Re: Refactoring the checkpointer's fsync request queue

On Sat, Mar 2, 2019 at 9:35 AM Shawn Debnath <sdn(at)amazon(dot)com> wrote:
> On Fri, Mar 01, 2019 at 03:03:19PM -0500, Robert Haas wrote:
> > On Fri, Mar 1, 2019 at 2:36 PM Shawn Debnath <sdn(at)amazon(dot)com> wrote:
> > > I disagree, at least with combining and retaining enums. Encoding all
> > > the possible request types with the current, planned and future SMGRs
> > > would cause a sheer explosion in the number of enum values.
> >
> > How big of an explosion would it be?
>
> 4 enum values x # of smgrs; currently md, soon undo and slru so 12 in
> total. Any future smgr addition will expand this further.

It's not so much the "explosion" that bothers me. I think we should
have a distinct sync requester enum, because we need a way to index
into the table of callbacks. How exactly you pack the two enums into
compact space seems like a separate question; doing it with two words
would obviously be wasteful, but it should be possible stuff them into
(say) a single uint8_t, uint16_t or whatever will pack nicely in the
request struct and allow the full range of request types (4?) + the
full range of sync requesters (which we propose to expand to 3 in the
forseeable future). Now perhaps the single enum idea was going to
involve explicit values that encode the two values SYNC_REQ_CANCEL_MD
= 0x1 | (0x04 << 4) so you could still extract the requester part, but
that's just the same thing with uglier code.

--
Thomas Munro
https://enterprisedb.com