Hello,
I'm not sure if these ideas were circulated before or not.
We use auto_explain a lot to investigate slow/problematic queries.
One of the main issues with its usability comes from the fact that EXPLAIN output is logged rather than returned to the caller in some way. If you have a large cluster with lots of replicas, there is also an extra inconvenience of log accumulation, search, etc.
Why not have an option to return EXPLAIN results as a NoticeResponse instead? That would make its usage more convenient.
Another thing is tangentially related...
I think it may be good to have a number of options to generate significantly shorter output similar to EXPLAIN. EXPLAIN is great, but sometimes people need more concise and specific information, for example total number of buffers and reads by certain query (this is pretty common), whether or not we had certain nodes in the plan (seq scan, scan of certain index(es)), how bad was cardinality misprediction on certain nodes, etc. It's not totally clear yet what would be the best way to define those rules, but I think we can come up with something reasonable. Logging or returning shorter messages like that can cause less overhead than logging full EXPLAIN and can potentially allow for better query monitoring overall.
Do you see any potential issues with implementing those? Of course there should be more details, like what kind of configuration parameters to add, how to define rules for the 2nd case, etc. Just wanted to check if there are any objections in general.
Thank you,
-Vladimir Churyukin.