06-reference

practical data modeling mma book club week3 chapters 7 9

2026-10-29·reference·source: Practical Data Modeling (Substack)·by Joe Reis

"Week 3 — Counting Tickets Without Counting Them Twice" — @practicaldatamodeling

Why this is in the vault

Third live-cohort post in the Mixed Model Arts re-read (Chapters 7–9, discussion target Nov 4), direct sequel to Week 1 and Week 2: it moves from "are these the same entity" to "what happens when you join and count across entities that legitimately relate many-to-many" — the exact mechanism behind the fan-out failure already filed in this vault, now told as a discussion exercise rather than a SQL postmortem.

The core argument

Chapters 7–9 cover relationships, grain, and aggregation as one connected unit: Ch 7 gives the ways to represent connections between entities, Ch 8 forces you to declare what one record actually represents, and Ch 9 asks whether the groups and measures you're computing still mean what you think once you've joined across a relationship. Reis's running example: a support ticket that mentions two products and contains ten chat messages. Join Tickets, TicketProducts, and the message transcript together and count rows, and the ticket count balloons — impressive, and wrong, because the customer contacted the business about exactly one ticket. He extends the same logic to rates: a team that resolves 9 of 10 tickets and a team that resolves 1 of 2 don't have a combined resolution rate of 70% just because that's the average of 90% and 50% — you have to preserve the underlying numerators and denominators (10 of 12, not an average of two percentages) to recompute a correct rolled-up rate.

The discussion questions push the same point: what does one record represent in Tickets vs. TicketProducts vs. a message-level transcript table (three different grains); how do you label product-level ticket counts so a reader understands why they won't sum to the distinct ticket total; and which components of a resolution rate have to survive a roll-up so it can be recalculated rather than re-averaged. Logistics (Nov 4 target week, paid-subscriber live session details) are pure scheduling and skipped here.

Mapping against Ray Data Co

The concrete connection: this is the narrative version of the exact mechanism already mapped in 2026-10-20-practical-data-modeling-fan-out-dashboard-double-counting — Reis explicitly built that standalone piece out of these same two chapters (8 and 9), and this book-club post is the slower, discussion-oriented walkthrough of the same ticket/product/message example. What it adds that the fan-out note didn't flag: the resolution-rate averaging trap is a distinct failure mode from row-count inflation, and it's one RDCO's own reporting surfaces are exposed to in a way the fan-out note's "fan-out skills" framing didn't cover. Any place a skill computes a per-batch rate (an audit pass rate, a critic pass rate, a weekly success percentage) and then rolls multiple batches up by averaging those per-batch percentages — rather than carrying forward the raw numerator/denominator and re-dividing — produces a mathematically wrong aggregate, silently, with no error thrown, which is precisely the failure class feedback_workflow_agent_output_integrity already tracks. The vault-note-schema changelog itself models the correct discipline without naming it: the I12 audit-failure tracking there is written as "Jun 17/73 fail → Jul 0/92 → Aug 2/60" — counts carried forward, never averaged rates — which is the right pattern to defend explicitly as new skills add their own "pass rate" reporting (/self-review, /vault-health, any critic that reports a weekly score) rather than assuming it'll be gotten right by convention.

Related