Horton

Re: BitmapHeapScan streaming read user and prelim refactoring

Tomas Vondra <tomas@vondra.me>

View thread View on postgr.es
From: Tomas Vondra <tomas@vondra.me>
To: Andres Freund <andres@anarazel.de>, Jakub Wartak <jakub.wartak@enterprisedb.com>
Cc: Robert Haas <robertmhaas@gmail.com>, Melanie Plageman <melanieplageman@gmail.com>, Nazir Bilal Yavuz <byavuz81@gmail.com>, Dilip Kumar <dilipbalaut@gmail.com>, Heikki Linnakangas <hlinnaka@iki.fi>, Thomas Munro <thomas.munro@gmail.com>, Pg Hackers <pgsql-hackers@postgresql.org>
Date: 2025-02-14T17:36:37Z
Lists: pgsql-hackers 
On 2/14/25 18:14, Andres Freund wrote:
> Hi,
> 
> On 2025-02-14 10:04:41 +0100, Jakub Wartak wrote:
>> Is there any reason we couldn't have new pg_test_iorates (similiar to
>> other pg_test_* proggies), that would literally do this and calibrate
>> best e_io_c during initdb and put the result into postgresql.auto.conf
>> (pg_test_iorates --adjust-auto-conf) , that way we would avoid user
>> questions on how to come with optimal value?
> 
> Unfortunately I think this is a lot easier said than done:
> 
> - The optimal depth depends a lot on your IO patterns, there's a lot of things
>   between fully sequential and fully random.
> 
>   You'd really need to test different IO sizes and different patterns. The
>   test matrix for that gets pretty big.
> 
> - The performance characteristics of storage heavily changes over time.
> 
>   This is particularly true in cloud environments, where disk/VM combinations
>   will be "burstable", allowing higher throughput for a while, but then not
>   anymore.  Measureing during either of those states will not be great for the
>   other state.
> 
> - e_io_c is per-query-node, but impacts the whole system. If you set
>   e_io_c=1000 on a disk with a metered IOPS of say 1k/s you might get a
>   slightly higher throughput for a bitmap heap scan, but also your commit
>   latency in concurrent will go through the roof, because your fdatasync()
>   will be behind a queue of 1k reads that all are throttled.
> 

All of this is true, ofc, but maybe it's better to have a tool providing
at least some advice? I'd imagine pg_test_fsync is affected by many of
those issues too (considering both are about I/O).

I'd definitely not want initdb to do this automatically, though. Getting
good numbers is fairly expensive (in time and I/O), can be flaky, etc.

But maybe having a tool that gives you a bunch of numbers, as input for
manual tuning, would be good enough?

As you say, it's not just about the hardware (and how that changes over
time because of "burst" credits etc.), but also about the workload.
Would it be possible to track something, and adjust this dynamically
over time? And then adjust the prefetch distance in some adaptive way?

Perhaps it's a bit naive, but say we know what % of requests is handled
from cache, how long it took, etc. We opportunistically increase the
prefetch distance, and check the cache hit ratio after a while. Did it
help (% increased, consumed less time), maybe try another step. If not,
maybe try prefetching less?


regards

-- 
Tomas Vondra

Commits

  1. Fix bitmapheapscan incorrect recheck of NULL tuples

  2. Increase default maintenance_io_concurrency to 16

  3. Separate TBM[Shared|Private]Iterator and TBMIterateResult

  4. Improve read_stream.c advice for dense streams.

  5. Increase default effective_io_concurrency to 16

  6. Delay extraction of TIDBitmap per page offsets

  7. Add lossy indicator to TBMIterateResult

  8. Move BitmapTableScan per-scan setup into a helper

  9. Add and use BitmapHeapScanDescData struct

  10. Fix bitmap table scan crash on iterator release

  11. Bitmap Table Scans use unified TBMIterator

  12. Add common interface for TBMIterators

  13. Make table_scan_bitmap_next_block() async-friendly

  14. Move EXPLAIN counter increment to heapam_scan_bitmap_next_block

  15. Refactor tidstore.c iterator buffering.

  16. BitmapHeapScan: Remove incorrect assert and reset field

  17. Change BitmapAdjustPrefetchIterator to accept BlockNumber

  18. BitmapHeapScan: Use correct recheck flag for skip_fetch

  19. BitmapHeapScan: Push skip_fetch optimization into table AM

  20. BitmapHeapScan: postpone setting can_skip_fetch

  21. BitmapHeapScan: begin scan after bitmap creation

  22. Fix EXPLAIN Bitmap heap scan to count pages with no visible tuples

  23. Remove redundant snapshot copying from parallel leader to workers

  24. Remove some obsolete smgrcloseall() calls.

  25. Remove the "snapshot too old" feature.

  26. Compute XID horizon for page level index vacuum on primary.