Index Advisor: How to Test Missing Index Insights

We recommend testing recommendations in pre-production or staging environments first before deploying changes to production. If possible, it is advisable to use a copy of the production database for your tests, otherwise you may not see a representative performance improvement or query plan change.

To perform a test and validate whether a recommendation improves performance:

  1. Pick one or more of the affected queries mentioned by the Index Advisor
  2. Run “EXPLAIN (ANALYZE, BUFFERS) SELECT …” for each of the queries, twice
  3. Note down the runtime of the second invocation (to avoid cold caches affecting the test)
  4. Additionally, note down the sum of the top level numbers of “shared blocks read” and “shared blocks hit” (this represents how much data the query was loading)
  5. Create the recommended index by copying the “CREATE INDEX” command
  6. Re-run the EXPLAIN (ANALYZE, BUFFERS) test, and compare the numbers

Note that the Index Advisor cost improvement is an estimatealways test and verify whether recommendations actually improve performance. In a warm cache test, as described above, it is likely that the measured improvement for the query timing is less than expected, as data in cache can be accessed quickly, even for larger amounts of data. The modeled execution cost improvement (the cost=... in the first line of plan output) should be closer, but may still vary due to the large number of variables in play.

The “shared buffers read” and “shared buffers hit” counts (combined together) give you a better measurement of what the true order of magnitude improvement of a change is, and reflects the cold cache performance better.

You can also use the query statistics and EXPLAIN plans that pganalyze collects to verify production performance after deploying a change. Index Advisor runs every 24h, and will resolve a missing index by watching for changes in the schema. Resolved issues can be viewed by navigating to the Resolved tab on the Missing Index check page.


Couldn't find what you were looking for or want to talk about something specific?
Start a conversation with us →