Reproducible before clever
Infrastructure defined in code and applied from a pipeline removes the class of incident that begins with someone remembering what they changed last month.
Boring deploys
Small, frequent, reversible. A deploy that ships one change is diagnosable; a deploy that ships forty is an investigation.
- Automated rollback path exercised on a normal day, not first in an incident.
- Health checks that fail loudly instead of degrading quietly.
- Migrations that are backwards compatible for one release.
The observability minimum
Request rate, error rate and duration per endpoint, plus structured logs with a correlation id. That set answers the first three questions of nearly every incident.