Notes
Lessons from systems that were slow, stateful, distributed, or simply more complicated than they first appeared.
One small change. The whole pipeline.
How a change classifier selected only the build and test jobs each edit needed, cutting build execution for isolated changes by 65–90% while keeping a full route for uncertain changes.
Read note · 7 minThe database was 3 GB. So why was the 32 GB disk full?
How a stalled Debezium replication slot retained 30 GB of PostgreSQL WAL, why quieter environments were more vulnerable, and why doubling the disk was only recovery, not a fix.
Read note · 4 minZero functions, two different causes
How a missing container image and a stale management API produced the same signal, and why deployment verification had to follow the work into the runtime.
Read note · 5 minCI is not a monitoring system
Why we moved simple probes out of Azure DevOps, consolidated checks that needed application context, and cut scheduled agent usage by roughly 70% for the workloads we changed.
Read note · 4 minReducing Thousands of Vulnerability Findings Without Hiding the Risk
How cleaning stale container images, remediating what was running, tightening access, and documenting exceptions made vulnerability management more useful.
Read note · 7 minThe Fastest Query Is Sometimes the One You Never Make
How projecting smaller read models, then caching representations instead of entity graphs, brought slow customer reports back under their performance target.
Read note · 6 min
