Thread: Some memory not freed at the exit of RelationBuildPartitionDesc()

Some memory not freed at the exit of RelationBuildPartitionDesc()

From
amul sul
Date:
Hi,

In RelationBuildPartitionDesc(), a memory space that use to gather partitioning
bound info wasn't free at the end.  This might not a problem because this
allocated memory will eventually be recovered when the top-level context is
freed, but the case when a partitioned table having 1000s or more partitions and
this partitioned relation open & close, and its cached entry invalidated in loop
then we'll have too may call to RelationBuildPartitionDesc() which will keep
wasting some space with every loop. 

For a demonstration purpose, I did the following changes to
heap_drop_with_catalog() and tried to drop a partitioned table having 5000
partitions(attached create script) which hit OOM on a machine in no time:

diff --git a/src/backend/catalog/heap.c b/src/backend/catalog/heap.c
index b7bcdd9d0f..6b7bc0d7ae 100644
--- a/src/backend/catalog/heap.c
+++ b/src/backend/catalog/heap.c
@@ -1842,6 +1842,8 @@ heap_drop_with_catalog(Oid relid)
        parentOid = get_partition_parent(relid);
        LockRelationOid(parentOid, AccessExclusiveLock);
 
+       rel = relation_open(parentOid, NoLock);
+       relation_close(rel, NoLock);
        /*
         * If this is not the default partition, dropping it will change the
         * default partition's partition constraint, so we must lock it.



I think we should do all partitioned bound information gathering and
calculation in temporary memory context which can be released at the end of
RelationBuildPartitionDesc(), thoughts/Comments?  

I did the same in the attached patch and the aforesaid OOM issue is disappeared.

Regards,
Amul
Attachment

Re: Some memory not freed at the exit of RelationBuildPartitionDesc()

From
Amit Langote
Date:
Hi Amul,

On Thu, Aug 8, 2019 at 4:15 PM amul sul <sulamul@gmail.com> wrote:
>
> Hi,
>
> In RelationBuildPartitionDesc(), a memory space that use to gather partitioning
> bound info wasn't free at the end.  This might not a problem because this
> allocated memory will eventually be recovered when the top-level context is
> freed, but the case when a partitioned table having 1000s or more partitions and
> this partitioned relation open & close, and its cached entry invalidated in loop
> then we'll have too may call to RelationBuildPartitionDesc() which will keep
> wasting some space with every loop.
>
> For a demonstration purpose, I did the following changes to
> heap_drop_with_catalog() and tried to drop a partitioned table having 5000
> partitions(attached create script) which hit OOM on a machine in no time:
>
> diff --git a/src/backend/catalog/heap.c b/src/backend/catalog/heap.c
> index b7bcdd9d0f..6b7bc0d7ae 100644
> --- a/src/backend/catalog/heap.c
> +++ b/src/backend/catalog/heap.c
> @@ -1842,6 +1842,8 @@ heap_drop_with_catalog(Oid relid)
>         parentOid = get_partition_parent(relid);
>         LockRelationOid(parentOid, AccessExclusiveLock);
>
> +       rel = relation_open(parentOid, NoLock);
> +       relation_close(rel, NoLock);
>         /*
>          * If this is not the default partition, dropping it will change the
>          * default partition's partition constraint, so we must lock it.
>
>
> I think we should do all partitioned bound information gathering and
> calculation in temporary memory context which can be released at the end of
> RelationBuildPartitionDesc(), thoughts/Comments?
>
> I did the same in the attached patch and the aforesaid OOM issue is disappeared.

Thanks for the patch.  This was discussed recently in the "hyrax vs.
RelationBuildPartitionDesc()" thread [1] and I think Alvaro proposed
an approach that's similar to yours.  Not sure why it wasn't pursued
though.  Maybe the reason is buried somewhere in that discussion.

Thanks,
Amit

[1] https://www.postgresql.org/message-id/flat/CA%2BTgmoY3bRmGB6-DUnoVy5fJoreiBJ43rwMrQRCdPXuKt4Ykaw%40mail.gmail.com



Re: Some memory not freed at the exit of RelationBuildPartitionDesc()

From
amul sul
Date:


On Thu, Aug 8, 2019 at 1:27 PM Amit Langote <amitlangote09@gmail.com> wrote:
Hi Amul,

