MLB Edge Bench

Methodology. Both columns issue parallel fetch() requests every 5 seconds from this browser. The direct path returns MLB's full GUMBO feed (~400–800 KB). The edge path returns a slimmed snapshot served from a Cloudflare Worker, backed by a 6-second caches.default cache that shares state across viewers in the same colo. Lower bytes also means faster decompression and parse on the client. p50/p95/p99 are computed over a rolling 60-sample window.

Caveat — these numbers are local to you. Latency observed here is shaped by your ISP, the route your packets take, and which Cloudflare colo Anycast happens to land you in. A user in Mumbai and a user in Newark will see very different absolute numbers from the same dashboard. Treat the spread between columns as the signal, not the absolute milliseconds. A production-grade version of this would log every request server-side via Cloudflare Analytics Engine and report percentiles across all real users — that's on the roadmap.

Cache hit rate matters. The edge column surfaces hit rate because latency wins are misleading without it. If most requests are misses, the Worker is still paying the upstream cost — the colo cache fronts it, but a cold colo still hits MLB. With ~6-second TTL and steady polling, the hit rate climbs quickly and most viewers from the same colo get HIT responses served directly from edge memory. The miss latency is what the direct column would look like with a slimmer payload; the hit latency is what the edge actually buys you.

Why p99, not just average. p99 = the slowest 1 in 100 requests. For a live game feed updating every pitch, that's the user whose refresh stalls right when a runner crosses home. Average latency hides this — a service with great average and bad p99 still feels broken to the unlucky 1%. Tail latency is what fans actually notice, and it's the metric edge networks are designed to compress.