RFD 3701
Edge cache warmup strategy
We have repeated launch-day latency spikes in regions where the CDN has little historical traffic. This proposal explores a warmup plan that prefetches top index and RFD detail pages before public announcements. The rollout would start with one low-traffic region and compare p95 TTFB, cache hit ratio, and origin error rate over two release cycles. If we see a measurable improvement without origin saturation, we can standardize this as part of release prep and document guardrails for large content drops.
Discussion Comments
1. @lowlydba ↗ ·
I like the phased rollout and the focus on p95 instead of averages. One thing we should add is a hard stop condition based on origin saturation, not just latency. If cache warmup accidentally creates a surge of misses in a cold region, we need a circuit breaker that disables prefetch quickly and records the failure mode so future launches can avoid repeating it.
2. @lowlydba ↗ ·
Can we include a comparison between synthetic prefetch traffic and real user traffic shape? The current draft assumes they are close enough, but past incidents showed different object popularity distributions in the first hour. If we model that gap early, we can choose a safer warmup set and avoid over-optimizing for pages users rarely request at launch time.