postgres-ai/nancy

Prewarm in Nancy?

Nastradamus opened this issue · 1 comments

Do we need pg_prewarm support in Nancy? Or other techniques?

Questions to answer:

  • how does VACUUM ANALYZE (we do it before any run) affect OS file cache and shared buffers?
  • if we consider prewarming and the size of DB exceeds RAM significantly, how will we choose what to cache?
  • isn't it better to perform an experimental run multiple times (say, 4) to get statistically reliable results and just disregard the results of the first time? The question is how to do it faster (especially if the workload is not read-only), but if filesystem snapshots will help, maybe this approach will be the best one?