Re: [PoC] Improve dead tuple storage for lazy vacuum

On Wed, 7 Jul 2021 at 13:47, Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
>
> Hi all,
>
> Index vacuuming is one of the most time-consuming processes in lazy
> vacuuming. lazy_tid_reaped() is a large part among them. The attached
> the flame graph shows a profile of a vacuum on a table that has one index
> and 80 million live rows and 20 million dead rows, where
> lazy_tid_reaped() accounts for about 47% of the total vacuum execution
> time.
>
> [...]
>
> Overall, 'rtbm' has a much better lookup performance and good memory
> usage especially if there are relatively many dead tuples. However, in
> some cases, 'intset' and 'array' have a better memory usage.

Those are some great results, with a good path to meaningful improvements.

> Feedback is very welcome. Thank you for reading the email through to the end.

The current available infrastructure for TIDs is quite ill-defined for
TableAM authors [0], and other TableAMs might want to use more than
just the 11 bits in use by max-BLCKSZ HeapAM MaxHeapTuplesPerPage to
identify tuples. (MaxHeapTuplesPerPage is 1169 at the maximum 32k
BLCKSZ, which requires 11 bits to fit).

Could you also check what the (performance, memory) impact would be if
these proposed structures were to support the maximum
MaxHeapTuplesPerPage of 1169 or the full uint16-range of offset
numbers that could be supported by our current TID struct?

Kind regards,

Matthias van de Meent

[0] https://www.postgresql.org/message-id/flat/0bbeb784050503036344e1f08513f13b2083244b.camel%40j-davis.com