Cinch is in public alpha.
// PRODUCT
Performance
Same-region benchmarks from production. Redis TCP reads at 1.4ms p50. Wake from S3 in 33ms.
Steady-state latency
1,000 iterations, active database, same Fly.io region (IAD)
| Protocol | Read p50 | Write p50 | Read p99 | Write p99 |
|---|---|---|---|---|
| Redis TCP | 1.4ms | 1.9ms | 3.7ms | 7.4ms |
| Redis REST | 2.3ms | 2.8ms | 6.1ms | 8.8ms |
| SQLite | 5.4ms | 5.4ms | 9.2ms | 10.7ms |
| Graph / Bolt | 44.9ms | 45.3ms | 83.9ms | 52.0ms |
Wake latency
10 iterations, database woken from stopped or archived state
| Scenario | REST p50 | TCP p50 | TCP p99 |
|---|---|---|---|
| Wake from stop | 4.4ms | 10.7ms | 11.6ms |
| Wake from archive (S3) | 4.6ms | 32.6ms | 63.1ms |
Three storage tiers
HOT
Active
Database loaded in memory. Serving reads and writes at full speed.
1–3ms reads
WARM
Stopped
Data on local NVMe. No memory used. Wakes on first request.
5–11ms wake
COLD
Archived
Data in S3-compatible storage. No local disk used. Restored on demand.
30–63ms wake
METHODOLOGY
- Benchmark app deployed on Fly.io in the same region (IAD) as the engine
- Steady-state: 1,000 iterations per protocol, pre-warmed database
- Wake: 10 iterations, database stopped/archived between each
- Timing via
time.perf_counter()(Python, microsecond precision) - Benchmark code is open source in the
benchmark/directory
Fast enough. 85% cheaper.
Sub-2ms reads. Sub-second cold starts. All on NVMe instead of RAM.
GET STARTED →