performance of IN (subquery)

[Kevin Murphy]

I'm using PG 7.4.3 on Mac OS X.

I am disappointed with the performance of queries like 'select foo from
bar where baz in (subquery)', or updates like 'update bar set foo = 2
where baz in (subquery)'. PG always seems to want to do a sequential
scan of the bar table. I wish there were a way of telling PG, "use the
index on baz in your plan, because I know that the subquery will return
very few results". Where it really matters, I have been constructing
dynamic queries by looping over the values for baz and building a
separate query for each one and combining with a UNION (or just
directly updating, in the update case). Depending on the size of the
bar table, I can get speedups of hundreds or even more than a thousand
times, but it is a big pain to have to do this.

Any tips?

Thanks,
Kevin Murphy

Illustrated:

The query I want to do is very slow:

select bundle_id from build.elements
where elementid in (
SELECT superlocs_2.ele ment_id
FROM superlocs_2 NATURAL JOIN bundle_superloc s_2
WHERE bundle_superloc s_2.protobundle _id = 1);
-----------
7644
7644
(2 rows)
Time: 518.242 ms
The subquery is fast:

SELECT superlocs_2.ele ment_id
FROM superlocs_2 NATURAL JOIN bundle_superloc s_2
WHERE bundle_superloc s_2.protobundle _id = 1;
------------
41209
25047
(2 rows)
Time: 3.268 ms
And using indexes on the main table is fast:

select bundle_id from build.elements
where elementid in (41209, 25047);
-----------
7644
7644
(2 rows)
Time: 2.468 ms

The plan for the slow query:

egenome_test=# explain analyze select bundle_id from build.elements
where elementid in (
SELECT superlocs_2.ele ment_id
FROM superlocs_2 NATURAL JOIN bundle_superloc s_2
WHERE bundle_superloc s_2.protobundle _id = 1);
egenome_test-# egenome_test(# egenome_test(# egenome_test(#
QUERY PLAN
\

------------------------------------------------------------------------
-------------------------------------------------------------
Hash Join (cost=70.33..72 .86 rows=25 width=4) (actual
time=583.051..5 83.059 rows=2 loops=1)
Hash Cond: ("outer".elemen t_id = "inner".element id)
-> HashAggregate (cost=47.83..47 .83 rows=25 width=4) (actual
time=0.656..0.6 58 rows=2 loops=1)
-> Hash Join (cost=22.51..47 .76 rows=25 width=4) (actual
time=0.615..0.6 25 rows=2 loops=1)
Hash Cond: ("outer".superl oc_id = "inner".superlo c_id)
-> Seq Scan on superlocs_2 (cost=0.00..20. 00 rows=1000
width=8) (actual time=0.004..0.0 12 rows=9 loops=1)
-> Hash (cost=22.50..22 .50 rows=5 width=4) (actual
time=0.076..0.0 76 rows=0 loops=1)
-> Seq Scan on bundle_superloc s_2
(cost=0.00..22. 50 rows=5 width=4) (actual time=0.024..0.0 33 rows=2
loops=1)
Filter: (protobundle_id = 1)
-> Hash (cost=20.00..20 .00 rows=1000 width=8) (actual
time=581.802..5 81.802 rows=0 loops=1)
-> Seq Scan on elements (cost=0.00..20. 00 rows=1000 width=8)
(actual time=0.172..405 .243 rows=185535 loops=1)
Total runtime: 593.843 ms
(12 rows)
---------------------------(end of broadcast)---------------------------
TIP 6: Have you searched our list archives?

http://archives.postgresql.org

[Marc G. Fournier]

On Thu, 26 Aug 2004, Kevin Murphy wrote:

I'm using PG 7.4.3 on Mac OS X.

I am disappointed with the performance of queries like 'select foo from bar
where baz in (subquery)', or updates like 'update bar set foo = 2 where baz
in (subquery)'. PG always seems to want to do a sequential scan of the bar
table. I wish there were a way of telling PG, "use the index on baz in your
plan, because I know that the subquery will return very few results". Where
it really matters, I have been constructing dynamic queries by looping over
the values for baz and building a separate query for each one and combining
with a UNION (or just directly updating, in the update case). Depending on
the size of the bar table, I can get speedups of hundreds or even more than a
thousand times, but it is a big pain to have to do this.

