Bad query plan with high-cardinality column

[Alexander Staubo]

I have a problem with a query that is planned wrong. This is my schema:

create table comments (
id serial primary key,
conversation_id integer,
created_at timestamp
);
create index comments_conversation_id_index on comments (conversation_id);
create index comments_created_at_index on comments (created_at);

The table has 3.5M rows, and 650k unique values for "conversation_id", where the histogram goes up to 54000 rows for the most frequent ID, with a long tail. There are only 20 values with a frequency of 1000 or higher. The "created_at" column has 3.5M distinct values.

Now, I have this query:

select comments.id from comments where
conversation_id = 3975979 order by created_at limit 13

This filters about 5000 rows and returns the oldest 13 rows. But the query is consistently planned wrong:

Limit (cost=0.00..830.53 rows=13 width=12) (actual time=3174.862..3179.525 rows=13 loops=1)
Buffers: shared hit=2400709 read=338923 written=21
-> Index Scan using comments_created_at_index on comments (cost=0.00..359938.52 rows=5634 width=12) (actual time=3174.860..3179.518 rows=13 loops=1)
Filter: (conversation_id = 3975979)
Rows Removed by Filter: 2817751
Buffers: shared hit=2400709 read=338923 written=21
Total runtime: 3179.553 ms

It takes anywhere between 3 seconds and several minutes to run, depending on how warm the disk cache is. This is the correct plan and index usage:

Limit (cost=6214.34..6214.38 rows=13 width=12) (actual time=25.471..25.473 rows=13 loops=1)
Buffers: shared hit=197 read=4510
-> Sort (cost=6214.34..6228.02 rows=5471 width=12) (actual time=25.469..25.470 rows=13 loops=1)
Sort Key: created_at
Sort Method: top-N heapsort Memory: 25kB
Buffers: shared hit=197 read=4510
-> Index Scan using comments_conversation_id_index on comments (cost=0.00..6085.76 rows=5471 width=12) (actual time=1.163..23.955 rows=5834 loops=1)
Index Cond: (conversation_id = 3975979)
Buffers: shared hit=197 read=4510
Total runtime: 25.500 ms

The problem for Postgres is obviously to estimate how many rows have a given "conversation_id" value, but I have confirmed that the value is correctly tracked in the histogram.

I'm at a loss how to explain why the planner thinks scanning a huge index that covers the entire table will ever beat scanning a small index that has 17% of the table's values. Even if the entire database were in RAM, this would hit way too much buffers unnecessarily. (I have determined that planner will consistently use the bad plan for higher-frequency values, and the good plan for lower-frequency values.) It will *always* be better to branch off the "comments_conversation_id_index" index.

Another curious thing: If I run ANALYZE repeatedly, the planner will sometimes, oddly enough, choose the correct plan. This behaviour actually seems related to effective_cache_size; if it's set small (128MB), the planner will sometimes favour the good plan, but if large (>= 2GB) it will consistently use the bad plan. Not sure if ANALYZE is changing anything or if it's just bad timing.

Things I have tried: I have bumped the statistics target up to 10000, but it does not help. I have also tried setting n_distinct for the column manually, since Postgres guesses it's 285k instead of 650k, but that does not help.

Our environment: 9.2.2, tweaked memory parameters (work_mem, sort_mem, shared_buffers, effective_cache_size), not touched planner/cost parameters. Problem also exists on 9.2.3.

[Tom Lane]

Post by Alexander Staubo
select comments.id from comments where
conversation_id = 3975979 order by created_at limit 13
I'm at a loss how to explain why the planner thinks scanning a huge
index that covers the entire table will ever beat scanning a small index
that has 17% of the table's values.

The reason is that the LIMIT may stop the query before it's scanned all
of the index. The planner estimates on the assumption that the desired
rows are roughly uniformly distributed within the created_at index, and
on that assumption, it looks like this query will stop fairly soon ...
but evidently, that's wrong. On the other hand, it knows quite well
that the other plan will require pulling out 5000-some rows and then
sorting them before it can return anything, so that's not going to be
exactly instantaneous either.

In this example, I'll bet that conversation_id and created_at are pretty
strongly correlated, and that most or all of the rows with that specific
conversation_id are quite far down the created_at ordering, so that the
search through the index takes a long time to run. OTOH, with another
conversation_id the same plan might run almost instantaneously.

If you're concerned mostly with this type of query then a 2-column index
on (conversation_id, created_at) would serve your purposes nicely. You
could likely even dispense with the separate index on conversation_id
alone.

regards, tom lane

[Alexander Staubo]

Post by Tom Lane
The reason is that the LIMIT may stop the query before it's scanned all
of the index. The planner estimates on the assumption that the desired
rows are roughly uniformly distributed within the created_at index, and
on that assumption, it looks like this query will stop fairly soon ...
but evidently, that's wrong. On the other hand, it knows quite well
that the other plan will require pulling out 5000-some rows and then
sorting them before it can return anything, so that's not going to be
exactly instantaneous either.
In this example, I'll bet that conversation_id and created_at are pretty
strongly correlated, and that most or all of the rows with that specific
conversation_id are quite far down the created_at ordering, so that the
search through the index takes a long time to run. OTOH, with another
conversation_id the same plan might run almost instantaneously.

