At a glance
- Did you know
- Decoding is memory-bound: each request holds its KV cache in GPU memory for as long as it generates.
- Rule of thumb
- Watch KV-cache memory, not just GPU utilisation; it sets how many requests decode at once.
- Watch out
- KV-cache memory fills; batches shrink and output slows.
Its settings in ScaleMySystems
What you can change on it in the lab, and what each number does.
- Concurrency (slots)
- Calls it can work on at once (per instance when autoscaling). Needed slots ≈ calls per second × seconds each takes. Extra calls wait in the queue, then are rejected.
- Queue capacity
- Calls allowed to wait for a free slot. A bigger queue absorbs bursts but adds waiting time; 0 rejects straight away.
- vCPU
- CPU cores available. Needed ≈ calls per second × CPU per request ÷ 1000. Above about 90% busy it slows down and may crash.
- CPU per request (ms)
- CPU time each call burns. Part of what the component is; make the code cheaper to lower it.
- Latency (p50) (ms)
- Typical time one call takes here, before any calls it makes itself. Slots are held for this long.
- Latency (p99) (ms)
- The slowest 1% of calls. Callers must set their timeout above this or they cut off healthy calls.
- Timeout to dependencies (ms)
- How long it waits for each call it makes before giving up. Set it above the dependency’s p99 but low enough that a dead dependency does not hold every slot.
- Error rate
- Share of calls that fail on their own (bugs, declines, flaky tools). An agent that calls it on every step multiplies this.
- Quota (req/min)
- Calls allowed per minute (an API key or provider limit). Calls over it get 429s. Needed = calls per second × 60, with headroom.
Problems that start with it
- Prefill is the bottleneck Serving is split into prefill and decode pools (disaggregated serving). Prompts got longer: the prefill pool can’t keep up, and time to first token suffers.
- Decode is the bottleneck Disaggregated serving with a small decode pool: answers got longer and decode slots fill up.
- ChatGPT: from launch build to real traffic A chat product on launch-week hardware: the GPU prefill and decode pools, the KV cache store, moderation and the chat API are an eighth of the size real traffic needs. 300 chats a second arrive with long prompts. Answers queue for 20 seconds and most fail.
Full lesson coming soon
Lessons are being added one component at a time: how it works, its settings, a live exercise to break and fix it, interview questions and a short check. See the Circuit breaker lesson for what one looks like.