On Thu, Aug 8, 2019 at 4:15 PM amul sul <sulamul@gmail.com> wrote:
>
> Hi,
>
> In RelationBuildPartitionDesc(), a memory space that use to gather partitioning
> bound info wasn't free at the end.  This might not a problem because this
> allocated memory will eventually be recovered when the top-level context is
> freed, but the case when a partitioned table having 1000s or more partitions and
> this partitioned relation open & close, and its cached entry invalidated in loop
> then we'll have too may call to RelationBuildPartitionDesc() which will keep
> wasting some space with every loop.
>
> For a demonstration purpose, I did the following changes to
> heap_drop_with_catalog() and tried to drop a partitioned table having 5000
> partitions(attached create script) which hit OOM on a machine in no time:
>
> diff --git a/src/backend/catalog/heap.c b/src/backend/catalog/heap.c
> index b7bcdd9d0f..6b7bc0d7ae 100644
> --- a/src/backend/catalog/heap.c
> +++ b/src/backend/catalog/heap.c
> @@ -1842,6 +1842,8 @@ heap_drop_with_catalog(Oid relid)
>         parentOid = get_partition_parent(relid);
>         LockRelationOid(parentOid, AccessExclusiveLock);
>
> +       rel = relation_open(parentOid, NoLock);
> +       relation_close(rel, NoLock);
>         /*
>          * If this is not the default partition, dropping it will change the
>          * default partition's partition constraint, so we must lock it.
>
>
> I think we should do all partitioned bound information gathering and
> calculation in temporary memory context which can be released at the end of
> RelationBuildPartitionDesc(), thoughts/Comments?
>
> I did the same in the attached patch and the aforesaid OOM issue is disappeared.

Thanks for the patch.  This was discussed recently in the "hyrax vs.
RelationBuildPartitionDesc()" thread [1] and I think Alvaro proposed
an approach that's similar to yours.  Not sure why it wasn't pursued
though.  Maybe the reason is buried somewhere in that discussion.

Thanks,
Amit

[1] https://www.postgresql.org/message-id/flat/CA%2BTgmoY3bRmGB6-DUnoVy5fJoreiBJ43rwMrQRCdPXuKt4Ykaw%40mail.gmail.com

Oh, quite similar, thanks Amit for pointing that out.

Look like "hyrax vs.RelationBuildPartitionDesc()" is in discussion for the master
branch only, not sure though, but we need the similar fix for the back branches as well.

Regards,
Amul
 

Re: Some memory not freed at the exit of RelationBuildPartitionDesc()

From
Amit Langote
Date:
On Thu, Aug 8, 2019 at 5:33 PM amul sul <sulamul@gmail.com> wrote:
> On Thu, Aug 8, 2019 at 1:27 PM Amit Langote <amitlangote09@gmail.com> wrote:

>> Thanks for the patch.  This was discussed recently in the "hyrax vs.
>> RelationBuildPartitionDesc()" thread [1] and I think Alvaro proposed
>> an approach that's similar to yours.  Not sure why it wasn't pursued
>> though.  Maybe the reason is buried somewhere in that discussion.
>
> Oh, quite similar, thanks Amit for pointing that out.
>
> Look like "hyrax vs.RelationBuildPartitionDesc()" is in discussion for the master
> branch only, not sure though, but we need the similar fix for the back branches as well.

Well, this is not a bug as such, so it's very unlikely that a fix like
this will be back-patched.  Also, if this becomes an issue only for
more than over 1000 partitions, then it's not very relevant for PG 10
and PG 11, because we don't recommend using so many partitions with
them.  Maybe we can consider fixing PG 12 though.

Thanks,
Amit



Re: Some memory not freed at the exit of RelationBuildPartitionDesc()

From
David Fetter
Date:
On Thu, Aug 08, 2019 at 05:42:21PM +0900, Amit Langote wrote:
> On Thu, Aug 8, 2019 at 5:33 PM amul sul <sulamul@gmail.com> wrote:
> > On Thu, Aug 8, 2019 at 1:27 PM Amit Langote <amitlangote09@gmail.com> wrote:
> 
> >> Thanks for the patch.  This was discussed recently in the "hyrax vs.
> >> RelationBuildPartitionDesc()" thread [1] and I think Alvaro proposed
> >> an approach that's similar to yours.  Not sure why it wasn't pursued
> >> though.  Maybe the reason is buried somewhere in that discussion.
> >
> > Oh, quite similar, thanks Amit for pointing that out.
> >
> > Look like "hyrax vs.RelationBuildPartitionDesc()" is in discussion for the master
> > branch only, not sure though, but we need the similar fix for the back branches as well.
> 
> Well, this is not a bug as such, so it's very unlikely that a fix like
> this will be back-patched.  Also, if this becomes an issue only for
> more than over 1000 partitions, then it's not very relevant for PG 10
> and PG 11, because we don't recommend using so many partitions with
> them.  Maybe we can consider fixing PG 12 though.