That's right. So I created a composite index, and not only does this make the plan correct, but the planner now chooses a much more efficient plan than the previous index that indexed only on "conversation_id":

Limit (cost=0.00..30.80 rows=13 width=12) (actual time=0.042..0.058 rows=13 loops=1)
Buffers: shared hit=8
-> Index Scan using index_comments_on_conversation_id_and_created_at on comments (cost=0.00..14127.83 rows=5964 width=12) (actual time=0.039..0.054 rows=13 loops=1)
Index Cond: (conversation_id = 3975979)
Buffers: shared hit=8
Total runtime: 0.094 ms


Is this because it can get the value of "created_at" from the index, or is it because it can know that the index is pre-sorted, or both?

Very impressed that Postgres can use a multi-column index for this. I just assumed, wrongly, that it couldn't. I will have to go review my other tables now and see if they can benefit from multi-column indexes.

Thanks!

[Tom Lane]

Post by Alexander Staubo
...
Is this because it can get the value of "created_at" from the index, or is it because it can know that the index is pre-sorted, or both?

What it knows is that leading index columns that have equality
constraints are effectively "don't-cares" for ordering purposes.
So in general, an indexscan on an index on (x,y) will be seen to
provide the ordering required by any of these queries:

select ... order by x;
select ... order by x,y;
select ... where x = constant order by x,y;
select ... where x = constant order by y;

Your query is an example of the last pattern. So the planner sees that
the bare indexscan, with no additional sort step, can satisfy the query,
and then its cost estimate for that with the effects of the LIMIT will
be less than for the other possible plans. There's no need to scan and
then sort thousands of rows, and there's no need to read through a
hard-to-guess-but-certainly-large number of irrelevant index entries.
The relevant part of the index is a small number of adjacent entries
that are already in the right order.

regards, tom lane

[Kevin Grittner]

   create table comments (
     id serial primary key,
     conversation_id integer,
     created_at timestamp
   );
   create index comments_conversation_id_index on comments (conversation_id);
   create index comments_created_at_index on comments (created_at);

I suspect you would be better off without those two indexes, and
instead having an index on (conversation_id, created_at).  Not just
for the query you show, but in general.

   select comments.id from comments where
     conversation_id = 3975979 order by created_at limit 13
This filters about 5000 rows and returns the oldest 13 rows. But
[planner thinks it will be cheaper to read index in ORDER BY
sequence and filter rows until it has 13 than to read 5471 rows
and sort them to pick the top 13 after the sort.]

In my experience these problems come largely from the planner not
knowing the cost of dealing with each tuple.  I see a lot less of
this if I raise cpu_tuple_cost to something in the 0.03 to 0.05
range.
--
Kevin Grittner
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

[Alexander Staubo]

Post by Kevin Grittner
I suspect you would be better off without those two indexes, and
instead having an index on (conversation_id, created_at). Not just
for the query you show, but in general.

Indeed, that solved it, thanks!

Post by Kevin Grittner
In my experience these problems come largely from the planner not
knowing the cost of dealing with each tuple. I see a lot less of
this if I raise cpu_tuple_cost to something in the 0.03 to 0.05
range.

Is this something I can just frob a bit without worrying about it adversely impacting database performance across the board, or should I be very careful and do lots of testing on a staging box first?

[Kevin Grittner]

Post by Alexander Staubo

Post by Kevin Grittner
In my experience these problems come largely from the planner
not knowing the cost of dealing with each tuple. I see a lot
less of this if I raise cpu_tuple_cost to something in the 0.03
to 0.05 range.

Is this something I can just frob a bit without worrying about it
adversely impacting database performance across the board, or
should I be very careful and do lots of testing on a staging box
first?

If possible, I would recommend trying it with the old indexes and
seeing whether it causes it to choose the better plan.  (Of course,
you're not going to beat the plan you get with the two-column index
for this query, but it might help it better cost the other
alternatives, which would be a clue that it makes your overall
costing model more accurate and would have a more general benefit.)
You can play with settings like this in a single session without
affecting any other sessions.

I always recommend testing a change like this in staging and
closely monitoring after deploying to production, to confirm the
overall benefit and look for any odd cases which might suffer a
performance regression.  For this particular change, I have never
seen a negative effect, but I'm sure that it's possible to have a
scenario where it isn't helpful.

Personally, I have changed this setting many times and have often
noted that 0.02 was not enough to cause choice of an optimal plan,
0.03 was always enough to do it if adjusting this setting was going
to help at all, and boosting it to 0.05 never caused further plan
changes in the cases I tested.  I have never seen such increases
cause less optimal plan choice.

If you try this, please post your results.

--
Kevin Grittner
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company