in one command? That would be > handy, > > although there's some potential of shooting your foot with replication > > slots; if the backup is aborted for some reason, the slot remains. > > I'd very much like to do that, but I think we need to build in some > defenses against the footgun. I unfortunately ran out of time to > implement it. > > There now exists the concept of a 'ephemeral' replication slot. Such > slots are dropped upon release/error. I'm not entirely sure yet how to > string that together with pg_basebackup, but I think it should be > possible. > > The easiest way would probably to add a SLOT parameter to BASE_BACKUP > that creates the slot, marks it as ephemeral for the duration, and only > persists it at the end. The problem is that that still leaves a small > window at the end when the server has finished the base backup, but the > client fails before finishing. > There's also the question how we want to make it cooperate with the > forked process that receives WAL. It's a reasonable wish to want it to > use the slot so resources are only reserved as much as necessary... > > Greetings, > > Andres Freund > > -- > Andres Freund http://www.2ndQuadrant.com/ > PostgreSQL Development, 24x7 Support, Training & Services > -- "Wouldn't the sentence 'I want to put a hyphen between the words Fish and And and And and Chips in my Fish-And-Chips sign' have been clearer if quotation marks had been placed before Fish, and between Fish and and, and and and And, and And and and, and and and And, and And and and, and and and Chips, as well as after Chips?"