
Software Architecture Insights · Tuesday · 9 min
The Limits Nobody Graphed
0:00-9:30
transcript
show notes
GitHub published five incidents for August 2026. Three of them are the same failure.
On the 17th, seven and a half hours of degraded service: 29,000 organizations, 4.8 million requests failed or slow, 56% of front-door requests failing at the peak. Three weeks, three incidents, one blind spot — capacity ran out in places nobody was graphing.
Your autoscaler scales the thing you are observing. Saturation happens in the thing you don't.
In this episode:
- The August 17th outage, and the two beside it on the 6th and the 26th
- Sidecar concurrency, network flow tables, database primary headroom: limits with no dashboard
- Why these limits are dangerous: someone else's layer, hard rather than soft, invisible
- Retry amplification, and how one data center's capacity problem became an authentication outage everywhere
- Protective throttles set so high they engaged after the damage was done
- Two lists to build this week: your real hard limits, and your retry code
Links:
links3