mirror of
https://github.com/hanzoai/datastore.git
synced 2026-09-26 13:55:07 +00:00
Nikolai KochetovandClaude Opus 4.6
16a6c5fa1f
Benchmark `ReplacingMergeTree` FINAL with a concentrated filter (`category = 1`, 50k out of 10M keys) and evenly distributed updates (every 100th key has version 2, defeating `split_intersecting`). Uses separate partitions to prevent merges from combining parts. Query reads full payload via `cityHash64(payload)`. Compares `query_plan_optimize_lazy_final = 0` (median 154ms) vs `= 1` (median 117ms) — 24% improvement. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
51 lines
3.1 KiB
XML
51 lines
3.1 KiB
XML
<test>
|
|
<!--
|
|
Test lazy FINAL optimization for ReplacingMergeTree.
|
|
The optimization builds a set of primary key values from filtered rows
|
|
and uses it for index analysis to prune parts/granules in the FINAL read.
|
|
It helps when:
|
|
- Updates are evenly distributed (defeating split_intersecting optimization)
|
|
- The filter narrows to a small key range (enabling PK pruning)
|
|
- The query reads full column data
|
|
-->
|
|
<settings>
|
|
<max_insert_threads>1</max_insert_threads>
|
|
</settings>
|
|
|
|
<create_query>
|
|
CREATE TABLE lazy_final_perf
|
|
(
|
|
part UInt8,
|
|
key UInt64,
|
|
version UInt64,
|
|
category UInt8,
|
|
payload String
|
|
)
|
|
ENGINE = ReplacingMergeTree(version)
|
|
PARTITION BY part
|
|
ORDER BY key
|
|
SETTINGS index_granularity = 256
|
|
</create_query>
|
|
|
|
<!-- 10 partitions, 1M rows each, non-overlapping key ranges.
|
|
category=1 only for keys 0..49999 (in partition 0). -->
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 0, number, 1, if(number < 50000, 1, 0), repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 1, number + 1000000, 1, 0, repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 2, number + 2000000, 1, 0, repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 3, number + 3000000, 1, 0, repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 4, number + 4000000, 1, 0, repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 5, number + 5000000, 1, 0, repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 6, number + 6000000, 1, 0, repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 7, number + 7000000, 1, 0, repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 8, number + 8000000, 1, 0, repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT 9, number + 9000000, 1, 0, repeat('x', 200) FROM numbers(1000000)</fill_query>
|
|
<!-- Evenly distributed updates across all partitions: every 100th key gets version 2. -->
|
|
<fill_query>INSERT INTO lazy_final_perf SELECT toUInt8(number * 100 / 1000000), number * 100, 2, if(number * 100 < 50000, 1, 0), repeat('y', 200) FROM numbers(100000)</fill_query>
|
|
|
|
<!-- Concentrated filter + full payload read. Optimization should prune to ~2 parts. -->
|
|
<query>SELECT count(), max(cityHash64(payload)) FROM lazy_final_perf FINAL WHERE category = 1 SETTINGS query_plan_optimize_lazy_final = 0</query>
|
|
<query>SELECT count(), max(cityHash64(payload)) FROM lazy_final_perf FINAL WHERE category = 1 SETTINGS query_plan_optimize_lazy_final = 1, max_rows_for_lazy_final = 10000000</query>
|
|
|
|
<drop_query>DROP TABLE IF EXISTS lazy_final_perf</drop_query>
|
|
</test>
|