Engineering notes / caching

Why we jitter cache TTLs

A synchronized-TTL stampede in March made it clear that uniform expiry is a liability at scale, and jittered TTLs combined with request coalescing at the cache layer are now the standard mitigation. Neither change is exotic on its own; together they moved p99 origin load from a recurring incident to a non-issue.

The rest of the win came from elsewhere. Edge response times improved considerably after last quarter’s cache changes, and parsing overhead on the misses that remained dropped once a streaming approach replaced the buffered one we started with.