Re: Freeze the inserted tuples during CTAS?

On Wed, Mar 10, 2021 at 3:57 PM Paul Guo <guopa(at)vmware(dot)com> wrote:
>
> > On Mar 3, 2021, at 1:35 PM, Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> >> On Sun, Feb 21, 2021 at 4:46 PM Paul Guo <guopa(at)vmware(dot)com> wrote:
> >> Attached is the v2 version that fixes a test failure due to plan change (bitmap index scan -> index only scan).
>
> > I think this is a good idea.
>
> > BTW, how much does this patch affect the CTAS performance? I expect
> > it's negligible but If there is much performance degradation due to
> > populating visibility map, it might be better to provide a way to
> > disable it.
>
> Yes, this is a good suggestion. I did a quick test yesterday.
>
> Configuration: shared_buffers = 1280M and the test system memory is 7G.
>
> Test queries:
> checkpoint;
> \timing
> create table t1 (a, b, c, d) as select i,i,i,i from generate_series(1,20000000) i;
> \timing
> select pg_size_pretty(pg_relation_size('t1'));
>
> Here are the running time:
>
> HEAD : Time: 10299.268 ms (00:10.299) + 1537.876 ms (00:01.538)
> Patch : Time: 12257.044 ms (00:12.257) + 14.247 ms
>
> The table size is 800+MB so the table should be all in the buffer. I was surprised
> to see the patch increases the CTAS time by 19.x%, and also it is not better than
> "CTAS+VACUUM" on HEAD version. In theory the visibility map buffer change should
> not affect that much. I looked at related code again (heap_insert()). I believe
> the overhead could decrease along with some discussed CTAS optimization
> solutions (multi-insert, or raw-insert, etc).
>
> I tested 'copy' also. The COPY FREEZE does not involve much overhead than COPY
> according to the experiement results as below. COPY uses multi-insert. Seems there is
> no other difference than CTAS when writing a new table.
>
> COPY TO + VACUUM
> Time: 8826.995 ms (00:08.827) + 1599.260 ms (00:01.599)
> COPY TO FREEZE + VACUUM
> Time: 8836.107 ms (00:08.836) + 13.581 ms
>
> So maybe think about doing freeze in CTAS after optimizing the CTAS performance
> later?

Thank you for testing. That's interesting.

I've also done some benchmarks for CTAS (2GB table creation) and got
similar results:

Patched : 44 sec
HEAD : 34 sec

Since CREATE MATERIALIZED VIEW is also internally treated as CTAS, I
got similar results even for CREATE MATVIEW.

After investigation, it seems to me that the cause of performance
degradation is that heap_insert() set PD_ALL_VISIBLE when inserting a
tuple for the first time on the page (L2133 in heapam.c). This
requires every subsequent heap_insert() to pin a visibility map buffer
(see RelationGetBufferForTuple()). This problem doesn't exist in
heap_multi_insert() since it reads vm buffer once to fill the heap
page with tuples.

Given such relatively big performance degradation, it seems to me that
we should do some optimization for heap_insert() first. Otherwise, we
will end up imposing those costs on all users.

> By the way, ‘REFRESH MatView’ does freeze by default. Matview is quite similar to CTAS.
> I did test it also and the conclusion is similar to that of CTAS. Not sure why FREEZE was
> enabled though, maybe I missed something?

I’m not sure the reason why setting visibility map and PD_ALL_VISIBLE
during REFRESH MATVIEW is enabled by default. By commit 7db0cd2145 and
39b66a91b, heap_insert and heap_multi_insert() sets visibility map
bits if HEAP_INSERT_FROZEN is specified. Looking at the commit
messages, those changes seem to be intended to COPY FREEZE but they
indeed affect REFRESH MATVIEW as well. But I could not find discussion
and mention about REFRESH MATVIEW both in those threads and commit
messages.

One reason that might justify such behavior would be materialized
views are read-only. Since visibility map bits never be cleared, even
if it costs here is some cost to set visibility map during the refresh
setting VM bits and PD_ALL_VISIBLE at creation might win. On the other
hand, a table created by CTAS is read-write. The user might not want
to pay a cost during creating a table if the table is updated
frequently after creation. Not sure. Being said that, I think this
performance degradation of REFRESH MATVIEW could be a problem. There
is no way to avoid the degradation and we also can rely on autovacuum
to set visibility map bits on materialized views. I'll start a new
thread to discuss that.

Regards,

--
Masahiko Sawada
EDB: https://www.enterprisedb.com/