On Thu, 11 Mar 2021 at 16:16, John Naylor <john.nay...@enterprisedb.com> wrote:
>
> On Thu, Mar 11, 2021 at 9:46 AM Matthias van de Meent
> <boekewurm+postg...@gmail.com> wrote:
>
> > Regarding the 2% slack logic, could we change it to use increments of
> > line pointers instead? That makes it more clear what problem this
> > solution is trying to work around; that is to say line pointers not
> > (all) being truncated away. The currently subtracted value accounts
>
> That makes sense.
>
> > ...
> > - if (len + saveFreeSpace > MaxHeapTupleSize)
> > + if (len + saveFreeSpace > maxPaddedFsmRequest)
> > ...
> > - targetFreeSpace = Max(len, MaxHeapTupleSize - (MaxHeapTupleSize * 2 /
> > 100));
> > + targetFreeSpace = Max(len, maxPaddedFsmRequest);
> > ...
>
> If we have that convenient constant, it seems equivalent (I think) and a bit
> more clear to write it this way, but I'm not wedded to it:
>
> if (len + saveFreeSpace > MaxHeapTupleSize &&
> len <= maxPaddedFsmRequest)
> {
> ...
> targetFreeSpace = maxPaddedFsmRequest;
> }
+ else if (len > maxPaddedFsmRequest
+ {
+ /* request len amount of space; it might still fit on
not-quite-empty pages */
+ targetFreeSpace = len;
+ }
If this case isn't added, the lower else branch will fail to find
fitting pages for len > maxPaddedFsmRequest tuples; potentially
extending the relation when there is actually still enough space
available.
> else
> targetFreeSpace = len + saveFreeSpace;
> Also, should I write a regression test for it? The test case is already
> available, just no obvious place to put it.
I think it would be difficult to write tests that exhibit the correct
behaviour on BLCKSZ != 8196. On the other hand, I see there are some
tests that explicitly call out that they expect BLCKSZ to be 8192, so
that has not really been a hard block before.
The previous code I sent had initial INSERT + DELETE + VACUUM. These
statements can be replaced with `INSERT INTO t_failure (b) VALUES
(repeat('1', 95)); VACUUM;` for simplicity. The vacuum is still needed
to populate the FSM for the new page.
With regards,
Matthias van de Meent