Any tips?

Thanks,
Kevin Murphy

Illustrated:

The query I want to do is very slow:

select bundle_id from build.elements
where elementid in (
SELECT superlocs_2.ele ment_id
FROM superlocs_2 NATURAL JOIN bundle_superloc s_2
WHERE bundle_superloc s_2.protobundle _id = 1);
-----------
7644
7644
(2 rows)
Time: 518.242 ms

what field type is protobundle_id? if you typecast the '1' to be the
same, does the index get used?

Email: sc*****@hub.org Yahoo!: yscrappy ICQ: 7615664

---------------------------(end of broadcast)---------------------------
TIP 3: if posting/reading through Usenet, please send an appropriate
subscribe-nomail command to ma*******@postg resql.org so that your
message can get through to the mailing list cleanly

[Paul Tillotson]

Kevin Murphy wrote:

------------------------------------------------------------------------
-------------------------------------------------------------
Hash Join (cost=70.33..72 .86 rows=25 width=4) (actual
time=583.051..5 83.059 rows=2 loops=1)
Hash Cond: ("outer".elemen t_id = "inner".element id)
-> HashAggregate (cost=47.83..47 .83 rows=25 width=4) (actual
time=0.656..0.6 58 rows=2 loops=1)
-> Hash Join (cost=22.51..47 .76 rows=25 width=4) (actual
time=0.615..0.6 25 rows=2 loops=1)
Hash Cond: ("outer".superl oc_id = "inner".superlo c_id)
-> Seq Scan on superlocs_2 (cost=0.00..20. 00 rows=1000 width=8)
(actual time=0.004..0.0 12 rows=9 loops=1)
-> Hash (cost=22.50..22 .50 rows=5 width=4) (actual time=0.076..0.0 76
rows=0 loops=1)
-> Seq Scan on bundle_superloc s_2 (cost=0.00..22. 50 rows=5 width=4)
(actual time=0.024..0.0 33 rows=2 loops=1)
Filter: (protobundle_id = 1)
-> Hash (cost=20.00..20 .00 rows=1000 width=8) (actual
time=581.802..5 81.802 rows=0 loops=1)
-> Seq Scan on elements (cost=0.00..20. 00 rows=1000 width=8) (actual
time=0.172..405 .243 rows=185535 loops=1)
The planner thinks that the sequential scan on elements will return 1000
rows, but it actually returned 185000. Did you ANALYZE this table recently?

Afterthought: It would be nice if the database was smart enough to
analyze a table of its own accord when a sequential scan returns more
than, say, 20 times what it was supposed to.

Paul
Total runtime: 593.843 ms
(12 rows)
---------------------------(end of broadcast)---------------------------
TIP 6: Have you searched our list archives?

http://archives.postgresql.org

---------------------------(end of broadcast)---------------------------
TIP 7: don't forget to increase your free space map settings

[Arthur Ward]

> Afterthought: It would be nice if the database was smart enough to

analyze a table of its own accord when a sequential scan returns more
than, say, 20 times what it was supposed to.

I've wondered on several occasions if there is any good reason for PG not
to automatically perform an analyze concurrently with a seq scan as it's
happening. That way, no extra disk IO is needed and the stats could say
up-to-date for almost free.

Any hackers around who can say why this might be a bad idea, or is it one
of those things that just needs a volunteer? (I'm not; at least not now.)

---------------------------(end of broadcast)---------------------------
TIP 8: explain analyze is your friend

[Greg Stark]

"Arthur Ward" <aw************ **@dominionscie nces.com> writes:

Any hackers around who can say why this might be a bad idea, or is it one
of those things that just needs a volunteer? (I'm not; at least not now.)

a) that would make plans change spontaneously. I hate being paged in the
middle of the night because some query is suddenly being slow when it had been
performing fine before.

b) Not all sequential scans will actually complete the scan. There could be a
limit imposed or a the sequential scan could be inside a EXISTS. In that case
the scan could be aborted at any point.

What I do think would be easy to do would be to keep statistics on the expense
of various components of the cost estimates. cpu_*_cost, random_page_cos t
effective_cache _size, ought to be values that can be solved for empirically
from the timing results.

But that still doesn't have to be done on every query. There's a trade-off
between work done on every query to plan queries and the benefit. Gathering
statistics and storing them on every sequential scan is way too much work
slowing down every query for minimal gain.

