$ ls ./blog # notes and case studies
Writing
Notes and case studies from building production systems. What was tried, what cost, and what changed.
- Draining, not stopping
ECS scale-in was killing 2-hour translation jobs. So workers drain before they stop.
- One engine, three providers
Rate limits caused 65% of review failures. Treating them as "retry later" stopped provider outages from becoming ours.
- 7.33M rows per call
One status broadcast ran 1,013 queries and pinned Aurora at 100% CPU. The fix was not an index. EXPLAIN showed each call scanning 7.33M rows.
- The verdict lives in code, not the prompt
Almost trusted a plausible explanation. So pass/fail had to live in tested code.