Leader track 🌙 · six sessions, no SQL
For whoever funds the work and reads the report. What data downtime costs, which signals are worth paying for, and the three numbers that say whether the programme is working.
What breaks, and who finds out first
Data downtime as a number rather than a feeling: what a wrong table costs while people use it, and the discovery gap between the team that owns it and the stakeholder who reports it.
📡 Leader session 2Five signals, five ways to find out first
Freshness, volume, schema, distribution and lineage: what each one sees, what each one is blind to, and which four cost nothing but a query.
📏 Leader session 3Quality is a dimension with a number
DAMA-DMBOK's six dimensions against ISO/IEC 25012's fifteen, and how to pick the three your business actually pays for rather than measuring all of them badly.
🔔 Leader session 4Alert or noise: the real cost of a threshold
Why the second month of a monitoring programme is where it dies, what a false page costs, and how to read a coverage number that was bought with a muted channel.
🪜 Leader session 5Coverage without headcount
Tiering tables so the 20 percent that carries the business gets watched properly and the long tail gets watched cheaply. The decision that keeps this affordable.
📈 Leader session 6From firefighting to a trust score
The 90-day plan and the three numbers to report: incidents caught before a stakeholder, mean hours to detect, and false pages per week.
Builder track 🔧 · eight sessions, one monitor set
One table, orders_daily at Bellwether, instrumented signal by signal. Session 8 runs the finished set against 30 nights of history and counts what it caught.
The metadata you already have
Four numbers every warehouse already stores, the SQL that pulls them, and the anatomy of a monitor that somebody actually answers at 07:00.
⏱️ Builder session 2Freshness: on-time is a distribution
Why a fixed deadline is the wrong shape for the question, trailing percentiles that learn the table's real habits, and the full-rebuild trap.
🧮 Builder session 3Volume and uniqueness
Same-weekday medians against seasonality, robust bands with MAD, and the duplicate rate that catches what a healthy row count hides.
🧬 Builder session 4Schema change: the break you can see coming
Column-set diffs, type widening, the difference between an additive change and a contract break, and why the revert pages too.
📊 Builder session 5Distribution drift from metadata only
Null rates, cardinality, min and max, segment share. How far you get without reading a row, and the point where this becomes a different discipline.
✅ Builder session 6Expectations as code
Great Expectations Core end to end: Data Context, Batch Definition, Suite, Validation Definition, Checkpoint and Actions, wired to the metrics table.
🕸️ Builder session 7Lineage and blast radius
Column-level lineage with OpenLineage, the impact query that turns an alert into a triage decision, and the reverse-ETL job nobody documented.
🎛️ Builder session 8The monitor bench
Thirty nights, five real defects, seven monitors, three numbers. Hair-trigger everything and detection falls back to where it started.
Choose your path 🗺️
Three doors into the same practice. Both tracks start at their own session 1, because the vocabulary is what makes the rest arguable.
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 detecting quiet failures should be exact about which of its own numbers are measured.
| Claim | Status |
|---|---|
| The 30 nights of load metadata | Generated once from a fixed seed and identical for every learner. Row counts, load lateness, duplicate rates, null rates and column sets, with a real weekday and weekend pattern |
| The five defects | Injected into specific nights: a 230-minute late load, a load holding 41 percent of a normal count, a duplicate rate of 8.4 percent, a dropped column, and a null rate of 11.4 percent. On a defect night every other signal is held at a quiet baseline, so credit for a catch can only go to the monitor watching that pillar |
| The ladder: 0, 1, 2, 2, 3, 4, 5 of 5 caught | Computed by the bench in your browser, night by night. Trailing percentiles, same-weekday medians, duplicate rates, column-set diffs and standard deviations are all real arithmetic on that history |
| Eight false pages from a fixed volume band | Arithmetic. Weekend volume is about 45 percent of a weekday, which sits outside a plus-or-minus 30 percent band on all eight weekend nights in the window |
| Tightening everything: 2 of 5, 25 false pages, 15.8 hours | The counts and the thresholds are real. The consequence runs through one modelled rule, below |
| The muted channel | Modelled, not measured. Six or more false pages in a trailing week is treated as a channel nobody reads, so the page fires and the stakeholder still finds the problem first. Labelled as a model on the widget itself |
| Great Expectations component names | Read from the GX Core 1.x documentation: Data Context, Data Source, Data Asset, Batch Definition, Expectation, Expectation Suite, Validation Definition, Checkpoint, Actions, Data Docs |
| DAMA-DMBOK's six dimensions and ISO/IEC 25012 | Both named as published, including the point that uniqueness appears in DAMA and not in the ISO inherent set. The ISO standard itself is paywalled, so wording travels through close secondary sources and is flagged where it does |
| Bellwether | A fictional grocery delivery company. The failure patterns 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.