2026-06-11 · Engineering
Edge reads: what actually sits in the cache
Read nodes are not full copies of the data. They are caches with a clear rule set: what is admitted, what is evicted, and how a range request is satisfied from one replica.
Admission
An object enters a read node's cache after two requests within five minutes — the classic two-hit rule. One-off reads never displace hot media, which keeps the cache useful during traffic spikes.
Eviction
Eviction is by bytes, not by count, weighted by last-access time. A 4 GB video and a 40 KB thumbnail do not compete on equal terms, so the node tracks byte cost per hit rather than hits alone.
Range requests
Players almost always ask for the first few hundred kilobytes, then seek. A range request is satisfied from a single replica: if the range is partially cached, the node fetches the missing span from the origin node and stitches it, rather than refetching the whole object. Median time to first byte from a warm node across the EU is under 30 ms, and a cold range costs one extra round trip inside the region.