LLM cost per request tripled because retries and prompt bloat, not traffic

Lance_Saul_85 · reddit · 2026-07-21

A production team found its cost per request tripled even though traffic stayed flat — and the culprit was retries plus prompt bloat, not demand.

• After instrumentation, they discovered a timeout retry path that could fire three full-price calls for one request in some failure modes.
• They also found their shared system prompt had silently grown to nearly 4k tokens over months as people kept appending guardrails without removing older ones.
• The retry bug is fixed; the remaining problem is organizational: nobody wants to delete another person’s prompt line, so the team is asking for a more boring, automatic process to keep the prompt under control.
• The post is a practical reminder that LLM spend often comes from hidden engineering debt, not just usage growth.

Original post →

More from Infra

• Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
• 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
• Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
• Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11
• LLM Serving Metrics Thread: Why TPOT and Uptime Make or Break User Experience — abhijithneil · 2026-09-11
• PlanetScale launches sharded Postgres: 768 servers acting as one, 1PB scale — dhruv2038 · 2026-09-11

Infra channel →