Horton

Re: index prefetching

Peter Geoghegan <pg@bowt.ie>

View thread View on postgr.es
From: Peter Geoghegan <pg@bowt.ie>
To: Andres Freund <andres@anarazel.de>
Cc: Tomas Vondra <tomas@vondra.me>, Alexandre Felipe <o.alexandre.felipe@gmail.com>, Thomas Munro <thomas.munro@gmail.com>, Nazir Bilal Yavuz <byavuz81@gmail.com>, Robert Haas <robertmhaas@gmail.com>, Melanie Plageman <melanieplageman@gmail.com>, PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>, Georgios <gkokolatos@protonmail.com>, Konstantin Knizhnik <knizhnik@garret.ru>, Dilip Kumar <dilipbalaut@gmail.com>
Date: 2026-02-17T17:16:23Z
Lists: pgsql-hackers

Attachments

On Mon, Feb 16, 2026 at 11:48 AM Andres Freund <andres@anarazel.de> wrote:
> Yes, It's hurting quite substantially. For well correlated index scans it
> prevents readahead from becoming aggressive enough even on a local low latency
> SSD. Which means it'll not even be even remotely aggressive enough on a
> networked block store.

I agree that the current heuristics (which were invented recently) are
too conservative. I overfit the heuristics to my current set of
adversarial queries, as a stopgap measure.

> Note that there is pretty much *no* readhead, because the yields happen more
> frequently than a io_combine_limit sized IO can be formed.

ISTM that we need the yields to better cooperate with whatever's
happening on the read stream side.

> With the yielding logic disabled:

> The comment seems to say it's about avoiding to look very into the future when
> using index only scans that just need a few heap lookups. Certainly an
> important goal.

The main motivation for yielding is to deal with things like merge
joins fed by at least one plain index scan, and plain scans for an
"ORDER BY .... LIMIT N" query. The LIMIT N is also addressed by the
"Use ExecSetTupleBound hint during index scans" patch, albeit
imperfectly. This is also important with index-only scans, though only
when heap fetches are actually required in the first place (most
index-only scans will never create a read stream in the first place).

I attach an example of where disabling the yield mechanism hurts
instead of helping, to give you a sense of the problems in this area.
Notice that the inner side of the merge join ("Index Scan using
prefetch_customers_pkey on prefetch_customers c") requires the same
number of buffer hits + buffer reads as master with yielding enabled
-- that allows the patch to shave off 7% of master's execution time.
Whereas without yielding, there's quite a lot more buffer hits (for
index page reads) + buffer misses (for heap page reads) -- which hurts
us. Without yielding, the patch takes about 13% longer to execute the
query.

Just to be clear, I'm not arguing that this is the right trade-off.
And I understand that it's just not feasible to completely prevent
such plan shapes from reading more data than strictly necessary -- a
certain amount of that seems like a "cost of doing business". But it
seems important that the added costs of speculatively reading later
batches/index leaf pages never exceeds some fixed threshold.

The non-yielding EXPLAIN ANALYZE has an inner-index scan that does
~60% more work than master/than the yielding variant of the patch.
It's not so much that the amount of extra work we'll perform is
excessive (though it is); what really concerns me is that the amount
of extra work is *indeterminate*. There's likely to be other queries
that are much slower still, at least in the absence of some kind of
yielding mechanism. And so ISTM that we need a better yield mechanism
-- one that is better attuned to the need to maintain an adequate
prefetch distance on the read stream side.

> One thing that does confuse me about the yielding logic is that it seems to
> actually put a cap on ever looking more than two batches ahead (with a bit of
> fuzziness)? Why support more batches then (INDEX_SCAN_MAX_BATCHES == 128)?
> Isn't two batches too low for anything with some correlation on even remotely
> higher latency storage?

FWIW that's not actually true; we *can* still use more than 2 batches.
My instrumentation confirms this. Though it seems to top out at about
4 batches total with the current yield logic/with my test suite
queries.

-- 
Peter Geoghegan

Commits

  1. read stream: Split decision about look ahead for AIO and combining

  2. read_stream: Only increase read-ahead distance when waiting for IO

  3. aio: io_uring: Trigger async processing for large IOs

  4. heapam: Keep buffer pins across index scan resets.

  5. heapam: Track heap block in IndexFetchHeapData.

  6. Move heapam_handler.c index scan code to new file.

  7. Rename heapam_index_fetch_tuple argument for clarity.

  8. Optimize fast-path FK checks with batched index probes

  9. read_stream: Prevent distance from decaying too quickly

  10. read_stream: Issue IO synchronously while in fast path

  11. bufmgr: Return whether WaitReadBuffers() needed to wait

  12. aio: io_uring: Allow IO methods to check if IO completed in the background

  13. bufmgr: Make UnlockReleaseBuffer() more efficient

  14. Add fake LSN support to hash index AM.

  15. Make IndexScanInstrumentation a pointer in executor scan nodes.

  16. Use fake LSNs to improve nbtree dropPin behavior.

  17. Move fake LSN infrastructure out of GiST.

  18. Use simplehash for backend-private buffer pin refcounts.

  19. nbtree: Avoid allocating _bt_search stack.

  20. bufmgr: Fix use of wrong variable in GetPrivateRefCountEntrySlow()

  21. Conditional locking in pgaio_worker_submit_internal

  22. Reduce ExecSeqScan* code size using pg_assume()

  23. Fix rare bug in read_stream.c's split IO handling.

  24. Remove HeapBitmapScan's skip_fetch optimization

  25. Optimize nbtree backwards scans.

  26. Fix multiranges to behave more like dependent types.

  27. Add EXPLAIN (MEMORY) to report planner memory consumption

  28. Optimize nbtree backward scan boundary cases.

  29. Increment xactCompletionCount during subtransaction abort.

  30. Add nbtree Valgrind buffer lock checks.

  31. Add nbtree high key "continuescan" optimization.

  32. Reduce pinning and buffer content locking for btree scans.

  33. Teach btree to handle ScalarArrayOpExpr quals natively.