On 19 November 2026, from 14:10 to 15:20 UTC, ElevenLabs refused about 10% of new ElevenAgents conversations overall.
Their write-up for incident INC-465 explains what happened. A server hosting one of their caches reached its limit for new network connections, and connections to that cache timed out. ElevenAgents kept its active-call counts in that cache. Calls that ended during that window weren't consistently removed from the counts. The counts reached customers' concurrency limits, and new conversations were turned away.
ElevenLabs cleared the stale counts at 15:17 UTC. It has since raised the connection limit on every server that hosts these caches.
The day before was the larger part of the same incident: the voice caching cluster kept restarting with an empty cache, and requests failed across most products. I covered that side, and why the miss path matters more than the hit path, in The cache hit isn't the interesting part.
The 19 November problem was different. The cache was holding a live count that the business used to decide who gets served. When updates to that count didn't land, customers were refused because of calls that had already ended.
The fixes in the write-up focus on capacity and limits: larger caching clusters, limits inside their services, the raised connection limit, and a review of the caching layer across their engineering teams.
The write-up doesn't say whether the active-call count now expires on its own, or whether it can be rebuilt from a record of which calls are actually live. I don't know, and I won't guess.
I worked on an image pipeline for a US consumer app that served more than 50 million images a day. The edge cache and R2 only ever held copies. The origin stayed the source of truth and generated any variant that was missing. When an image changed, a key inventory in D1 listed every derived object to purge, so invalidation was an indexed operation instead of a reconstruction during an incident. Image-path origin CPU went from around 90% to below 15%.
Every copy in those layers could be regenerated from the origin.
If a cache or delivery problem like this is sitting in your business, I take that work through Developer Labs.