Обсуждение: 9.4 logical replication - walsender keepalive replies
In 9.4 we've the below block of code to walsender.c as
/* * We only send regular messages to the client for full decoded * transactions, but a synchronous replication and
walsendershutdown * possibly are waiting for a later location. So we send pings * containing the flush location every
nowand then.
*/
if (MyWalSnd->flush < sentPtr && !waiting_for_ping_response)
{ WalSndKeepalive(true); waiting_for_ping_response = true;
}
I am finding that my logical replication reader is spending a tremendous
amount of time sending feedback to the server because a keep alive reply
was requested. My flush pointer is smaller than sendPtr, which I see as
the normal case (The client hasn't confirmed all the wal it has been
sent). My client queues the records it receives and only confirms when
actually processes the record.
So the sequence looks something like
Server Sends LSN 0/1000
Server Sends LSN 0/2000
Server Sends LSN 0/3000
Client confirms LSN 0/2000
I don't see why all these keep alive replies are needed in this case
(the timeout value is bumped way up, it's the above block that is
triggering the reply request not something related to timeout)
Steve
Hi Steve,
On 2014-06-30 11:40:50 -0400, Steve Singer wrote:
> In 9.4 we've the below block of code to walsender.c as
>
> /*
> * We only send regular messages to the client for full decoded
> * transactions, but a synchronous replication and walsender shutdown
> * possibly are waiting for a later location. So we send pings
> * containing the flush location every now and then.
> */
> if (MyWalSnd->flush < sentPtr && !waiting_for_ping_response)
> {
> WalSndKeepalive(true);
> waiting_for_ping_response = true;
> }
>
>
> I am finding that my logical replication reader is spending a tremendous
> amount of time sending feedback to the server because a keep alive reply was
> requested. My flush pointer is smaller than sendPtr, which I see as the
> normal case (The client hasn't confirmed all the wal it has been sent). My
> client queues the records it receives and only confirms when actually
> processes the record.
>
> So the sequence looks something like
>
> Server Sends LSN 0/1000
> Server Sends LSN 0/2000
> Server Sends LSN 0/3000
> Client confirms LSN 0/2000
> I don't see why all these keep alive replies are needed in this case (the
> timeout value is bumped way up, it's the above block that is triggering the
> reply request not something related to timeout)
Right. I thought about this for a while, and I think we should change
two things. For one, don't request replies here. It's simply not needed,
as this isn't dealing with timeouts. For another don't just check ->flush
< sentPtr but also && ->write < sentPtr. The reason we're sending these
feedback messages is to inform the 'logical standby' that there's been
WAL activity which it can't see because they don't correspond to
anything that's logically decoded (e.g. vacuum stuff).
Would that suit your needs?
Greetings,
Andres Freund
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services
On 07/06/2014 10:11 AM, Andres Freund wrote: > Hi Steve, > Right. I thought about this for a while, and I think we should change > two things. For one, don't request replies here. It's simply not needed, > as this isn't dealing with timeouts. For another don't just check ->flush > < sentPtr but also && ->write < sentPtr. The reason we're sending these > feedback messages is to inform the 'logical standby' that there's been > WAL activity which it can't see because they don't correspond to > anything that's logically decoded (e.g. vacuum stuff). > Would that suit your needs? > > Greetings, Yes I think that will work for me. I tested with the attached patch that I think does what you describe and it seems okay. > Andres Freund >
Вложения
On 07/14/2014 01:19 PM, Steve Singer wrote: > On 07/06/2014 10:11 AM, Andres Freund wrote: >> Hi Steve, > >> Right. I thought about this for a while, and I think we should change >> two things. For one, don't request replies here. It's simply not needed, >> as this isn't dealing with timeouts. For another don't just check >> ->flush >> < sentPtr but also && ->write < sentPtr. The reason we're sending these >> feedback messages is to inform the 'logical standby' that there's been >> WAL activity which it can't see because they don't correspond to >> anything that's logically decoded (e.g. vacuum stuff). >> Would that suit your needs? >> >> Greetings, > > Yes I think that will work for me. > I tested with the attached patch that I think does what you describe > and it seems okay. > > Any feedback on this? Do we want that change for 9.4, or do we want something else? > >> Andres Freund >> > > >
On 2014-08-11 17:22:27 -0400, Steve Singer wrote: > On 07/14/2014 01:19 PM, Steve Singer wrote: > >On 07/06/2014 10:11 AM, Andres Freund wrote: > >>Hi Steve, > > > >>Right. I thought about this for a while, and I think we should change > >>two things. For one, don't request replies here. It's simply not needed, > >>as this isn't dealing with timeouts. For another don't just check > >>->flush > >>< sentPtr but also && ->write < sentPtr. The reason we're sending these > >>feedback messages is to inform the 'logical standby' that there's been > >>WAL activity which it can't see because they don't correspond to > >>anything that's logically decoded (e.g. vacuum stuff). > >>Would that suit your needs? > >> > >>Greetings, > > > >Yes I think that will work for me. > >I tested with the attached patch that I think does what you describe and > >it seems okay. > > > > > > > Any feedback on this? Do we want that change for 9.4, or do we want > something else? I plan to test and apply it in the next few days. Digging myself from under stuff from before my holiday right now... Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 2014-08-11 23:52:32 +0200, Andres Freund wrote: > On 2014-08-11 17:22:27 -0400, Steve Singer wrote: > > On 07/14/2014 01:19 PM, Steve Singer wrote: > > >On 07/06/2014 10:11 AM, Andres Freund wrote: > > >>Hi Steve, > > > > > >>Right. I thought about this for a while, and I think we should change > > >>two things. For one, don't request replies here. It's simply not needed, > > >>as this isn't dealing with timeouts. For another don't just check > > >>->flush > > >>< sentPtr but also && ->write < sentPtr. The reason we're sending these > > >>feedback messages is to inform the 'logical standby' that there's been > > >>WAL activity which it can't see because they don't correspond to > > >>anything that's logically decoded (e.g. vacuum stuff). > > >>Would that suit your needs? > > >> > > >>Greetings, > > > > > >Yes I think that will work for me. > > >I tested with the attached patch that I think does what you describe and > > >it seems okay. > > > > > > > > > > > > Any feedback on this? Do we want that change for 9.4, or do we want > > something else? > > I plan to test and apply it in the next few days. Digging myself from > under stuff from before my holiday right now... Committed. Thanks and sorry for the delay. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services