LLM Ops Cookbook — 30 production recipes

The field manual for LLM workloads in production — every recipe tested on a live multi-provider gateway serving 33 models.

Stop 429 cascades with sliding-window pacing that doubles throughput on the same quota
Kill the silent 180-second hang: stream idle timeouts, circuit breakers, failover ladders with sticky windows
Cut cost 5x with model tiering — extraction on cheap, judgment on strong — plus per-task cost accounting
Swap models without gambling: golden-set regression gates and a four-task quality bench you can run in 15 minutes

¥39

≈ $19 · 支付宝或信用卡

Choose how to pay

支付宝 / 微信 — ¥39（扫码即付）

国内最快：扫码付款 → 填支付宝交易单号 → 几分钟内出下载链接。无需账号。

中文说明

LLM Ops Cookbook

30 篇 LLM 生产配方：限流节拍、JSON 修复、多供应商故障转移、模型降本、缓存、评测门禁，加两个真实事故复盘。

付款后怎么拿到：扫码付款 → 填支付宝交易单号 → 一般几分钟内自动出下载链接（营业时段更快）。关掉页面也能用订单号回来查。
交付物：30 篇 markdown + 目录 + 使用说明
跑不起来怎么办：回复订单说明你的环境，我修好或退款。一次付费，不限设备数，不再收费。

Not sure yet? Take the free sample first

Download 3 full recipes free — no payment, no signup, instant zip.
Recipe 001 (rate-limit pacing), 029 (the 15-second messaging window), 030 (the postmortem template).
If they save you one bad night, the other 27 are worth 19 USDT.

About paying

This store takes USDT or USDC on six EVM chains (BSC, Ethereum, Polygon, Arbitrum, Base, Optimism) — that is what keeps it account-free and instant. You need any wallet that can send USDT or USDC on one of those chains (Trust Wallet, MetaMask, Rabby, Binance, OKX). Pick the chain in the checkout and it shows the exact token address to send. The order page shows the exact amount and a scannable address; paying that exact amount unlocks the download automatically. No USDT? The free sample needs no payment at all — and every free tool on this site runs offline in your browser.

The six parts

Throughput — pacing, key rotation, failover ladders, bounded fan-out for 1000-item jobs
Contracts — JSON repair loops, tool-call normalization, streaming partial JSON, determinism
Cost engineering — model tiering, token budgets, cache layers, free-tier ladders
Reliability — circuit breakers, watchdogs, idempotent checkpoints, structured logs
Evaluation — golden sets, the four-task bench, latency percentiles, A/B routing
War stories — a live 16-channel gateway dissected: real numbers, six named bugs, two postmortems

What you get

30 recipes in markdown, each one page: Problem → Why it bites → Recipe (runnable code) → Checklist.
Node.js and Python implementations that run on free tiers first-class — no paid API key required to reproduce anything.
Windows-specific background-service recipes (job objects, BOM traps, hidden windows) you will not find anywhere else.

Who wrote it

Every recipe comes out of a live system: a free-tier gateway doing real daily work, scheduled jobs that must survive bad nights, and a messaging product with a 5-second reply window on 20–180s upstreams. The bugs are real, the numbers are real, the fixes shipped.