Skip to main content

How to tune work_mem

work_mem is used for sorting and hashing operations during query execution.

By default it's 4MB.

Due to its nature, the tuning of work_mem is quite tricky.

One of the possible approaches is explained here.

Rough tuning and "safe" values of work_mem​

First, apply rough optimization as described in Rough configuration tuning (80/20 rule; OLTP).

A query can "spend" work_mem multiple times (for multiple operations). But it is not allocated fully for each operation – an operation can need a lower amount of memory.

Therefore, it is hard to reliably predict, how much memory we'll need to use to handle a workload, without actual observation of the workload.

Moreover, in Postgres 13, new parameter was added, hash_mem_multiplier, adjusting the logic. The default is 2 in PG13-16. It means that max memory used by a single hash operation is 2 * work_mem.

Worth mentioning, understanding how much memory is used by a session in Linux is very tricky per se – see a great article by Andres Freund: Analyzing the Limits of Connection Scalability in Postgres (2020).

A safe approach would be:

  • estimate how much memory is free – subtracting shared_buffers, maintenance_work_mem, etc.,
  • then divide the estimated available memory by max_connections and additional number such as 4-5 (or more, to be on even safer side) – assuming that each backend will be using up to 4work_mem or 5work_mem. Of course, this multiplier itself is a very rough estimate – in reality, OLTP workloads usually are much less hungry on average (e.g., having a lot of PK lookups mean that average memory consumption is very low).

In practice, it can make sense to adjust work_mem to a higher value, but this needs to be done after understanding the behavior of Postgres under certain workload. The following steps are parts of iterative approach for further tuning.

Temp files monitoring​

Monitor how often temporary files are created and the size of these files. Source of proper info:

1) pg_stat_database, columns temp_files and temp_bytes.

2) Postgres logs – set log_temp_files to a low value, down to 0 (being careful with the observer effect).

3) temp_blks_read and temp_blks_written in pg_stat_statements.

The goal of further work_mem tuning is getting rid of temporary files completely, or minimizing the frequency of creation of them and the size.

Optimize queries​

If possible, consider optimizing queries that involve temporary files creation. Spend a reasonable amount of effort on this - if there is no obvious optimization, proceed to the next step.

Raising work_mem further: how much?​

If reasonable optimization efforts are done, it is time to consider raising work_mem further.

However, how much? The answer lies in the size of individual temporary files – from monitoring described above, we can find maximum and average temp file size, and estimate the required raise from there.

🎯 TODO: detailed steps

Raise work_mem for parts of workload​

It makes sense to raise it for individual queries. Consider two options:

  • set work_mem ... in individual sessions or transactions (set local ...), or
  • if you use various DB users for various parts of your workload, consider adjusting work_mem for particular users (e.g., for those that run analytical-like queries): alter user ... set work_mem ....

Raise work_mem globally​

Only if the previous steps are not suitable (e.g., it is hard to optimize queries and you cannot tune work_mem for parts of workload), then consider raising work_mem globally, evaluating OOM risks.

Iterate​

After some time, review data from monitoring to ensure that situation improved or decide to perform another iteration.

Extra: pg_get_backend_memory_contexts​

In Postgres 14, a new function pg_get_backend_memory_contexts() and corresponding view were added; see docs. This is helpful for detailed analysis of how current session works with memory, but the major limitation of this capability is that it works only with the current session.

🎯 TODO: How to use it.