--
greg
---------------------------(end of broadcast)---------------------------
TIP 7: don't forget to increase your free space map settings

[Tom Lane]

Paul Tillotson <pn***@shentel. net> writes:

The planner thinks that the sequential scan on elements will return 1000
rows, but it actually returned 185000. Did you ANALYZE this table recently?
Or either of the other ones? All those scan costs look like defaults
:-(
Afterthought: It would be nice if the database was smart enough to
analyze a table of its own accord when a sequential scan returns more
than, say, 20 times what it was supposed to.

I've thought about this before. One simple trick would be to get rid
of the current pg_class reltuples/relpages fields in favor of a
tuples-per-page estimate, which could be multiplied by
RelationGetNumb erOfBlocks() during planning. In the absence of any
ANALYZE data the tuples-per-page estimate might be pretty bogus, but
it couldn't be off by more than an order of magnitude or so either way.
And in any case we'd have a guaranteed up-to-date number of blocks.

The objections that could be raised to this are (AFAICS) two:

1. Adding at least an lseek() kernel call per table, and per index, to
every planning operation. I'm not sure this would be significant,
but I'm not sure it wouldn't be, either.

2. Instability of plans. Right now, the planner will not change plans
underneath you --- you have to issue an explicit VACUUM or ANALYZE
to change the terms of discussion. That would stop being true if
physical file size were always taken into account. Maybe this is a
problem, or maybe it isn't ... as someone who likes to be able to
debug planner behavior without actually creating umpteen-gig test
tables, my world view may be a bit skewed ...

It's certainly doable if we decide the pluses outweigh the minuses.
Thoughts?

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 1: subscribe and unsubscribe commands go to ma*******@postg resql.org

[Greg Stark]

Tom Lane <tg*@sss.pgh.pa .us> writes:

I've thought about this before. One simple trick would be to get rid of the
current pg_class reltuples/relpages fields in favor of a tuples-per-page
estimate, which could be multiplied by RelationGetNumb erOfBlocks() during
planning.
This would do something interesting to one of the problem cases I have now. I
have trouble testing a particular batch job that generates a large amount of
precalculated denormalized data.

That's because the data is empty when the job starts, so the plpgsql function
that handles the job caches plans based on an empty table. But as the job
proceeds the data grows and I'm afraid the cached plan may start performing
poorly.

In order to test it I need to run it once, analyze, then reload the function,
truncate the data and run it another time. And hope I generated good
representative test data.

I'm thinking of changing to a non-plpgsql implementation. But that's only half
the issue. I'm not about to run analyze in the middle of the data generation
(which wouldn't work anyways since it's in a transaction). So I can't really
get good statistics for this job, not until we're actually in a steady state
in production.

Sometimes I wonder whether I wouldn't rather a more predictable system with
less focus on statistics. The resulting plans would be more predictable and
predictability is a good thing in production systems...
In the absence of any ANALYZE data the tuples-per-page estimate might be
pretty bogus, but it couldn't be off by more than an order of magnitude or
so either way.
I don't see why it couldn't. If you have a table badly in need of vacuuming
(or had one at the time of the last analyze) it could be off by way more than
an order of magnitude.

For that matter, a table that had undergone many deletes and then been
vacuumed would not change length for a long time afterward even as many new
inserts are performed. Until the table is analyzed the estimated table size
could be off by an arbitrary factor.
The objections that could be raised to this are (AFAICS) two:

1. Adding at least an lseek() kernel call per table, and per index, to
every planning operation. I'm not sure this would be significant,
but I'm not sure it wouldn't be, either.
That seems like something that could be addressed with enough time. A single
value per table, couldn't it be stored in shared memory and only updated
whenever the heap is extended or truncated? Even if you have thousands of
tables that would only be a few kilobytes of shared memory.
2. Instability of plans. Right now, the planner will not change plans
underneath you --- you have to issue an explicit VACUUM or ANALYZE
to change the terms of discussion. That would stop being true if
physical file size were always taken into account. Maybe this is a
problem, or maybe it isn't ... as someone who likes to be able to
debug planner behavior without actually creating umpteen-gig test
tables, my world view may be a bit skewed ...

