PERFORMANCE

Numbers we can reproduce.

Every figure on this page comes from a measured run - YCSB 0.17, published hardware, raw output in the repo. No projected numbers.

By the numbers

Results that speak for themselves.

Benchmarked on real cloud hardware — GCP N2 nodes with 8 NVMe SSDs, 64 cores, 480 GB RAM.

85%
Reduction in read latency
From 33ms down to 4.9ms P99
10×
Less unnecessary rewriting
Compaction only touches keys, never values
3–5×
Lower infrastructure cost
vs DynamoDB at equivalent scale
1B+
Keys on a 3-node cluster
256 GB cache · 3 TB NVMe per node

P99 read latency at 1 billion keys — head to head

Lower is better. Sustained mixed read/write load. Equivalent hardware.

— ms (P99) →
Cassandra
Vendor-published typicals · wide-column · JVM GC pauses included
~80ms
~80ms P99
LSM-tree DB
RocksDB-class store · compaction stalls dominate the tail
~33ms
~33ms P99
DynamoDB
AWS-published single-digit-ms · per-request billing
~12ms
~12ms P99
VeltrixDB · miss
Measured · one NVMe read via io_uring · no compaction in the read path
4.9ms
4.9ms · cache miss
VeltrixDB · hit
Measured · LIRS DRAM cache · zero disk I/O
0.3ms
0.3ms · cache hit
SystemP99 read latencyvs VeltrixDB missSource
Cassandra~80 ms16× slowerVendor-published typical
LSM-tree DB (RocksDB-class)~33 ms6.7× slowerVendor-published typical
DynamoDB~12 ms2.4× slowerAWS-published
VeltrixDB · cache miss4.9 msbaselineMeasured (YCSB 0.17)
VeltrixDB · cache hit0.3 ms16× fasterMeasured (YCSB 0.17)
VeltrixDB achieves <5ms P99 even on cache misses because values are read directly from NVMe via zero-copy io_uring — no compaction, no page-cache eviction, no surprises.
Transparency

Live cluster, live metrics.

A snapshot from one of our internal benchmark clusters. Every number on this page traces back to a metric you can read in Prometheus.

veltrixdb · cluster-prod-01 · us-east-1 Snapshot · May 21 2026 · 14:42 UTC · uptime 187 d
Active keys
1.04B
↗ +2.1M / hr
Reads scrubbed (lifetime)
108B+
14.3k/s sustained
Writes processed
3.42B+
WAL group-commit 80µs
P99 read latency
4.9 ms
stable for 187 d
Read latency · P50 / P95 / P99last 24h
Ops/sec per shard1024 shards
Cache hit rate 94.2% · Write amp 1.0× · SLA breaches 0 Exposed via /metrics · Prometheus + Grafana ship in the Helm chart
Active Development · June 2026

What's new in the engine.

Five features live in the current build — each one targeting a specific latency or cost problem. Every change ships with a benchmark gate so you can see exactly what it does to the numbers.

01
Shipped
I/O Engine
Zero-interrupt polling I/O
VeltrixDB's I/O layer runs in polling mode — the engine continuously monitors the completion queue rather than waiting for kernel interrupts. Under sustained write load this eliminates wakeup overhead entirely, flattening tail latency at the I/O boundary where most databases lose P99 predictability.
100%
0
syscalls
per read op
02
Shipped
Garbage Collection
Cold-aware garbage collection
VeltrixDB's GC engine ranks VLog entries by access temperature before deciding what to compact. Hot data is never relocated during a GC pass — only cold entries move. GC runs without competing for the NVMe bandwidth live reads depend on, and emergency compaction runs have dropped to zero in sustained benchmarks.
100%
0
GC emergency
runs · bench verified
03
Shipped
CPU Scheduler
Dedicated compaction core
VeltrixDB pins its compaction thread to a reserved CPU core. Compaction work never contends with request-handling threads for scheduler time — the intermittent P99 tail spikes that appear under heavy writes on shared-core designs simply don't exist in this architecture.
100%
±3%
P99 jitter
under load
04
Shipped
Resource Control
eBPF-backed GC throttle
A built-in BPF program watches CPU throttle counters in real time and signals VeltrixDB's GC scheduler before a throttle event hits the write path. Sustained write peaks stay smooth — hard stalls become gradual backpressure, and the GC emergency counter stays at zero even on resource-constrained nodes.
100%
0
stalls in
60-min stress
05
Active
Storage Tiering
Cloud cold-tier offload — S3, GCS, Azure
VeltrixDB's tiering engine automatically moves data that hasn't been accessed in over 5 minutes to object storage, rate-limited to avoid I/O contention. The NVMe index stays hot. Cold data lands on S3, GCS, or Azure at a fraction of NVMe cost — no manual configuration required, and the GC engine is aware of tiered entries so it never relocates them unnecessarily.
88%
10×
cheaper cold
storage vs NVMe
Sustained writes
110k ops/s
8 workers · 200K keys · 128B values
P99 write latency
162ms
per 1024-key batch · 158µs/key amortised
Packing density
160B
bytes/record · gate ≤ 182.4 B
Gate PASS
GC emergency gate
0
emergency runs in 60-min stress test
Gate PASS

Don't trust our numbers. Reproduce them.

git clone github.com/VeltrixDB/veltrixdb && ./scripts/bench.sh
Hard pass/fail gates on packing density and GC pressure · exit 0 = the numbers held on your hardware