Week 24 · Season 7 · Alive
Running for real
You can monitor FRIDAY for real people, catch drift early, swap its model with a gate and a canary, run an incident and write a blameless postmortem, and keep its logs private and its labels honest.
This week
- Green tests, angry users: monitoring
- The slow leak: drift
- Swap the engine mid-flight: migration and canary releases
- 2:07 a.m.: incidents and blameless postmortems
- Logs that leak: privacy and AI rules
- Go deeper: SLOs, error budgets, shadow traffic, EU AI Act
How you learn
Read a week of FRIDAY's log for quiet failures, score drift by hand, and rehearse the incident routine before a real one happens.
What you build
Dashboards and alerts for FRIDAY: four signals and an SLO on one screen, alarms that wake you only for real problems, logs with private data removed, and a runbook you've practised.
Stretch: Move FRIDAY to a new model behind a gate, shadow traffic and a canary, with a rollback rule written first, and write up what you saw.
What you ship
Project 7: The public demo. FRIDAY live for strangers, with a system diagram, a cost sheet and a demo you can defend.
- OPS.md with four signals, an SLO and two alert rules, plus a dashboard
- Redacted logs with a 30-day timer, and tests that prove it
- A runbook and a blameless postmortem of a practice incident
Lessons
- Green tests, angry users
After this deck you can explain why FRIDAY can pass every test and still fail real people, name the four golden signals of monitoring, and use quiet user signals and an online eval to find what's going wrong.
- The slow leak
After this deck you can explain drift (the world moves, FRIDAY stays the same), compare this week's questions with last month's using one drift score, and set an alarm that rings for a real slide, not for one bad day.
- Swap the engine mid-flight
After this deck you can plan a model migration: pin a dated model, gate the new one on your own test cases slice by slice, and roll it out as a canary release with a rollback rule written first.
- 2:07 a.m.
After this deck you can run an incident in the right order (take charge, stop the harm, tell people, take notes, then find the cause) and write a blameless postmortem that fixes the system instead of blaming a person.
- Logs that leak
After this deck you can find personal data in FRIDAY's logs, apply redaction before anything is written, keep less for less time, and explain the basics of the GDPR and the EU AI Act (with official sources).
- Go deeper: budgets, shadows and the rulebook
After this deck you can write a service level objective for FRIDAY, turn it into an error budget and a burn-rate alarm, test a new model with shadow traffic, and find your way around the official EU AI Act sources.
Labs
Python in the browser, next to the lessons.
Weekend build
- FRIDAY, week 24: dashboards and alerts
One screen that says whether FRIDAY is OK for real people, alarms that wake you only for real problems, and logs that keep no secrets.
Free links for this week
Hand-picked official docs, papers and talks.
- Google SRE: Monitoring distributed systems · docs
- Google SRE workbook: Alerting on SLOs · docs
- Google SRE workbook: Canarying releases · docs
- Google SRE: Postmortem culture · docs
- Google SRE: Managing incidents · docs
- Evidently docs (drift) · docs
- OpenTelemetry docs · docs
- EU AI Act (official text) · docs
- AI Act Service Desk: implementation timeline · docs
- GDPR (official text) · docs