Avoiding Goodhart's Law - Use SLO's as Tools Not Cudgels
summary
Hospital lab system story shows a real Goodhart failure: a monitor for queue length >100 becomes the target while transactions time out before the queue, so patient impact continues. Uses APM like AppDynamics, Datadog, New Relic and percentiles not averages, builds nested SLIs (code → infra → customer), sets a 28-day sliding SLO with weekly error‑budget tracking, negotiates SLAs that include refunds or freezes, and chooses meaningful customer‑facing metrics (repeat lookups, 5‑minute guarantee vs 10‑minute nurse run).
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:38#about_marco
- 3:14Labs Processing: MQ Flow
- 3:39Labs Processing: Queue Failure
- 4:50Labs Processing: TX Failure
- 7:12The SLI, SLO, SLA Model
- 10:25Code Example
- 14:12Dimensions
- 15:13Infrastructure Example
- 19:10Business & CX Example
- 23:09Prepare to Engage
- 26:36Negotiation Flow
- 30:47Avoiding Goodhart's Law
- 32:42Links & More
related talks
32m9 Things You Should Do When Starting to Use SLOs
18mSLOs as an Organisational 'Check Engine' Light
47mCultivating Production Excellence
15mShit, our Site is Down! Let's build a culture of reliability
14mRun, Walk, Crawl, or How We Failed Our Way to SLO Readiness
30mReliable Observability at Scale: Error Budgets for 1,000+
20m9 SLIs; Oh My!
23mSLO's for Openstack - the API for Engineering Teams