Clean code practices, system design, architecture patterns, and development methodologies
Debuggability is not an afterthought. The system properties that make an incident faster to diagnose — and how to bake them in at design time rather than retrofit them after the first bad page.
Software bills of materials are required for some government contracts and increasingly expected for any product with security claims. Generating a useful SBOM for embedded C firmware.
Head-based sampling at 1% misses the rare slow requests you actually care about. Tail-based sampling solves this and introduces a different set of operational challenges.
Event sourcing solves some auditability problems cleanly and creates new ones around schema evolution and event replay performance. An honest retrospective.
Layer caching, test parallelization, and selective test execution moved our GitHub Actions build from 22 to 7 minutes. The one optimization we thought would help but didn't.
We migrated to OpenTelemetry from three different proprietary SDKs. The cardinality explosion from naively porting existing metrics nearly took down our Prometheus cluster.
Deep dive into microservices architecture with practical design patterns, implementation strategies, and real-world examples for building scalable distributed systems.
Redis-backed job queues are fast but lose jobs on restart without persistence configuration most tutorials skip. Postgres-backed queues with SKIP LOCKED solve the durability problem.
Bazel's hermetic builds eliminate the 'it works on my machine' problem for a repo that includes both embedded C and Python test harness code. The setup that worked and the rules_cc pain.
Serverless functions open a new database connection on every invocation. PgBouncer in transaction mode solves this elegantly for Postgres. RDS Proxy adds IAM auth but introduces latency.
The Istio / Envoy sidecar adds measurable latency. On p99 for our mix of RPCs, the overhead was 4 ms — acceptable for some services and not for others.
Feature flags accumulate. After eighteen months and three product releases, our flag system had 140 active flags and three engineers who understood any given flag's interaction graph.