Case Studies Highlights Summary
Every highlight bullet from the 8 pages in this section, gathered on one page and grouped by the page it came from.
- A case study documents a decision made under real constraints, not a template to copy verbatim
- The constraints section matters more than the resulting architecture diagram
- Case studies decay as your context diverges from theirs - check the date and the stack version first
- Metrics separate evidence from anecdote - adjectives without numbers aren't a case study, they're marketing
- Modular monolith with clear package boundaries
- Pydantic v2 at API and worker edges
- Postgres + Alembic expand-contract migrations
- Celery for async jobs; structlog + OTel
- Docker image Python 3.14, uv lockfile
- Versioned datasets and model artifacts
- Training pipeline separate from serving API
- Batch features vs online features documented
- Eval gates before promote
- Drift monitoring on prod predictions
- Idempotent stages with watermark checkpoints
- Orchestrator owns schedule and retries
- Polars/pandas for transform; load to warehouse
- Data contracts on schema changes
- Observability on row counts and freshness
- Strangler fig over big-bang rewrite
- Extract domain services before microservices
- Expand-contract migrations throughout
- Module boundaries enforced with import-linter
- CFR dropped after removing circular imports
- Migrate reads first behind flag
- Async SQLAlchemy pool tuning critical
- Never block loop with sync ORM
- p95 dropped 65% at 3k RPS
- Rollback was flag off, not redeploy
- Benchmark production-like data volume
- Measure p50/p95/p99, not averages
- Include warm-up and steady-state phases
- Record pool, CPU, and GC metrics
- Compare before/after with same harness
- Organized by domain: web, data, ML, ops
- Each lesson links to guardrail
- Sourced from real incident patterns
- Onboarding required reading
- Reviewed quarterly for new entries
Revisado por Chris St. John·Última actualización: 31 jul 2026