On Fri, Feb 24, 2012 at 3:14 AM, Daniel Farina <daniel@heroku.com> wrote:
> At ExecutorFinish (already hookable) all NodeKeys remembered by an
> extension should be invalidated, as that memory is free and ready to
> be used again.
I think this statement is false; I thought it was true because *not*
invalidating gives me correct-appearing but unsound results.
Firstly, it should be ExecutorEnd, not ExecutorFinish, but this
doesn't work because those run in the exec_simple_query path twice:
once in the planner and then again in PortalCleanup, it's important
that invalidation only occur at the end, and not twice. So there's a
problem of a kind resulting from this experiment.
--
fdr