Your P99 TTFT spikes every few minutes on an otherwise...

A P99 TTFT spike in this context is typically caused by a KV eviction storm. When the decode batch fills available blocks, admission control pauses new requests. When a large batch of long sessions completes, it releases a sudden burst of blocks. This triggers the admission of a large queue of new requests at once, causing the prefill stage to become saturated. The fix is to implement admission rate limiting, which spreads re-admissions over multiple scheduling steps rather than unblocking the entire queue at once.