Architecture NotesPractical Monitoring
Monitor the service, not the server.
If you run anything in production, read this once. It rewires how you think about monitoring.
If you touch operations or maintenance, read Practical Monitoring (Mike Julian, O'Reilly). It teaches the essence of what to monitor, in the language of people who have carried a pager.
Readable
O'Reilly books have a reputation: excellent but dense, easy to abandon halfway. This one is different — remarkably readable, and the more production experience you have, the harder its points land.
The core line
What's wrong with 100% CPU? If the service is healthy, nothing. Collect metrics in detail for diagnosis — but that is not what you monitor. Monitor whether the service is alive, not the server.
See the service
- 100% CPU with normal responses and no user pain is not an incident
- A health check should prove "queries reach the database and return correctly," not "the server process is up"
Keep operations humane
- Alert only on things that need action — daily noise trains people to ignore alerts
- Rotate on-call — a single permanent first responder burns out
One step past "100% CPU means an incident," toward "is the service healthy for the user?" This book sharpened where the Observe of 3.4.1 Observability should point.
Practical Monitoring: Effective Strategies for the Real World — Mike Julian, O'Reilly Media. (Japanese edition: 『入門 監視』, O'Reilly Japan.)