Thread: [HACKERS] Dropping a partitioned table takes too long
$SUBJECT, if the table has, say, 2000 partitions. The main reason seems to be that RelationBuildPartitionDesc() will be called that many times within the same transaction, which perhaps we cannot do much about right away. But one thing we could do is to reduce the impact of memory allocations it does. They are currently leaked into the caller's context, which may not be reset immediately (such as PortalHeapMemory). Instead of doing it in the caller's context, use a temporary context that is deleted before returning. Attached is a patch for that. On my local development VM, `drop table table_with_2000_partitions` finished in 27 seconds with the patch instead of more than 20 minutes that it currently takes. Thoughts? Adding this to the open items list. Thanks, Amit PS: this was actually mentioned by Ragnar Ouchterlony who reported some bugs back in declarative partitioning in January [1] [1] https://www.postgresql.org/message-id/17d89e08-874b-c1b1-aa46-12d5afb26235%40agama.tv -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers
Attachment
The attached patch try to replace 'heap_open' with 'LockRelationOid' when locking parent table.
It improved dropping a table with 7000 partitions.2017-04-25 15:07 GMT+08:00 Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>:
$SUBJECT, if the table has, say, 2000 partitions.
The main reason seems to be that RelationBuildPartitionDesc() will be
called that many times within the same transaction, which perhaps we
cannot do much about right away. But one thing we could do is to reduce
the impact of memory allocations it does. They are currently leaked into
the caller's context, which may not be reset immediately (such as
PortalHeapMemory). Instead of doing it in the caller's context, use a
temporary context that is deleted before returning. Attached is a patch
for that. On my local development VM, `drop table
table_with_2000_partitions` finished in 27 seconds with the patch instead
of more than 20 minutes that it currently takes.
Thoughts?
Adding this to the open items list.
Thanks,
Amit
PS: this was actually mentioned by Ragnar Ouchterlony who reported some
bugs back in declarative partitioning in January [1]
[1]
https://www.postgresql.org/message-id/17d89e08-874b-c1b1- aa46-12d5afb26235%40agama.tv
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
--
Attachment
On Tue, Apr 25, 2017 at 12:37 PM, Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> wrote: > $SUBJECT, if the table has, say, 2000 partitions. > > The main reason seems to be that RelationBuildPartitionDesc() will be > called that many times within the same transaction, which perhaps we > cannot do much about right away. But one thing we could do is to reduce > the impact of memory allocations it does. They are currently leaked into > the caller's context, which may not be reset immediately (such as > PortalHeapMemory). Instead of doing it in the caller's context, use a > temporary context that is deleted before returning. Attached is a patch > for that. On my local development VM, `drop table > table_with_2000_partitions` finished in 27 seconds with the patch instead > of more than 20 minutes that it currently takes. > > Thoughts? > I am not able to undestand why does changing memory context cause so much difference in execution time? The way this patch uses the memory context in this patch, it's possible that in future we will allocate something in temporary context and then refer it from a longer context. Instead, may be we should free things specially or change memory context only when allocating those things. -- Best Wishes, Ashutosh Bapat EnterpriseDB Corporation The Postgres Database Company
Hi, On 2017/04/25 20:07, 高增琦 wrote: > > 2017-04-25 15:07 GMT+08:00 Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>: > >> $SUBJECT, if the table has, say, 2000 partitions. >> >> The main reason seems to be that RelationBuildPartitionDesc() will be >> called that many times within the same transaction, which perhaps we >> cannot do much about right away. But one thing we could do is to reduce >> the impact of memory allocations it does. They are currently leaked into >> the caller's context, which may not be reset immediately (such as >> PortalHeapMemory). Instead of doing it in the caller's context, use a >> temporary context that is deleted before returning. Attached is a patch >> for that. On my local development VM, `drop table >> table_with_2000_partitions` finished in 27 seconds with the patch instead >> of more than 20 minutes that it currently takes. > > The attached patch try to replace 'heap_open' with 'LockRelationOid' when > locking parent table. > It improved dropping a table with 7000 partitions. Your patch seems to be a much better solution to the problem, thanks. Regards, Amit
