The usecase that I think might be useful is to have a database client send metadata along with a query.
This partially is possible today by setting application_name, but that is a separate request, it would be great if that could be sent along with the query in one go.
Another option to pass metadata is to add a comment (/* .. */), but a comment cannot be set for a prepared statement, because the statement is prepared first and then later invoked on runtime, which executes a query that is fixed.
observability frameworks like OpenTelemetry support tracing through all layers of a stack, and trace_ids can even be passed into sql with extensions like sqlcommenter. however sqlcommenter puts the trace_id into a comment which effectively breaks all the utility of pg_stat_statements since each query has a unique trace_id.
There are some other use-cases: 1) RO-RW routing. Users can pass target-session-attrs to the server within query labels to hint about its need. Useful when some kind of proxy (odyssey,pgbouncer,spqr,pgpool II, pgcat, pg_doorman) is used. 2) pg_comment_stats uses comments in the query to accumulate statistics. [0]
Thinking a bit more, my root issue is specifically around pg_stat_statements so maybe it’s also solvable with some changes to how query jumbling is done
But that topic seems like one where we’d never get consensus
Should query jumbling for calculating query_id be customizable somehow? How would we resolve multiple concurrent opinions about how queries should be jumbled (eg if comment_stats needs different tweaks than sqlcommenter)? Was there previous discussion about this already? I’ll need to go search mailing list history a bit
-Jeremy
Sent from my TI-83
Jeremy,
Thanks for this. I am hoping to get consensus of ideas for changes to the protocol.
