Обсуждение: Primary not sending to synchronous standby
Hi,
I've noticed that if the primary is started and then a base backup is immediately taken from it and started as as a synchronous standby, it doesn't replicate and the primary hangs indefinitely when trying to run any WAL-generating statements. It only recovers when either the primary is restarted (which has to use a fast shutdown otherwise it also hangs forever), or the standby is restarted.Here's a way of reproducing it:
-------------------------------
mkdir -p -m 0700 primary standby1
initdb -N -k -D primary -E 'UTF8'
-------------------------------
mkdir -p -m 0700 primary standby1
initdb -N -k -D primary -E 'UTF8'
cat << PRIMARYCONFIG >> primary/postgresql.conf
shared_buffers = 8MB
logging_collector = on
log_line_prefix = '%m - %u - %d'
synchronous_standby_names = 'standby1'
max_connections = 8
wal_level = 'hot_standby'
port = 5530
max_wal_senders = 3
wal_keep_segments = 6
shared_buffers = 8MB
logging_collector = on
log_line_prefix = '%m - %u - %d'
synchronous_standby_names = 'standby1'
max_connections = 8
wal_level = 'hot_standby'
port = 5530
max_wal_senders = 3
wal_keep_segments = 6
PRIMARYCONFIG
cat << PRIMARYHBA >> primary/pg_hba.conf
local replication rep_user trust
host replication rep_user 127.0.0.1/32 trust
host replication rep_user ::1/128 trust
PRIMARYHBA
local replication rep_user trust
host replication rep_user 127.0.0.1/32 trust
host replication rep_user ::1/128 trust
PRIMARYHBA
pg_ctl start -D primary
psql -p 5530 -h localhost -c 'SET SESSION synchronous_commit TO 'off';CREATE USER rep_user REPLICATION;;' -d postgres
pg_basebackup -x -D standby1 -h localhost -p 5530 -U rep_user
port = 5531
hot_standby = on
STANDBYCONFIG
hot_standby = on
STANDBYCONFIG
cat << STANDBYRECOVERY >> standby1/recovery.conf
standby_mode = 'on'
recovery_target_timeline = 'latest'
primary_conninfo = 'host=127.0.0.1 user=rep_user port=5530 application_name=standby1'
STANDBYRECOVERY
standby_mode = 'on'
recovery_target_timeline = 'latest'
primary_conninfo = 'host=127.0.0.1 user=rep_user port=5530 application_name=standby1'
STANDBYRECOVERY
pg_ctl -D standby1 start
-------------------------------
-------------------------------
Note that if you run the commands one by one, there isn't a problem. If you run it as a script, the standby doesn't connect to the primary. There aren't any errors reported by either the standby or the primary. The primary's wal sender process reports the following:
wal sender process rep_user 127.0.0.1(45243) startup waiting for 0/3000158
wal sender process rep_user 127.0.0.1(45243) startup waiting for 0/3000158
Anyone know why this would be happening? And if this could be a problem in other scenarios?