This is what I'm afraid of. As a OLTP application programmer -- web sites, I
admit -- I care a lot more about plan stability than finding optimal plans.

An well written OLTP application will only have a fixed number of queries that
are executed repeatedly with different parameters. I don't care how long the
queries take as long as they're always "fast enough".

Every time a new plan is tried there's a chance that it will be wrong. It only
takes one wrong plan out of the hundreds of queries to bring down the entire
application.

Ideally I would want a guarantee that every query would *always* result in the
same plan. Once I've tested them and approved the plans I want to know that
only those approved plans will ever run, and I want to be present and be able
to verify new plans before they go into production.

I doubt I'm going to convince anyone today, but I think there will be a
gradual change in mindset as the new binary protocol becomes more popular. And
postgres takes over some of mysql's web mindshare.

--
greg
---------------------------(end of broadcast)---------------------------
TIP 4: Don't 'kill -9' the postmaster

[Gaetano Mendola]

Tom Lane wrote:

2. Instability of plans. Right now, the planner will not change plans
underneath you --- you have to issue an explicit VACUUM or ANALYZE
to change the terms of discussion. That would stop being true if
physical file size were always taken into account. Maybe this is a
problem, or maybe it isn't ... as someone who likes to be able to
debug planner behavior without actually creating umpteen-gig test
tables, my world view may be a bit skewed ...

Did you forgot the autovacuum daemon ? I didn't see anyone bitten or
paged during the night for autovacuum daemon job.
Regards
Gaetano Mendola

[Jon Lapham]

Tom Lane wrote:

I've thought about this before. One simple trick would be to get rid
of the current pg_class reltuples/relpages fields in favor of a
tuples-per-page estimate, which could be multiplied by
RelationGetNumb erOfBlocks() during planning. In the absence of any
ANALYZE data the tuples-per-page estimate might be pretty bogus, but
it couldn't be off by more than an order of magnitude or so either way.
And in any case we'd have a guaranteed up-to-date number of blocks.

The objections that could be raised to this are (AFAICS) two:
[snip]
2. Instability of plans. Right now, the planner will not change plans
underneath you --- you have to issue an explicit VACUUM or ANALYZE
to change the terms of discussion. That would stop being true if
physical file size were always taken into account. Maybe this is a
problem, or maybe it isn't ... as someone who likes to be able to
debug planner behavior without actually creating umpteen-gig test
tables, my world view may be a bit skewed ...

It's certainly doable if we decide the pluses outweigh the minuses.
Thoughts?

My first reaction is to wonder if this would give performance exactly
equal to running a true ANALYZE in every situation? If not, then you
would end up with an automated pseudo-ANALYZE (performance-wise).

In my opinion, it is almost a feature that non-ANALYZE-d tables give
such horrendous performance, it kicks you in the butt to do some
thinking about when to correctly deal with ANALYZEing.

So, in short, I think it is a huge win if we could have automatic
ANALYZE with true ANALYZE performance, but a huge loss if the automatic
ANALYZE performance is not exactly as good as a true ANALYZE.

--
-**-*-*---*-*---*-*---*-----*-*-----*---*-*---*-----*-----*-*-----*---
Jon Lapham <la****@jandr.o rg> Rio de Janeiro, Brasil
Personal: http://www.jandr.org/
***-*--*----*-------*------------*--------------------*---------------
---------------------------(end of broadcast)---------------------------
TIP 9: the planner will ignore your desire to choose an index scan if your
joining column's datatypes do not match

[Kevin Murphy]

Thanks all for the reminders about analyzing, and I apologize for wasting
everyone's time. The main table did indeed need to be analyzed (I had
truncated it and repopulated it with "insert ... select" but forgotten to
analyze). The other tables are very small temporary tables, and I assumed,
whether correctly or not, that analyzing would not be helpful for them.

All this is happening in the context of an algorithm in a PL/PGSQL function.
The three temporary tables are reused thousands of times. I wasn't sure if
it would be better to truncate them between uses to keep them small or just
allow them to grow. Unfortunately for the Tree of Knowledge, performance is
now more than adequate, so I may not do this experiment.

Thanks,
Kevin Murphy

---------------------------(end of broadcast)---------------------------
TIP 2: you can get off all lists at once with the unregister command
(send "unregister YourEmailAddres sHere" to ma*******@postg resql.org)