Lessons learned: 40k requests a day on a single 24GB card

What self-hosted LLM serving cost us, what it bought us, and the four things I would change tomorrow.

The short version

continuous batching tripled throughput while p99 latency moved by 40ms

What we tried

Move	What happened
Bigger context window	Cost up 3x, accuracy flat
Prompt rewrite on its own	+4% on the golden set, then regressed two weeks later
Swapping vLLM for a boring alternative	Same quality, fewer moving parts

What broke

KV-cache memory, not model weights, is what actually OOMs your node

Behind almost every incident was a default we inherited and never questioned. A token limit too small. A timeout copied from an HTTP tutorial. A retry loop with no ceiling.

By the numbers

• p50 looked fine the entire time; p99 is where users actually lived
• 60% of "the model is wrong" tickets were retrieval returning the wrong document
• Rolling back a prompt was faster than fixing one — every single time

What I would do differently

1. Instrument first: inputs, outputs, latency, token counts, on day one
2. Build a 200-example golden set before the second feature, not after the fifth
3. Make every external call idempotent and deadline-aware
4. Read my own traces for an hour a week, on the calendar

The uncomfortable lesson

Most of what we shipped was ordinary backend engineering wearing an AI hat. That is it. No framework required.