• Timing guidelines for queries
• Cold and warm cache
• Slow list views and APIs

Query performance guidelines

This document describes various guidelines to follow when optimizing SQL queries.

When you are optimizing your SQL queries, there are two dimensions to pay attention to:

1. The query execution time. This is paramount as it reflects how the user experiences GitLab.
2. The query plan. Optimizing the query plan is important in allowing queries to independently scale over time. Realizing that an index keeps a query performing well as the table grows before the query degrades is an example of why we analyze these plans.

Timing guidelines for queries

• When analyzing your query’s performance, pay attention to if the time you are seeing is on a cold or warm cache. These guidelines apply for both cache types.
• When working with batched queries, change the range and batch size to see how it effects the query timing and caching.
• If an existing query is not performing well, make an effort to improve it. If it is too complex or would stall development, create a follow-up so it can be addressed in a timely manner. You can always ask the database reviewer or maintainer for help and guidance.

Cold and warm cache

When evaluating query performance it is important to understand the difference between cold and warm cached queries.

The first time a query is made, it is made on a “cold cache”. Meaning it needs to read from disk. If you run the query again, the data can be read from the cache, or what PostgreSQL calls shared buffers. This is the “warm cache” query.

When analyzing an EXPLAIN plan, you can see the difference not only in the timing, but by looking at the output for Buffers by running your explain with EXPLAIN(analyze, buffers). Database Lab automatically includes these options.

If you are making a warm cache query, you see only the shared hits.

For example, using Database Lab:

Shared buffers:
  - hits: 36467 (~284.90 MiB) from the buffer pool
  - reads: 0 from the OS file cache, including disk I/O

Or in the explain plan from psql:

Buffers: shared hit=7323

If the cache is cold, you also see reads.

Using Database Lab:

Shared buffers:
  - hits: 17204 (~134.40 MiB) from the buffer pool
  - reads: 15229 (~119.00 MiB) from the OS file cache, including disk I/O

In psql:

Buffers: shared hit=7202 read=121

Slow list views and APIs

We often build filtered list views and APIs in GitLab which need to have many different filter and sorting options. All these options are usually encapsulated in finders and exposed by API/GraphQL arguments. While we have many possible pagination performance optimizations , there is often no way to make all combinations of sorting and filtering performant. Attempts to make many options performant might involve adding too many indexes which sacrifices performance of our primary database. This is only justified for common use cases and should not be considered as a way to make all permutations of filter and sort performant. What this means practically is that there will likely be filtered views and API requests that timeout when certain sorting or filtering options are applied. We still allow them to be added by teams where they benefit certain customers with specific combinations of filtering/sorting, but we need to accept that they will timeout for some users.