RedBoxDb Benchmark Results Platform: Windows x64 | Compiler: MinGW g++ | Config: Release
CPU Feature: AVX2 enabled
Threads: 12
Indexing: K-Means Clustering (K=10, Probes=1)
RNG Seed: 42 (fixed <-- results are reproducible)
Parameter
Value
Vectors
100,000
Dimensions
128
Queries per search test
1,000
Data size
48.8 MB
Metric
Value
Time
0.099 s
Throughput
~1,010,000 vectors/sec
Data written
48.8 MB
RNG generation excluded from timing.
[2/6] Search Latency <-- Single Nearest Neighbor
Hot-cache performance (dataset paged in by Bench 1)
Metric
Value
QPS
~1,140 queries/sec
Min
0.685 ms
Avg
0.874 ms
P50
0.821 ms
P95
1.163 ms
P99 ~1.46 ms
Max
2.106 ms
[3/6] Search Latency <-- Top-10 Nearest Neighbors (search_N)
Metric
Value
K
10
QPS
~1,062 queries/sec
Min
0.708 ms
Avg
0.939 ms
P50
0.867 ms
P95
1.325 ms
P99 ~1.61 ms
Max
2.202 ms
search_N is only ~7% slower than single search on average <-- priority queue overhead remains negligible vs scan cost.
[4/6] Update Throughput <-- O(1) via id_to_index
Metric
Value
Updates
1,000
Throughput
~276,000 updates/sec
Min
0.000 ms
Avg
0.003 ms
P50
0.003 ms
P95
0.004 ms
P99 ~0.005 ms
Max
0.017 ms
Direct hash lookup via id_to_index <-- no linear scan of stored vectors.
[5/6] Mixed Workload <-- 70% Search / 20% Insert / 10% Delete
Metric
Value
Total ops
10,000
Searches
6,959
Inserts
2,043
Deletes
998
Total time
~1.24 s
Throughput
~8,040 ops/sec
[6/6] Search Under Heavy Deletion <-- 40% of DB Deleted
Metric
Value
Inserted vectors
100,000
Deleted vectors
40,000
Live rows
60,000
QPS
~1,980 queries/sec
Min
0.404 ms
Avg
0.502 ms
P50
0.482 ms
P95
0.609 ms
P99 ~0.86 ms
Max
1.198 ms
Compare against Bench 2 to estimate deleted_flags overhead under heavy tombstoning.
Operation
Throughput
P99 Latency
Insert
~1,010,000 /sec
—
Search (top-1)
~1,140 QPS
~1.46 ms
Search (top-10)
~1,062 QPS
~1.61 ms
Update (indexed)
~276,000 /sec
~0.005 ms
Mixed workload
~8,040 ops/sec
—
Heavy deletion search
~1,980 QPS
~0.86 ms
b