Adding an index can decrease SELECT performance (original) (raw)

© Laurenz Albe 2018

• An example
• A new index makes things go sour
  • What happened?
• What can we do to get the better plan?

We all know that you have to pay a price for a new index you create — data modifying operations will become slower, and indexes use disk space. That's why you try to have no more indexes than you actually need.

But most people think that SELECT performance will never suffer from a new index. The worst that can happen is that the new index is not used.

However, this is not always true, as I have seen more than once in the field. I'll show you such a case and tell you what you can do about it.

An example

We will experiment with this table:

| | CREATE TABLE skewed ( sort integer NOT NULL, category integer NOT NULL, interesting boolean NOT NULL);INSERT INTO skewed SELECT i, i%1000, i>50000 FROM generate_series(1, 1000000) i;CREATE INDEX skewed_category_idx ON skewed (category);VACUUM (ANALYZE) skewed; | | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |

We want to find the first twenty interesting rows in category 42:

| | EXPLAIN (ANALYZE, BUFFERS)SELECT * FROM skewedWHERE interesting AND category = 42ORDER BY sortLIMIT 20; | | ------------------------------------------------------------------------------------------------------------ |

This performs fine:

PostgreSQL uses the index to find the 1000 rows with category 42, filters out the ones that are not interesting, sorts them and returns the top 20. 5 milliseconds is fine.

A new index makes things go sour

Now we add an index that can help us with sorting. That is definitely interesting if we often have to find the top 20 results:

| | CREATE INDEX skewed_sort_idx ON skewed (sort); | | --------------------------------------------------- |

And suddenly, things are looking worse:

| | QUERY PLAN------------------------------------------------------------- Limit (cost=0.42..736.34 rows=20 width=9) (actual time=21.658..28.568 rows=20 loops=1) Buffers: shared hit=374 read=191 -> Index Scan using skewed_sort_idx on skewed (cost=0.42..33889.43 rows=921 width=9) (actual time=21.655..28.555 rows=20 loops=1) Filter: (interesting AND (category = 42)) Rows Removed by Filter: 69022 Buffers: shared hit=374 read=191 Planning time: 0.507 ms Execution time: 28.632 ms | | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |

What happened?

PostgreSQL thinks that it will be faster if it examines the rows in sort order using the index until it has found 20 matches. But it doesn't know how the matching rows are distributed with respect to the sort order, so it is not aware that it will have to scan 69042 rows until it has found its 20 matches (see Rows Removed by Filter: 69022 in the above execution plan).

What can we do to get the better plan?

PostgreSQL v10 has added extended statistics to track how the values in different columns are correlated, but that does not track the distributions of the values, so it will not help us here.

There are two workarounds:

1. Drop the index that misleads PostgreSQL. If that is possible, it is a simple solution. But usually one cannot do that, because the index is either used to enforce a unique constraint, or it is needed by other queries that benefit from it.
2. Rewrite the query so that PostgreSQL cannot use the offending index. Of the many possible solutions for this, I want to present two:
   • A subquery with OFFSET 0:
     | | SELECT *FROM (SELECT * FROM skewed WHERE interesting AND category = 42 OFFSET 0) qORDER BY sortLIMIT 20; |
     | -------------------------------------------------------------------------------------------------------------- |
     This makes use of the fact that OFFSET and LIMIT prevent a subquery from being “flattened”, even if they have no effect on the query result.
   • Using an expression as sort key:
     | | SELECT * FROM skewedWHERE interesting AND category = 42ORDER BY sort + 0LIMIT 20; |
     | -------------------------------------------------------------------------------------- |
     This makes use of the fact that PostgreSQL cannot deduce that sort + 0 is the same as sort. Remember that PostgreSQL is extensible, and you can define your own + operator!

In order to receive regular updates on important changes in PostgreSQL, subscribe to our newsletter, or follow us on Twitter, Facebook, or LinkedIn.