Thom
Hi, On 2015-02-23 15:25:57 +0000, Thom Brown wrote: > I've noticed that if the primary is started and then a base backup is > immediately taken from it and started as as a synchronous standby, it > doesn't replicate and the primary hangs indefinitely when trying to run any > WAL-generating statements. It only recovers when either the primary is > restarted (which has to use a fast shutdown otherwise it also hangs > forever), or the standby is restarted. > > Here's a way of reproducing it: > ... > Note that if you run the commands one by one, there isn't a problem. If > you run it as a script, the standby doesn't connect to the primary. There > aren't any errors reported by either the standby or the primary. The > primary's wal sender process reports the following: > > wal sender process rep_user 127.0.0.1(45243) startup waiting for 0/3000158 > > Anyone know why this would be happening? And if this could be a problem in > other scenarios? Given that normally a walsender doesn't wait for syncrep I guess this is the above backend just did authentication. If you gdb into the walsender, what's the backtrace? We previously had discussions about that being rather annoying; I unfortunately don't remember enough of the thread to reference it here. If it really is this, I think we should add some more smarts about only enabling syncrep once a backend is fully up and maybe even remove it from more scenarios during commits generally (e.g. if no xid was assigned and we just pruned something). Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 2015-02-23 16:38:44 +0100, Andres Freund wrote: > I unfortunately don't remember enough of the thread to reference it > here. Found the right keywords. The threads below http://archives.postgresql.org/message-id/369698E947874884A77849D8FE3680C2%40maumau and http://www.postgresql.org/message-id/5CF4ABBA67674088B3941894E22A0D25@maumau Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 23 February 2015 at 15:38, Andres Freund <andres@2ndquadrant.com> wrote:
#0 0x00007f66d1725940 in poll () from /lib/x86_64-linux-gnu/libc.so.6
#1 0x0000000000617faa in WaitLatchOrSocket ()
#2 0x000000000064741b in SyncRepWaitForLSN ()
#3 0x00000000004bbf8f in CommitTransaction ()
#4 0x00000000004be135 in CommitTransactionCommand ()
#5 0x0000000000757679 in InitPostgres ()
#6 0x0000000000675032 in PostgresMain ()
#7 0x00000000004617ef in ServerLoop ()
#8 0x0000000000627c9c in PostmasterMain ()
#9 0x000000000046223d in main ()
-- Hi,
On 2015-02-23 15:25:57 +0000, Thom Brown wrote:
> I've noticed that if the primary is started and then a base backup is
> immediately taken from it and started as as a synchronous standby, it
> doesn't replicate and the primary hangs indefinitely when trying to run any
> WAL-generating statements. It only recovers when either the primary is
> restarted (which has to use a fast shutdown otherwise it also hangs
> forever), or the standby is restarted.
>
> Here's a way of reproducing it:
> ...
> Note that if you run the commands one by one, there isn't a problem. If
> you run it as a script, the standby doesn't connect to the primary. There
> aren't any errors reported by either the standby or the primary. The
> primary's wal sender process reports the following:
>
> wal sender process rep_user 127.0.0.1(45243) startup waiting for 0/3000158
>
> Anyone know why this would be happening? And if this could be a problem in
> other scenarios?
Given that normally a walsender doesn't wait for syncrep I guess this is
the above backend just did authentication. If you gdb into the
walsender, what's the backtrace?
#0 0x00007f66d1725940 in poll () from /lib/x86_64-linux-gnu/libc.so.6
#1 0x0000000000617faa in WaitLatchOrSocket ()
#2 0x000000000064741b in SyncRepWaitForLSN ()
#3 0x00000000004bbf8f in CommitTransaction ()
#4 0x00000000004be135 in CommitTransactionCommand ()
#5 0x0000000000757679 in InitPostgres ()
#6 0x0000000000675032 in PostgresMain ()
#7 0x00000000004617ef in ServerLoop ()
#8 0x0000000000627c9c in PostmasterMain ()
#9 0x000000000046223d in main ()
Thom
On 23 February 2015 at 15:42, Andres Freund <andres@2ndquadrant.com> wrote:
On 2015-02-23 16:38:44 +0100, Andres Freund wrote:
> I unfortunately don't remember enough of the thread to reference it
> here.
Found the right keywords. The threads below
http://archives.postgresql.org/message-id/369698E947874884A77849D8FE3680C2%40maumau
and
http://www.postgresql.org/message-id/5CF4ABBA67674088B3941894E22A0D25@maumau
Yes, this seems to be virtually the same issue reported. The trace looks the same except for RecordTransactionCommit.
