Re: Unhappy about API changes in the no-fsm-for-small-rels patch

On Thu, May 2, 2019 at 4:57 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> On Thu, May 2, 2019 at 12:39 PM John Naylor <john(dot)naylor(at)2ndquadrant(dot)com> wrote:
> >
> Okay, I have updated the patch to incorporate your changes and call
> relcache invalidation at required places. I have updated comments at a
> few places as well. The summarization of this patch is that (a) it
> moves the local map to relation cache (b) performs the cache
> invalidation whenever we create fsm (either via backend or vacuum),
> when some space in a page is freed by vacuum (provided fsm doesn't
> exist) or whenever the local map is cleared (c) additionally, we clear
> the local map when we found a block from FSM, when we have already
> tried all the blocks present in cache or when we are allowed to create
> FSM.
>
> If we agree on this, then we can update the README accordingly.
>
> Can you please test/review?

There isn't enough time. But since I already wrote some debugging
calls earlier (attached), I gave it a brief spin, I found this patch
isn't as careful as HEAD making sure we don't try the same block twice
in a row. If you insert enough tuples into an empty table such that we
need to extend, you get something like this:

DEBUG: Not enough space on block 0
DEBUG: Now trying block 0
DEBUG: Not enough space on block 0
DEBUG: Updating local map for block 0

At this point, I'm sorry to say, but I'm in favor of reverting. There
just wasn't enough time to redesign and debug a feature like this
during feature freeze.

--
John Naylor https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services