V3 delta-search dashboard

Final 2026-09-14 experiment: one million keywords per distribution, three independent processes per mode on Redis 8.10.1 and Valkey 9.1.2. Each ratio compares the median delta-search result with the same full-rebuild baseline. The validation report explains the workload, release gates, and reproduction commands.

The dashboard also includes the archived V3 baseline and R2/R3 measurements below. They use a different benchmark schema, so their metadata is shown separately from the final delta-search release gate.

FailedRelease gate

410 / 1110Search p95 failures

30 / 30Ready-time passes

2 / 6Peak-RSS failures

By distribution

A ratio above its gate fails. RSS is the process high-water mark after compaction.

Server	Keywords	p95 failures	Worst p95 ≤ 1.10×	Worst ready ≤ 0.30×	Peak RSS ≤ 1.00×
Redis	Shared	79 / 185	2.47×	0.164×	0.827×
Redis	Diverse	66 / 185	2.21×	0.090×	0.925×
Redis	Korean	57 / 185	2.61×	0.091×	1.552×
Valkey	Shared	79 / 185	2.62×	0.164×	0.967×
Valkey	Diverse	77 / 185	2.43×	0.089×	0.941×
Valkey	Korean	52 / 185	1.99×	0.089×	1.527×

API p95 failures

Each API has 45 paired comparisons per server across inputs and operations.

API	Redis	Valkey
`Find`	17 / 45 17 / 45	17 / 45 17 / 45
`FindSet`	22 / 45 22 / 45	19 / 45 19 / 45
`FindIndex`	19 / 45 19 / 45	21 / 45 21 / 45
`FindMatches`	14 / 45 14 / 45	18 / 45 18 / 45
`Contains`	19 / 45 19 / 45	20 / 45 20 / 45
`FindStream`	23 / 45 23 / 45	26 / 45 26 / 45
`FindBatch`	18 / 45 18 / 45	20 / 45 20 / 45
`FindParallel`	10 / 45 10 / 45	9 / 45 9 / 45
`FindIndexParallel`	7 / 45 7 / 45	4 / 45 4 / 45
`Scan`	21 / 45 21 / 45	18 / 45 18 / 45
`MaskText`	17 / 45 17 / 45	17 / 45 17 / 45
`ReplaceText`	13 / 45 13 / 45	15 / 45 15 / 45

The raw evaluator results are in Redis JSON and Valkey JSON. `DeltaSearch` remains opt-in.

Benchmark sets

These archived measurements use the original scale-test schema and are shown as context, separate from the final delta-search gate.

Set	Date	Servers	Runs	Status	Configuration
V3 baseline	2026-09-06	8.10.1 / 9.1.2, source-built, prefetch-batch-max-size=0	36	Recorded	Original versioned scale test
R2/R3 verification	2026-09-06	8.10.1 / 9.1.2, source-built, prefetch-batch-max-size=0	36	Recorded	RefreshDebounce=20ms

The final delta-search JSON files and these two archived result sets remain available under test/benchmarks/results for reproducibility.