HeliosDB Nano — Pipeline Performance Baseline Report

HeliosDB Nano — Pipeline Performance Baseline Report

Date: 2026-02-13 Test: cargo test --test pipeline_performance_test -- --nocapture Hardware: Linux 5.14.0-611 x86_64

Phase Distribution Summary

In-Memory Mode

Phase	Time	% of Total
Parse	1.7ms	0.2%
Plan	1.8ms	0.2%
Execute	518ms	61.9%
Commit	314ms	37.4%
Other	1.1ms	0.1%

Persistent (RocksDB) Mode

Phase	Time	% of Total
Parse	1.9ms	0.2%
Plan	1.9ms	0.2%
Execute	759ms	70.2%
Commit	316ms	29.2%
Other	1.1ms	0.1%

Baseline Throughput (In-Memory)

Statement	Wall Time	Parse	Plan	Execute	Commit	Rows	ops/sec
CREATE TABLE	3.2ms	65us	45us	2.8ms	280us	0	306
ALTER TABLE ADD COL	4.2ms	67us	44us	817us	3.2ms	0	240
DROP TABLE	4.2ms	63us	39us	1.1ms	3.0ms	0	236
INSERT (single)	4.3ms	61us	38us	1.2ms	3.0ms	1	233
INSERT (bulk 100)	445ms	-	-	-	-	100	2
UPDATE (single)	7.7ms	57us	48us	4.3ms	3.3ms	1	131
UPDATE (bulk WHERE)	17.3ms	55us	39us	12.3ms	4.9ms	180	58
DELETE (single)	8.5ms	52us	42us	5.0ms	3.3ms	1	118
DELETE (bulk WHERE)	197ms	53us	40us	194ms	3.1ms	50	5
SELECT * (full scan)	7.1ms	59us	85us	6.8ms	0	1000	140
SELECT WHERE	5.5ms	72us	91us	5.3ms	0	180	184
SELECT WHERE id=	5.2ms	55us	79us	5.0ms	0	1	198
SELECT LIMIT 10	4.4ms	67us	79us	4.3ms	0	10	226
SELECT projection+filter	5.5ms	75us	87us	5.3ms	0	334	183
COUNT(*)	4.9ms	73us	96us	4.7ms	0	1	203
AVG/SUM/MIN/MAX	6.2ms	134us	104us	6.0ms	0	1	161
GROUP BY	6.8ms	124us	95us	6.6ms	0	50	145
GROUP BY + HAVING	6.4ms	112us	96us	6.2ms	0	50	155
ORDER BY DESC	12.9ms	73us	93us	12.8ms	0	1000	77
ORDER BY (multi-col)	13.0ms	104us	101us	12.8ms	0	1000	75
INNER JOIN	11.0ms	152us	147us	10.7ms	0	150	90
LEFT JOIN	11.8ms	127us	155us	11.5ms	0	150	85
CTE	6.3ms	152us	107us	6.0ms	0	1	158
Window (ROW_NUMBER)	5.8ms	129us	91us	5.5ms	0	100	175
UNION ALL	11.2ms	128us	146us	10.9ms	0	1000	90
IN (subquery)	23.0ms	130us	155us	22.7ms	0	200	43

Key Bottlenecks

1. Execution dominates reads (95-99% of wall time for SELECT)
2. Commit dominates writes in persistent mode (44-69% for DDL/DML)
3. Operator BUILD 4x slower than EXEC (117ms build vs 29ms exec)
4. Bulk UPDATE has 6.4x disk penalty (17ms in-memory vs 110ms persistent)
5. IN (subquery) is slowest query (23ms — subquery materialization)

Improvements Implemented

All 5 planned improvements have been implemented. See PERFORMANCE_IMPROVEMENTS.md for full results.

1. Plan/operator caching — LRU cache (256 entries) for SQL → LogicalPlan
2. Batch commit — execute_batch() wraps N statements in single transaction
3. Index-based point lookups — ART index for WHERE pk=value (12.8x faster)
4. SQL parse cache — LRU cache (512 entries) for SQL → AST Statement
5. RocksDB write path optimization — pipelined writes, buffer tuning