--
Thom
On 2015-02-23 15:48:25 +0000, Thom Brown wrote: > On 23 February 2015 at 15:42, Andres Freund <andres@2ndquadrant.com> wrote: > > > On 2015-02-23 16:38:44 +0100, Andres Freund wrote: > > > I unfortunately don't remember enough of the thread to reference it > > > here. > > > > Found the right keywords. The threads below > > > > http://archives.postgresql.org/message-id/369698E947874884A77849D8FE3680C2%40maumau > > and > > > > http://www.postgresql.org/message-id/5CF4ABBA67674088B3941894E22A0D25@maumau > > > > Yes, this seems to be virtually the same issue reported. The trace looks > the same except for RecordTransactionCommit. So, I proposed in http://www.postgresql.org/message-id/20140707155113.GB1136@alap3.anarazel.de that we make sequences assign a xid and only wait for syncrep when a xid is assigned. The biggest blocker was that somebody would have to do some code reviewing to find other locations that might need similar treatment. I did a, quick, grep for XLogInsert() and I think we're otherwise fine. There's some debatable cases: * XLOG_STANDBY_LOCK doesn't force a xid to be assigned. I think it's harmless though, as we really only need to wait for that to be replicated if the transaction did something relevant (i.e. catalog changes). And those will force xid assignment. * 2pc records don't assign a xid. But twophase.c does it's own waiting, so that's fine. * Plain vacuums will not trigger waits. But I think that's good. There's really no need to wait if all that's been done is some cleanup without visible consequences. * Fujii brought up that we might want to wait for XLOG_SWITCH - I don't really see why. * XLOG_RESTORE_POINT is a similar candidate - I don't see really valid arguments for making 2pc wait. The attached, untested, patch changes things so that we a) only wait for syncrep if we both wrote WAL and had a xid assigned b) use an async commit if we just had a xid assigned, without having written WAL, even if synchronous_commit = off c) acquire a xid when WAL logging sequence changes (arguable at least one of the xid assignments is redundant, but it doesn't cost anything, so ...) I think it makes sense to change a) and b) that way because there's no need to wait for WAL flushes/syncrep waits when all that happened is manipulations of temporary/unlogged tables or HOT pruning. It's slightly wierd that the on-disk flush and the syncrep wait essentially used two different mechanisms for deciding when to flush. Comments? This is obviously just a POC, but I think something like this does make a great deal of sense. Thom, does that help? Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services
Вложения
On 23 February 2015 at 16:53, Andres Freund <andres@2ndquadrant.com> wrote:
So, I proposed inOn 2015-02-23 15:48:25 +0000, Thom Brown wrote:
> On 23 February 2015 at 15:42, Andres Freund <andres@2ndquadrant.com> wrote:
>
> > On 2015-02-23 16:38:44 +0100, Andres Freund wrote:
> > > I unfortunately don't remember enough of the thread to reference it
> > > here.
> >
> > Found the right keywords. The threads below
> >
> > http://archives.postgresql.org/message-id/369698E947874884A77849D8FE3680C2%40maumau
> > and
> >
> > http://www.postgresql.org/message-id/5CF4ABBA67674088B3941894E22A0D25@maumau
> >
>
> Yes, this seems to be virtually the same issue reported. The trace looks
> the same except for RecordTransactionCommit.
http://www.postgresql.org/message-id/20140707155113.GB1136@alap3.anarazel.de
that we make sequences assign a xid and only wait for syncrep when a xid
is assigned. The biggest blocker was that somebody would have to do some
code reviewing to find other locations that might need similar
treatment.
I did a, quick, grep for XLogInsert() and I think we're otherwise
fine. There's some debatable cases:
* XLOG_STANDBY_LOCK doesn't force a xid to be assigned. I think it's
harmless though, as we really only need to wait for that to be
replicated if the transaction did something relevant (i.e. catalog
changes). And those will force xid assignment.
* 2pc records don't assign a xid. But twophase.c does it's own waiting,
so that's fine.
* Plain vacuums will not trigger waits. But I think that's good. There's
really no need to wait if all that's been done is some cleanup without
visible consequences.
* Fujii brought up that we might want to wait for XLOG_SWITCH - I don't
really see why.
* XLOG_RESTORE_POINT is a similar candidate - I don't see really valid
arguments for making 2pc wait.
The attached, untested, patch changes things so that we
a) only wait for syncrep if we both wrote WAL and had a xid assigned
b) use an async commit if we just had a xid assigned, without having
written WAL, even if synchronous_commit = off
c) acquire a xid when WAL logging sequence changes (arguable at least
one of the xid assignments is redundant, but it doesn't cost
anything, so ...)
I think it makes sense to change a) and b) that way because there's no
need to wait for WAL flushes/syncrep waits when all that happened is
manipulations of temporary/unlogged tables or HOT pruning. It's slightly
wierd that the on-disk flush and the syncrep wait essentially used two
different mechanisms for deciding when to flush.
Comments? This is obviously just a POC, but I think something like this
does make a great deal of sense.
