Learn with Phoebe · Data Ops & Security

Learn Data Reliability with Phoebe

Every reliability conversation that goes badly goes badly for the same reason: somebody made a promise before anybody made a measurement. Harborlight's team believed their orders pipeline was basically 99 percent. Counted by partition it was 95.14, and the difference is a month of promises the history could not keep. This course builds the contract properly: indicators for partitions, a target the history supports, an error budget that decides what ships, alerting that pages once per incident instead of thirty-five times, and the runbooks, backfills and postmortems that make a rota survivable. Six leader sessions on the promise, the budget and the humans; eight builder sessions on one pipeline, measured, promised, alerted, documented and replayed over thirty nights on call.

14sessions, two tracks
720hourly partitions on the bench
35pages from alerting on every breach
5pages from burn-rate alerting, all useful
45min per session

Leader track 📈 · six sessions, no SQL

For whoever funds the work and signs the settlement report. What a promise with a number changes, what the budget buys, and why the rota is the last decision rather than the first.

🤝 Leader session 1 · start here

A promise with a number

SLI, SLO and SLA as the reliability literature defines them, why 100 percent is the wrong target, and 95.14 against the 99 everybody believed.

▶ Start here 💰 Leader session 2

Error budgets buy decisions

One hundred minus the promise, as bad partitions allowed. 95 percent ends with 3 percent left; 99.9 is gone on night 4.

 📟 Leader session 3

On-call is a design, not a rota

35 pages a month with five useful is a design problem, not a staffing one. The humane rota and why it is also the effective one.

 🔍 Leader session 4

Postmortems that change something

Blameless as the entry condition, actions with an owner and a date, and the gap only five postmortems read together reveal.

 🧱 Leader session 5

Backfills and debt as reliability work

Why the backlog that never shrinks belongs in the budget conversation.

 📊 Leader session 6

Reporting reliability upward

Budget remaining, pages per week and toil, monthly, forever. The report that does not need a meeting.

the board version

Builder track 🔧 · eight sessions, one reliability contract

One pipeline, Harborlight's orders, measured, promised, alerted, documented and replayed. Session 8 lives the month under every target and policy.

📏 Builder session 1 · start here

SLIs for one pipeline

Freshness, correctness and coverage from the data-processing menu, specification against implementation, and 685 of 720.

▶ Start here 🎯 Builder session 2

SLOs from history, not hope

Start loose, tighten later. Four targets and the night each one runs out of budget.

 🔥 Builder session 3

Error-budget arithmetic and burn-rate alerts

Fast and slow windows, the rising edge, and 35 pages against 5 for the same five incidents.

 📖 Builder session 4

Runbooks a tired person can follow

The first three steps, what goes in, what stays out.

 ⏪ Builder session 5

Backfills without breaking downstream

Idempotency, watermarks, replay, and the settlement report that has already been signed.

 📦 Builder session 6

The incident package

Timeline in budget terms, impact, actions with owners and dates.

 🧹 Builder session 7

Toil: measuring and buying it down

What counts as toil, the toil budget, and what to automate first.

 🎛️ Builder session 8

The on-call bench

Thirty nights replayed under four targets and three policies. Two levers, and neither moves the other.

🔒 the capstone
start here foundations hands-on advanced and capstone

Choose your path 🗺️

Three doors into the same contract. Both tracks start at their own session 1.

💼 I fund this and sign the report L1 → L2 → L3 → L6 the promise, the budget, the rota, the monthly report
🔧 I carry the pager B1 → B2 → B3 → B4 → B8 measure, promise, alert once per incident, know what to do, then live the month
🎯 Everything, in order L1 → L4 → L5 → B1 → B5 → B6 → B7 → B8 from basically 99 to a contract the history supports

Your passport 🎫

Clear all three quiz questions in a session and it gets stamped. Fourteen stamps across both tracks, and they live in this browser only.

What is real here 🧾

A course about promises should be exact about which of its own numbers are measured.

ClaimStatus
The 720 partitions and 35 failuresFixed and identical for every learner: 30 nights, 24 hourly partitions, five injected incidents on nights 4, 9, 15, 21 and 27 with 3, 14, 6, 8 and 4 bad partitions
95.14 percent achieved685 of 720. Arithmetic
Budgets and exhaustion nights100 percent minus the target, times 720. 95 percent: 36 allowed, 3 percent left. 97: 21.6, out on night 15. 99: 7.2, night 9. 99.9: 0.72, night 4. Computed by the bench
35 pages against 5Threshold paging fires on every bad partition. Burn-rate paging fires when a 2-hour window burns at 14.4x or a 12-hour window at 3x, on the rising edge, never twice inside 12 hours. Those windows and rates are the bench's own stated choices
Actionable pagesDefined mechanically as the first page inside a real incident: 5 of 35, and 5 of 5. Nothing about the human is modelled
The two levers do not interactChange the target and the page counts stay the same; change the policy and the budget stays the same. Both visible on the bench
SRE definitionsQuoted from the SRE book chapter on SLOs and the SRE workbook chapter on implementing SLOs, as recorded in the course map. The workbook's dedicated alerting chapter is not quoted
HarborlightA fictional marketplace. The incidents are ordinary ones; the company is not real

The knowledge map 🧠

Both tracks at a glance - hover a session to spotlight its concepts, click any node to jump in.