Skip to content

Probe identity-keyed expression caches without cloning the key - #10615

Open
robert3005 wants to merge 3 commits into
developfrom
rk/exact-key-probes
Open

robert3005 wants to merge 3 commits into
developfrom
rk/exact-key-probes

Conversation

@robert3005

Copy link
Copy Markdown
Contributor

Every split of a scan prunes with each filter conjunct, and each layout
reader on the way looks the conjunct up in a cache keyed by tree
identity. Each lookup cloned the expression into a key, hashed the
options of its function, such as literal lists, and dropped the key
again. An ExactKey trait lets these maps be probed with a borrowed expression,
the hash uses only the identity of the children, and the struct reader
lends its cached partitioning instead of cloning it.

@robert3005 robert3005 added the changelog/performance A performance improvement label Oct 19, 2026
@codspeed

codspeed Bot commented Oct 19, 2026 •

Copy link
Copy Markdown

Merging this PR will improve performance by 25.23%

2 warnings

⚠️ Walltime results may be unreliable

Some benchmarks used the Walltime instrument outside CodSpeed Macro Runners, which can make the results inconsistent.

Learn more about Macro Runners

See affected benchmarks

⚠️ 1 benchmark is dominated by system calls

System calls cannot be simulated consistently and are left out of the measure. The reported value understates the real cost. Switch to the Walltime instrument to measure them.

Learn more about system calls

See affected benchmarks

⚡ 14 improved benchmarks
✅ 1968 untouched benchmarks
⏩ 57 skipped benchmarks1

Performance Changes

Mode Benchmark BASE HEAD Efficiency
⚡ Walltime compare_u8_neon 2.6 µs 1.6 µs +62.33%
⚡ Walltime compare_int_constant_neon 4.5 µs 3.2 µs +40.88%
⚡ Walltime compare_int_nullable_neon 5.7 µs 4.2 µs +37.8%
⚡ Walltime compare_u64_neon 5.1 µs 3.8 µs +35.13%
⚡ Walltime compare_int_neon 5.1 µs 3.8 µs +34.53%
⚡ Walltime compare_f32_neon 5.5 µs 4.1 µs +33.53%
⚡ Walltime compare_int_eq_neon 5 µs 3.8 µs +31.2%
⚡ Walltime infallible_bool_neon[i32, PerRowPerRow] 3.9 µs 3.3 µs +17%
⚡ Walltime compare_float_neon 6.9 µs 6 µs +16.14%
⚡ Walltime decode_blocked_bitpacked_chunked_neon[u32, 524288] 30.3 µs 27 µs +12.33%
⚡ Walltime scalar_subtract_neon 13.5 µs 12.1 µs +12.3%
⚡ Walltime deferred_bool_constant_neon[ConstantPerRow] 4.5 µs 4.1 µs +10.95%
⚡ Walltime infallible_bool_neon[i64, PerRowPerRow] 5.7 µs 5.1 µs +10.69%
⚡ Walltime infallible_bool_constant_neon[ConstantPerRow] 4.5 µs 4 µs +10.42%

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.


Comparing rk/exact-key-probes (4e1f8a4) with develop (3da99e9)

Open in CodSpeed

Footnotes

  1. 57 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

@robert3005
robert3005 marked this pull request as draft October 20, 2026 12:09
@robert3005
robert3005 force-pushed the rk/exact-key-probes branch 2 times, most recently from f34b6ff to 1251b6b Compare October 26, 2026 23:40
@robert3005
robert3005 marked this pull request as ready for review October 26, 2026 23:42
@github-actions

Copy link
Copy Markdown
Contributor

This PR has been marked as stale because it has been open for 14 days with no activity. Please comment or remove the stale label if you wish to keep it active, otherwise it will be closed in 7 days

@github-actions github-actions Bot added the stale This PR is stale and will be auto-closed soon label Nov 16, 2026
@robert3005 robert3005 removed the stale This PR is stale and will be auto-closed soon label Nov 16, 2026
robert3005 and others added 3 commits November 19, 2026 15:34
Every split of a scan prunes with each filter conjunct, and each layout
reader on the way looks the conjunct up in a cache keyed by tree
identity. Each lookup cloned the expression into a key, hashed the
options of its function, such as literal lists, and dropped the key
again. The clones touch reference counts that every scan thread shares,
and on TPC-DS q39 on DuckDB they were more than half of the pruning
time.

An ExactKey trait lets these maps be probed with a borrowed expression,
the hash uses only the identity of the children, and the struct reader
lends its cached partitioning instead of cloning it. Pruning in q39
takes a third of the samples it took before.

Signed-off-by: Robert Kruszewski <robert@spiraldb.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Robert Kruszewski <github@robertk.io>
DuckDB scans of local files reuse the layout readers of a file across
queries. The readers cache pruning results, partitioned expressions and
evaluated dictionary values by the identity of the expression, and every
query binds new expressions, so each query added entries that no later
lookup could hit. The file statistics reader held every filter it pruned,
which kept the expressions alive for all the caches below it. On
ClickBench q20 the process grew by about 30 MiB per query, and the pages
of each new allocation were faulted in and zeroed.

Key these caches with WeakBoundExprKey, which identifies an expression as
ExactBoundExpr does but holds its children weakly, and drop the entries of
dead expressions when inserting a new one. The weak reference keeps the
allocation of the children, so a later expression cannot take their
address and match a stale key. ExactKey now compares identities so that
both kinds of key can be probed with a borrowed expression.

Memory of 13 runs of q20 stays at 190 MiB instead of growing to 540 MiB,
and q20 and q21 take 16% and 21% less time.

Signed-off-by: Robert Kruszewski <robert@spiraldb.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015qFEJyt9gzF3mi2F91ZwML

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

changelog/performance A performance improvement

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant