Skip to content
WeftKitBeta
Performance

Numbers Don't Lie.
Benchmarks Do.

Every claim backed by reproducible benchmarks. Same hardware, default configs, published Docker images. Run them yourself.

~116 ns
In-Memory p99
WeftKitMem point read
~204 ns
Point get p99
WeftKitKV (1K keys)
~2M
ops/sec
WeftKitKV throughput
~234 μs
Flat search
100 vectors top-10
Head-to-Head

WeftKit vs. Industry Leaders

Direct comparisons against the most widely deployed alternatives for each database type.

WeftKitRel

SQL parse (complex)

vs. SQLite
Lower is better
WeftKit~5.1 μs
SQLite~15 μs
WeftKitRel

SQL parse (simple SELECT)

vs. SQLite
Lower is better
WeftKit~1.6 μs
SQLite~5 μs
WeftKitKV

Point get p99

vs. RocksDB
Lower is better
WeftKit~204 ns
RocksDB~15 μs
WeftKitKV

Point write p99

vs. RocksDB
Lower is better
WeftKit~360 ns
RocksDB~12 μs
WeftKitKV

Throughput (single thread)

vs. RocksDB
Higher is better
WeftKit~2M ops/sec
RocksDB~800K ops/sec
WeftKitMem

Point read p99

vs. Redis (network)
Lower is better
WeftKit~116 ns
Redis (network)~5 μs (network)
WeftKitMem

Throughput (single thread)

vs. Redis (network)
Higher is better
WeftKit~8.6M ops/sec
Redis (network)~500K ops/sec
WeftKitVec

HNSW search 2K vectors

vs. Qdrant
Lower is better
WeftKit~31 ms
Qdrant~10 ms
WeftKitVec

Flat search 100 vectors

vs. Qdrant
Lower is better
WeftKit~234 μs
Qdrant~1.5 ms
WeftKitDoc

Insert one

vs. MongoDB (embedded)
Lower is better
WeftKit~3.1 μs
MongoDB (embedded)~50 μs
WeftKitDoc

Insert 1000 docs

vs. MongoDB (embedded)
Lower is better
WeftKit~15.2 ms
MongoDB (embedded)~50 ms
WeftKitDoc

Find all (10K docs)

vs. MongoDB (embedded)
Lower is better
WeftKit~37 ms
MongoDB (embedded)~80 ms
WeftKitGraph

BFS 1000 vertices

vs. Neo4j
Lower is better
WeftKit~160 μs
Neo4j~500 μs
WeftKitGraph

Vertex creation p99

vs. Neo4j
Lower is better
WeftKit~2.5 μs
Neo4j~50 μs
Memory Efficiency

Footprint Per Module

Base memory usage at idle. Minimum configuration (no indexes, no cache) to maximum typical production deployment.

Min → Max Memory (MB)

WeftKitRel
4–88 MB
WeftKitVec
8–256 MB
WeftKitKV
6–28 MB
WeftKitDoc
8–32 MB
WeftKitGraph
12–128 MB
WeftKitMem
8–64 MB
WeftKitMD
4–32 MB
WeftKitFile
6–16 MB
Pool Manager

Connection Proxy Performance

io_uring-powered proxy handling 1M+ concurrent connections with sub-microsecond acquisition.

1M+
Concurrent connections
per instance
~3.3 ns
Protocol detect (Redis)
p99 latency
~39 ns
Cache hit lookup
p99 latency
~86 ns
Queue push + pop
p99 latency
< 1 KB
Idle connection memory
per connection
~744 ns
Health record
per success
Under the Hood

Kernel-Level Performance

Raw engine, security, and utility layer measurements via Criterion bench harness. These are the building blocks powering every database module.

