Robert Haas <robertmhaas@gmail.com> writes: > ... Second, we've generally made a > decision up until now that we don't want to have a hard dependency on > stable functions actually being stable. If they aren't, and for > example you're using index expressions, your queries may return wrong > answers, but you won't get weird internal error messages, and you > won't get a crash. I think the bar for this feature is the same.
Yeah, I agree --- wrong answers may be acceptable in such a case, but crashes or unintelligible error messages aren't. There are practical reasons for being tolerant here, notably that it's not that easy for users to get their volatility markings right.
In the case at hand, I think that means that allowing workers to do pruning would require hardening the parallel append code against the situation where their pruning results vary. Maybe, instead of passing the pruning results *down*, we could pass them *up* to the leader and then throw an error if they're different?
Note we’re talking here about “initial” pruning that occurs during ExecInitNode(). Workers are only launched during ExecGather[Merge]() which thereafter do ExecInitNode() on their copy of the the plan tree. So if we are to pass the pruning results for cross-checking, it will have to be from the leader to workers.
