Performance Benchmarks
Every gate run measures end-to-end performance using Playwright (headless Chromium) against a live Node.js server on a random port. Each scenario runs for 3 seconds of sustained load. Web Worker transport and pako compression are enabled (defaults).
Latest results
Loading benchmark data...
Throughput over time
Median latency over time
Main-thread impact: Worker vs no-worker
Main-thread blocking over time
Console errors & warnings over time
What's measured
Each benchmark connects a real browser client (datasole IIFE bundle in headless Chromium) to a live Node.js server via WebSocket and runs sustained load for 3 seconds:
| Scenario | Description |
|---|---|
| RPC echo | Sequential ds.rpc(TestRpc.Echo, payload) — full round-trip latency (wire name still echo) |
| RPC concurrent | 10 concurrent ds.rpc() calls per batch — multiplexed throughput |
| RPC small payload | ds.rpc() with a tiny JSON body (<256 B, below compression threshold) |
| RPC large JSON | ds.rpc() with randomized JSON (~1 KB, above compression threshold) — exercises pako |
| Binary frame streaming | Server pushes 1 KB binary-like frames at max rate — simulates audio/video metadata streaming |
| Two-way low-latency emit | Client emits, server echoes — measures full round-trip at emit speed (game ticks, trade confirms) |
| Server event receive | Server broadcasts at max rate, client counts received events |
| Live state sync | Server mutates state rapidly, client receives JSON Patch diffs |
| Client event emit | Client fires ds.emit() in a tight loop — fire-and-forget throughput |
| CRDT increment | Client increments a PN counter rapidly — CRDT op throughput |
| Mixed workload | Alternating RPC, emit, and add — combined throughput under mixed load |
Additionally, 6 main-thread comparison benchmarks run each workload twice (with and without Web Worker transport) and measure Long Tasks, rAF jitter, and jank frames to quantify the off-main-thread benefit.
Every benchmark also records the number of browser console errors and warnings emitted during the test. These are first-class correctness indicators — any non-zero count under sustained load signals a runtime issue (unhandled rejection, deprecation, failed assertion, etc.) that should be investigated.
All benchmarks run in isolation (separate Playwright process, own browser instance) as part of the CI gate on every push to main.