EngineBuffer pool pin (hit)
~23 ns
EngineMVCC begin + commit
~128 ns
EngineMVCC write + commit
~690 μs
EnginePage read (sequential 8K)
~476 ns
EngineWAL append (async)
~38 ns
SecurityAES-256-GCM encrypt 512 B
~5.1 μs
SecurityHMAC-SHA256 sign 4 KB
~12.4 μs
SecurityRBAC permission check
~28 ns
UtilityCRC32C 1 KB
~56 ns
UtilityxxHash64 64 KB
~4.4 μs
UtilityBLAKE3 hash 4 KB
~2.3 μs
UtilityLZ4 compress 8 KB
~1.1 μs
UtilityLZ4 decompress 8 KB
~1.8 μs
DocumentInsert one
~3.1 μs
DocumentInsert 1000 docs
~15.2 ms
DocumentAggregate (match + limit)
~686 μs
KeyValuePoint get (1K keys)
~204 ns
KeyValuePoint write
~360 ns
GraphBFS 1000 vertices
~160 μs
GraphDFS 500 vertices
~65 μs
GraphVertex creation
~2.5 μs
MarkdownFull document parse
~17 μs
MarkdownBM25 score (1K docs)
~18.6 μs
MarkdownRender with TOC
~22 μs
FileStoreUpload 4 KB chunk
~3.7 ms
FileStoreDownload 256 B
~59 μs
FileStoreGC cycle (1000 chunks)
~996 μs
VectorHNSW search (2K vectors)
~31 ms
VectorFlat search (100 vectors)
~234 μs
InMemoryPoint get (per op)
~116 ns
InMemoryPoint set (per op)
~151 ns
Docker Deployment

Direct vs. Containerized Performance

Docker overhead averages 4.3% across all operations. CPU-bound: ~3.5%, I/O-bound: ~6.2%. Maximum observed: 7.9%.

Docker Engine 24.x, debian:bookworm-slim, --cpus=2 --memory=4g2026-03-02
EngineBuffer pool pin (hit)
+4.3%
Direct~23 ns
Docker~24 ns
EngineMVCC begin + commit
+3.9%
Direct~128 ns
Docker~133 ns
EngineMVCC write + commit
+3.6%
Direct~690 μs
Docker~715 μs
EnginePage read (seq 8K)
+7.1%
Direct~476 ns
Docker~510 ns
EngineWAL append (async)
+7.9%
Direct~38 ns
Docker~41 ns
SecurityAES-256-GCM encrypt 512B
+2%
Direct~5.1 μs
Docker~5.2 μs
SecurityHMAC-SHA256 sign 4KB
+2.4%
Direct~12.4 μs
Docker~12.7 μs
UtilityCRC32C 1KB
+1.8%
Direct~56 ns
Docker~57 ns
UtilityLZ4 compress 8KB
+3.6%
Direct~1.1 μs
Docker~1.14 μs
KeyValuePoint get (1K keys)
+3.9%
Direct~204 ns
Docker~212 ns
KeyValuePoint write
+6.9%
Direct~360 ns
Docker~385 ns
InMemoryPoint get (per op)
+3.4%
Direct~116 ns
Docker~120 ns
InMemoryPoint set (per op)
+4.6%
Direct~151 ns
Docker~158 ns
DocumentInsert one
+6.5%
Direct~3.1 μs
Docker~3.3 μs
DocumentInsert 1000 docs
+5.9%
Direct~15.2 ms
Docker~16.1 ms
GraphBFS 1000 vertices
+4.4%
Direct~160 μs
Docker~167 μs
GraphVertex creation
+6%
Direct~2.5 μs
Docker~2.65 μs
MarkdownFull document parse
+2.9%
Direct~17 μs
Docker~17.5 μs
VectorFlat search (100 vec)
+3.4%
Direct~234 μs
Docker~242 μs
FileStoreUpload 4KB chunk
+6.8%
Direct~3.7 ms
Docker~3.95 ms
PoolProtocol detect (Redis)
+3%
Direct~3.3 ns
Docker~3.4 ns
PoolCache hit lookup
+2.6%
Direct~39 ns
Docker~40 ns
+3.5%
CPU-Bound Avg
+6.2%
I/O-Bound Avg
+4.3%
Overall Avg
+7.9%
Max Observed
Methodology

How We Benchmark

Transparency and reproducibility are non-negotiable. Every benchmark is fully reproducible and independently verifiable.

Environment

Measured inside the published Docker images on arm64 (Apple M-series) and x86-64 hosts with NVMe SSD storage. Images built with release profile + LTO.

Warmup

Criterion default: 3-second warmup, 100 iterations minimum. Cold-cache and warm-cache results reported separately.

Percentiles

p50, p95, p99, p999 latency reported. Charts show p99 (worst-case typical). Max latency excluded.

Competitors

Competitor numbers use published reference values under official default configurations. No detuning.

Reproducible

Every benchmark runs against the same published Docker tags listed on the Downloads page. Pull the exact same image and replay the workload to reproduce our numbers.

Disclaimer

All WeftKit numbers are measured against the published Docker images over the loopback wire protocol — the exact topology customers deploy. No in-process bench tricks.