Thom, does that help?
Yeah, this appears to eliminate the problem, at least in the case I reported.
Thanks
--
Thom
On 2015-02-23 17:53:59 +0100, Andres Freund wrote: > On 2015-02-23 15:48:25 +0000, Thom Brown wrote: > > On 23 February 2015 at 15:42, Andres Freund <andres@2ndquadrant.com> wrote: > > > > > On 2015-02-23 16:38:44 +0100, Andres Freund wrote: > > > > I unfortunately don't remember enough of the thread to reference it > > > > here. > > > > > > Found the right keywords. The threads below > > > > > > http://archives.postgresql.org/message-id/369698E947874884A77849D8FE3680C2%40maumau > > > and > > > > > > http://www.postgresql.org/message-id/5CF4ABBA67674088B3941894E22A0D25@maumau > > > > > > > Yes, this seems to be virtually the same issue reported. The trace looks > > the same except for RecordTransactionCommit. > > So, I proposed in > http://www.postgresql.org/message-id/20140707155113.GB1136@alap3.anarazel.de > that we make sequences assign a xid and only wait for syncrep when a xid > is assigned. The biggest blocker was that somebody would have to do some > code reviewing to find other locations that might need similar > treatment. > > I did a, quick, grep for XLogInsert() and I think we're otherwise > fine. There's some debatable cases: > > * XLOG_STANDBY_LOCK doesn't force a xid to be assigned. I think it's > harmless though, as we really only need to wait for that to be > replicated if the transaction did something relevant (i.e. catalog > changes). And those will force xid assignment. > * 2pc records don't assign a xid. But twophase.c does it's own waiting, > so that's fine. > * Plain vacuums will not trigger waits. But I think that's good. There's > really no need to wait if all that's been done is some cleanup without > visible consequences. > * Fujii brought up that we might want to wait for XLOG_SWITCH - I don't > really see why. > * XLOG_RESTORE_POINT is a similar candidate - I don't see really valid > arguments for making 2pc wait. > > > The attached, untested, patch changes things so that we > a) only wait for syncrep if we both wrote WAL and had a xid assigned > b) use an async commit if we just had a xid assigned, without having > written WAL, even if synchronous_commit = off > c) acquire a xid when WAL logging sequence changes (arguable at least > one of the xid assignments is redundant, but it doesn't cost > anything, so ...) > > I think it makes sense to change a) and b) that way because there's no > need to wait for WAL flushes/syncrep waits when all that happened is > manipulations of temporary/unlogged tables or HOT pruning. It's slightly > wierd that the on-disk flush and the syncrep wait essentially used two > different mechanisms for deciding when to flush. I think this patch is a clear improvement - and unless somebody protests I'm planning to push it sometime not too far away. What I'm wondering about is what we want to do with the back branches: It's really rather ugly that enabling syncrep can prevent you from logging in; especially if the connection that's blocked is the syncrep walsender. It'll resolve itself if you kill the connections often enough, but still, that's something you don't wan to have to do. The patch isn't particularly complex and should only make things efficient, so it'd be defensible to backpatch it. Thoughts? Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 2015-02-23 17:09:24 +0000, Thom Brown wrote: > On 23 February 2015 at 16:53, Andres Freund <andres@2ndquadrant.com> wrote: > > Comments? This is obviously just a POC, but I think something like this > > does make a great deal of sense. > > > > Thom, does that help? > Yeah, this appears to eliminate the problem, at least in the case I > reported. I've pushed a somewhat more evolved version of this after more testing. Greetings, Andres Freund --Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training & Services
On 26 February 2015 at 13:08, Andres Freund <andres@2ndquadrant.com> wrote:
On 2015-02-23 17:09:24 +0000, Thom Brown wrote:
> On 23 February 2015 at 16:53, Andres Freund <andres@2ndquadrant.com> wrote:
> > Comments? This is obviously just a POC, but I think something like this
> > does make a great deal of sense.
> >
> > Thom, does that help?
> Yeah, this appears to eliminate the problem, at least in the case I
> reported.
I've pushed a somewhat more evolved version of this after more testing.
Thanks. I'll give it another round of testing later.
--
Thom