Start With Data, Not Hunches
The query that feels slow is rarely the one that costs the most. A report that runs once a day for two seconds matters far less than a lookup that runs ten thousand times a minute at fifty milliseconds. Tuning works best when you rank queries by total time, which is mean latency multiplied by call count.
Where the Numbers Live
- Statistics views such as pg stat statements aggregate calls, total time, and mean time per normalized query.
- Application metrics and traces show latency from the caller side, including network and pool wait.
- The slow query log captures individual statements that cross a time threshold.
What to Look For
- High total time means a strong tuning target even if each call is quick.
- High mean time points at a single expensive statement.
- High call count can hint at an n plus one or a missing cache.
A Simple Loop
Measure, pick the top offender by total time, fix it, then measure again. The ranking shifts after every fix, so always re sort.
Key idea
Find slow queries by ranking on total time, mean latency times call count, so you spend effort where it changes overall load the most.