A fix for the thousands-of-partitions case would be very welcome for 12.

Best,
David.
-- 
David Fetter <david(at)fetter(dot)org> http://fetter.org/
Phone: +1 415 235 3778

Remember to vote!
Consider donating to Postgres: http://www.postgresql.org/about/donate



Re: Some memory not freed at the exit of RelationBuildPartitionDesc()

From
amul sul
Date:


On Sat, Aug 10, 2019 at 12:16 AM David Fetter <david@fetter.org> wrote:
On Thu, Aug 08, 2019 at 05:42:21PM +0900, Amit Langote wrote:
> On Thu, Aug 8, 2019 at 5:33 PM amul sul <sulamul@gmail.com> wrote:
> > On Thu, Aug 8, 2019 at 1:27 PM Amit Langote <amitlangote09@gmail.com> wrote:
>
> >> Thanks for the patch.  This was discussed recently in the "hyrax vs.
> >> RelationBuildPartitionDesc()" thread [1] and I think Alvaro proposed
> >> an approach that's similar to yours.  Not sure why it wasn't pursued
> >> though.  Maybe the reason is buried somewhere in that discussion.
> >
> > Oh, quite similar, thanks Amit for pointing that out.
> >
> > Look like "hyrax vs.RelationBuildPartitionDesc()" is in discussion for the master
> > branch only, not sure though, but we need the similar fix for the back branches as well.
>
> Well, this is not a bug as such, so it's very unlikely that a fix like
> this will be back-patched.  Also, if this becomes an issue only for
> more than over 1000 partitions, then it's not very relevant for PG 10
> and PG 11, because we don't recommend using so many partitions with
> them.  Maybe we can consider fixing PG 12 though.

A fix for the thousands-of-partitions case would be very welcome for 12.


Look like commit # d3f48dfae42 added the required fix but is enabled only for
the clobber-cache builds :(

Regards,
Amul


Re: Some memory not freed at the exit of RelationBuildPartitionDesc()

From
David Fetter
Date:
On Tue, Aug 13, 2019 at 03:08:34PM +0530, amul sul wrote:
> On Sat, Aug 10, 2019 at 12:16 AM David Fetter <david@fetter.org> wrote:
> 
> > On Thu, Aug 08, 2019 at 05:42:21PM +0900, Amit Langote wrote:
> > > On Thu, Aug 8, 2019 at 5:33 PM amul sul <sulamul@gmail.com> wrote:
> > > > On Thu, Aug 8, 2019 at 1:27 PM Amit Langote <amitlangote09@gmail.com>
> > wrote:
> > >
> > > >> Thanks for the patch.  This was discussed recently in the "hyrax vs.
> > > >> RelationBuildPartitionDesc()" thread [1] and I think Alvaro proposed
> > > >> an approach that's similar to yours.  Not sure why it wasn't pursued
> > > >> though.  Maybe the reason is buried somewhere in that discussion.
> > > >
> > > > Oh, quite similar, thanks Amit for pointing that out.
> > > >
> > > > Look like "hyrax vs.RelationBuildPartitionDesc()" is in discussion for
> > the master
> > > > branch only, not sure though, but we need the similar fix for the back
> > branches as well.
> > >
> > > Well, this is not a bug as such, so it's very unlikely that a fix like
> > > this will be back-patched.  Also, if this becomes an issue only for
> > > more than over 1000 partitions, then it's not very relevant for PG 10
> > > and PG 11, because we don't recommend using so many partitions with
> > > them.  Maybe we can consider fixing PG 12 though.
> >
> > A fix for the thousands-of-partitions case would be very welcome for 12.
> >
> >
> Look like commit # d3f48dfae42 added the required fix but is enabled only
> for
> the clobber-cache builds :(

I've got a real world multi-tenancy case that would really be helped
by this. Can we enable it for all builds, please?

Best,
David.
-- 
David Fetter <david(at)fetter(dot)org> http://fetter.org/
Phone: +1 415 235 3778

Remember to vote!
Consider donating to Postgres: http://www.postgresql.org/about/donate



Re: Some memory not freed at the exit of RelationBuildPartitionDesc()

From
Tom Lane
Date:
David Fetter <david@fetter.org> writes:
> On Tue, Aug 13, 2019 at 03:08:34PM +0530, amul sul wrote:
>> Look like commit # d3f48dfae42 added the required fix but is enabled only
>> for the clobber-cache builds :(

> I've got a real world multi-tenancy case that would really be helped
> by this. Can we enable it for all builds, please?

This sounds like nonsense to me.  As pointed out by the commit message,
that fix was just taking care of bloat observed with CCA on.

            regards, tom lane