While reviewing Etsuro-san's patch to force replanning after FDW option
changes, I noticed that there is a great lack of consistency about where
InvokeObjectPostAlterHook calls have been placed relative to other actions
such as forced relcache invals. I wonder exactly what expectations a
post-alter hook function could have about cache coherency, or indeed if
there's any clarity at all about what such a hook might do. For instance,
it looks to me like the hook would generally need to do a
CommandCounterIncrement in order to be able to "see" the update it's being
called for, and I'm unsure that that would be safe at every call site.
Is that supposed to be allowed, or are we expecting that object access
hooks are only going to do open-loop actions that don't rely on the
details of the change?
I suppose this fits in well with our grand tradition of not documenting
hooks at all, but for a set of hooks as invasive as these are, I think
we ought to do better.
regards, tom lane
