Best way to get result count before LIMIT was applied

Question

When paging through data that comes from a DB, you need to know how many pages there will be to render the page jump controls.

Currently I do that by running the query twice, once wrapped in a count() to determine the total results, and a second time with a limit applied to get back just the results I need for the current page.

This seems inefficient. Is there a better way to determine how many results would have been returned before LIMIT was applied?

I am using PHP and Postgres.

Answer 1

Pure SQL

Things have changed since 2008. You can use a window function to get the full count and the limited result in one query. (Introduced with PostgreSQL 8.4 in 2009).

SELECT foo
     , count(*) OVER() AS full_count
FROM   bar
WHERE  <some condition>
ORDER  BY <some col>
LIMIT  <pagesize>
OFFSET <offset>

Note that this can be considerably more expensive than without the total count. All rows have to be counted, and a possible shortcut taking just the top rows from a matching index may not be helpful any more.
Doesn't matter much with small tables or full_count <= OFFSET + LIMIT. Matters for a substantially bigger full_count.

Corner case: when OFFSET is at least as great as the number of rows from the base query, no row is returned. So you also get no full_count. Possible alternative:

• Run a query with a LIMIT/OFFSET and also get the total number of rows

Consider the sequence of events:

1. WHERE clause (and JOIN conditions, but not here) filter qualifying rows from the base table(s).

   (GROUP BY and aggregate functions would go here.)

2. Window functions are applied considering all qualifying rows (depending on the OVER clause and the frame specification of the function). The simple count(*) OVER() is based on all rows.

3. ORDER BY

   (DISTINCT or DISTINCT ON would go here.)

4. LIMIT / OFFSET are applied based on the established order to select rows to return.

LIMIT / OFFSET becomes increasingly inefficient with a growing number of rows in the table. Consider alternative approaches if you need better performance:

• Optimize query with OFFSET on large table

Alternatives to get final count

There are completely different approaches to get the count of affected rows (not the full count before OFFSET & LIMIT were applied). Postgres has internal bookkeeping how many rows where affected by the last SQL command. Some clients can access that information or count rows themselves (like psql).

For instance, you can retrieve the number of affected rows in plpgsql immediately after executing an SQL command with:

GET DIAGNOSTICS integer_var = ROW_COUNT;

Details in the manual.

Or you can use pg_num_rows in PHP. Or similar functions in other clients.

Related:

• Calculate number of rows affected by batch query in PostgreSQL

Answer 2

As I describe on my blog, MySQL has a feature called SQL_CALC_FOUND_ROWS. This removes the need to do the query twice, but it still needs to do the query in its entireity, even if the limit clause would have allowed it to stop early.

As far as I know, there is no similar feature for PostgreSQL. One thing to watch out for when doing pagination (the most common thing for which LIMIT is used IMHO): doing an "OFFSET 1000 LIMIT 10" means that the DB has to fetch at least 1010 rows, even if it only gives you 10. A more performant way to do is to remember the value of the row you are ordering by for the previous row (the 1000th in this case) and rewrite the query like this: "... WHERE order_row > value_of_1000_th LIMIT 10". The advantage is that "order_row" is most probably indexed (if not, you've go a problem). The disadvantage being that if new elements are added between page views, this can get a little out of synch (but then again, it may not be observable by visitors and can be a big performance gain).

Answer 3

You could mitigate the performance penalty by not running the COUNT() query every time. Cache the number of pages for, say 5 minutes before the query is run again. Unless you're seeing a huge number of INSERTs, that should work just fine.

Answer 4

Since Postgres already does a certain amount of caching things, this type of method isn't as inefficient as it seems. It's definitely not doubling execution time. We have timers built into our DB layer, so I have seen the evidence.

Answer 5

Seeing as you need to know for the purpose of paging, I'd suggest running the full query once, writing the data to disk as a server-side cache, then feeding that through your paging mechanism.

If you're running the COUNT query for the purpose of deciding whether to provide the data to the user or not (i.e. if there are > X records, give back an error), you need to stick with the COUNT approach.