On 2017/04/25 20:55, Ashutosh Bapat wrote: > On Tue, Apr 25, 2017 at 12:37 PM, Amit Langote > <Langote_Amit_f8@lab.ntt.co.jp> wrote: >> $SUBJECT, if the table has, say, 2000 partitions. >> >> The main reason seems to be that RelationBuildPartitionDesc() will be >> called that many times within the same transaction, which perhaps we >> cannot do much about right away. But one thing we could do is to reduce >> the impact of memory allocations it does. They are currently leaked into >> the caller's context, which may not be reset immediately (such as >> PortalHeapMemory). Instead of doing it in the caller's context, use a >> temporary context that is deleted before returning. Attached is a patch >> for that. On my local development VM, `drop table >> table_with_2000_partitions` finished in 27 seconds with the patch instead >> of more than 20 minutes that it currently takes. >> >> Thoughts? >> > I am not able to undestand why does changing memory context cause so > much difference in execution time? > > The way this patch uses the memory context in this patch, it's > possible that in future we will allocate something in temporary > context and then refer it from a longer context. Instead, may be we > should free things specially or change memory context only when > allocating those things. Actually, I am withdrawing this patch for time being, because a much direct and better solution has been offered upthread by GaoZengqi. Anyway, temporary context added by my patch would not contain any objects that will be accessed outside of RelationBuildPartitionDesc(). Remember that the content that we put into the longer-lived rd_partdesc uses the memory allocated in rd_pdcxt, not the temporary context that I proposed. Thanks, Amit
On Tue, Apr 25, 2017 at 10:05 PM, Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> wrote: >> The attached patch try to replace 'heap_open' with 'LockRelationOid' when >> locking parent table. >> It improved dropping a table with 7000 partitions. > > Your patch seems to be a much better solution to the problem, thanks. Does anyone wish to object to this patch as untimely? If not, I'll commit it. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On 4/26/17 12:15, Robert Haas wrote: > On Tue, Apr 25, 2017 at 10:05 PM, Amit Langote > <Langote_Amit_f8@lab.ntt.co.jp> wrote: >>> The attached patch try to replace 'heap_open' with 'LockRelationOid' when >>> locking parent table. >>> It improved dropping a table with 7000 partitions. >> >> Your patch seems to be a much better solution to the problem, thanks. > > Does anyone wish to object to this patch as untimely? > > If not, I'll commit it. Seems quite reasonable. -- Peter Eisentraut http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
Robert Haas <robertmhaas@gmail.com> writes:
> On Tue, Apr 25, 2017 at 10:05 PM, Amit Langote
> <Langote_Amit_f8@lab.ntt.co.jp> wrote:
>> Your patch seems to be a much better solution to the problem, thanks.
> Does anyone wish to object to this patch as untimely?
> If not, I'll commit it.
It's certainly not untimely to address such problems. What I'm wondering
is if we should commit both patches. Avoiding an unnecessary heap_open
is certainly a good thing, but it seems like the memory leak addressed
by the first patch might still be of concern in other scenarios.
regards, tom lane
On Wed, Apr 26, 2017 at 12:22 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote: > Robert Haas <robertmhaas@gmail.com> writes: >> On Tue, Apr 25, 2017 at 10:05 PM, Amit Langote >> <Langote_Amit_f8@lab.ntt.co.jp> wrote: >>> Your patch seems to be a much better solution to the problem, thanks. > >> Does anyone wish to object to this patch as untimely? > >> If not, I'll commit it. > > It's certainly not untimely to address such problems. What I'm wondering > is if we should commit both patches. Avoiding an unnecessary heap_open > is certainly a good thing, but it seems like the memory leak addressed > by the first patch might still be of concern in other scenarios. I will defer to you on that. If you think that patch is a good idea, please have at it. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
It seems that in 'load_relcache_init_file()', we forget to initialize 'rd_pdcxt' of relcache.
2017-04-27 0:33 GMT+08:00 Robert Haas <robertmhaas@gmail.com>:
On Wed, Apr 26, 2017 at 12:22 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> On Tue, Apr 25, 2017 at 10:05 PM, Amit Langote
>> <Langote_Amit_f8@lab.ntt.co.jp> wrote:
>>> Your patch seems to be a much better solution to the problem, thanks.
>
>> Does anyone wish to object to this patch as untimely?
>
>> If not, I'll commit it.
>
> It's certainly not untimely to address such problems. What I'm wondering
> is if we should commit both patches. Avoiding an unnecessary heap_open
> is certainly a good thing, but it seems like the memory leak addressed
> by the first patch might still be of concern in other scenarios.
I will defer to you on that. If you think that patch is a good idea,
please have at it.
--
On Wed, Apr 26, 2017 at 12:21 PM, Peter Eisentraut <peter.eisentraut@2ndquadrant.com> wrote: > On 4/26/17 12:15, Robert Haas wrote: >> On Tue, Apr 25, 2017 at 10:05 PM, Amit Langote >> <Langote_Amit_f8@lab.ntt.co.jp> wrote: >>>> The attached patch try to replace 'heap_open' with 'LockRelationOid' when >>>> locking parent table. >>>> It improved dropping a table with 7000 partitions. >>> >>> Your patch seems to be a much better solution to the problem, thanks. >> >> Does anyone wish to object to this patch as untimely? >> >> If not, I'll commit it. > > Seems quite reasonable. OK, done. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company
On Fri, Apr 28, 2017 at 6:12 AM, 高增琦 <pgf00a@gmail.com> wrote: > It seems that in 'load_relcache_init_file()', we forget to initialize > 'rd_pdcxt' of relcache. Fixed. Thanks. -- Robert Haas EnterpriseDB: http://www.enterprisedb.com The Enterprise PostgreSQL Company