RPC failover benchmark
How often Ethereum RPC calls succeed when an upstream returns HTTP 429 or 502, and how long requests take to finish. The controlled test and production observations measure different environments.
Controlled comparison
Measured UTC on darwin/arm64 with Atlas revision 54e0881f. Each row covers 100 simulated Ethereum mainnet eth_call requests in a local, single-process test.
| Setup | First upstream | Succeeded / total | p95 response | p99 response | p95 recovery |
|---|---|---|---|---|---|
| Single upstream | HTTP 429 | 0 / 100 | 11.04 ms | 11.7 ms | — |
| Sequential two-upstream fallback | HTTP 429 | 100 / 100 | 32.07 ms | 33.77 ms | 21.04 ms |
| RouteMesh Atlas | HTTP 429 | 100 / 100 | 32.67 ms | 35.73 ms | 21.49 ms |
| Single upstream | HTTP 502 | 0 / 100 | 11.03 ms | 11.04 ms | — |
| Sequential two-upstream fallback | HTTP 502 | 100 / 100 | 32.05 ms | 32.58 ms | 21.03 ms |
| RouteMesh Atlas | HTTP 502 | 100 / 100 | 32.3 ms | 33.82 ms | 21.23 ms |
Production observations
Ethereum mainnet HTTP eth_call batches from to UTC. The first recorded upstream attempt returned HTTP 429 or 502.
| Strategy | Batches | Success rate | Succeeded via another provider | Router regions | p95 response | p99 response |
|---|---|---|---|---|---|---|
| economy | 13,194 | 98.98% | 11,473 | 3 | 702.7 ms | 1,222.1 ms |
| performance | 113 | 98.23% | 111 | 2 | 97 ms | 460 ms |
Methodology
In the controlled test, the primary returns an error after a scripted 10 ms delay, and the backup returns a valid response after 20 ms. We run the Atlas request coordinator with those same simulated upstreams and compare it with a single call and a sequential retry. Response percentiles include errors. Recovery measures time from the first error to a successful response and is reported only for successful requests. Timing includes local process scheduling. There are no network or provider latency measurements in this test.
The production snapshot covers one complete UTC hour after at least one hour of telemetry ingestion and counts each recorded batch ID once. It includes strategies with at least 100 qualifying batches. Success means the batch was fulfilled with a 2xx HTTP status and no JSON-RPC error. The alternate provider column counts successful batches whose selected provider differs from the first attempted provider. Response latency includes all outcomes and measures time at the router, excluding client network transit. Attempts can overlap, so the production data describes observed outcomes rather than the incremental time or effect of a retry.