mirror of
https://github.com/hanzoai/datastore.git
synced 2026-09-26 13:55:07 +00:00
Nikita TaranovandClaude Opus 4.8
b77566dd60
Probes 50M rows against a ~1M-distinct set on a composite fixed key via (a, b) IN (subset), for a 4-byte key (two UInt16 -> keys32) and an 8-byte key (two UInt32 -> keys64). Locally, versus the keys128 baseline: keys32 0.129s -> 0.066s (1.95x), keys64 0.129s -> 0.084s (1.54x). Co-Authored-By: Claude Opus 4.8 <[email protected]>
16 lines
977 B
XML
16 lines
977 B
XML
<test>
|
|
<!--
|
|
A set over a composite fixed key that fits in 4 or 8 bytes now uses the packed keys32 / keys64
|
|
methods instead of collapsing to keys128. The per-key slot becomes 4 or 8 bytes instead of 16,
|
|
which shrinks the hash set (better cache behaviour) and uses cheaper hashing/comparison, so both
|
|
building the set and probing it get faster. The set has ~1M distinct keys and is probed by 50M
|
|
rows via `(a, b) IN (subset)`.
|
|
-->
|
|
|
|
<!-- keys32: two UInt16 columns (4 bytes). -->
|
|
<query>SELECT count() FROM numbers_mt(50000000) WHERE (toUInt16(number), toUInt16(intHash32(number))) IN (SELECT toUInt16(number), toUInt16(intHash32(number)) FROM numbers(1000000))</query>
|
|
|
|
<!-- keys64: two UInt32 columns (8 bytes). -->
|
|
<query>SELECT count() FROM numbers_mt(50000000) WHERE (toUInt32(number), toUInt32(intHash64(number))) IN (SELECT toUInt32(number), toUInt32(intHash64(number)) FROM numbers(1000000))</query>
|
|
</test>
|