I wrote:
> After sleeping on it, I think the right answer is to introduce the new
> error-message field (and not worry about 9.5). Will work on a patch
> for that, unless I hear objections pretty soon.
BTW, while I'm looking at this: what on god's green earth is
ThrowErrorData doing copying the supplied data into edata->assoc_context?
Surely it should be putting the data into the ErrorContext, where it's not
going to get flushed before it can be reported? (Note that in the sole
existing use-case, edata->assoc_context is going to have been set to
CurrentMemoryContext by pq_parse_errornotice, and I see no good reason to
assume that's very long-lived ... in fact, it looks like it's whatever
happens to be active when ProcessInterrupts is called, which means there's
probably a totally separate set of problems here having to do with
potential leaks into long-lived contexts.)
I have little use for the name of that function either, as it's not
necessarily going to "throw" anything. Maybe ReportErrorUsingData,
or something like that?
regards, tom lane
