06-reference

practical data modeling turf wars cross train

2026-06-02·reference·source: Practical Data Modeling·by Joe Reis

"The Turf Wars Are Over. Time to Cross-Train" — Joe Reis

Why this is in the vault

This is the capstone manifesto for Joe Reis's "Mixed Model Arts" (MMA) framework — the 16-chapter book the vault has tracked all spring (see the mma-ch1 through ch16 series). Where the chapters built the toolkit piece by piece, this essay states the thesis in one breath: the methodology holy wars (Kimball vs Inmon, Data Vault, One Big Table, knowledge graphs) are obsolete because the consumer changed. Machines are now first-class consumers of data, and no single 1990s-era modeling camp owns the answers for that game. It belongs in the vault as the definitive statement of a framework Ray Data Co already treats as a tracked author's flagship idea, and because its "know which game you're in, cross-train the rest" posture is a direct analog to RDCO's own solo-operator-wears-all-hats reality.

The core argument

Reis uses a mixed-martial-arts analogy as the spine. A pure boxer dominates a boxing ring, but loses in a cage where takedowns and kicks are legal — not because boxing got worse, but because a different game showed up with different rules. Data modeling has run its own thirty-year style war (relational vs everything, Kimball vs Inmon, Data Vault infighting, One Big Table, the "knowledge"/graph crowd vs everyone on AI). His read: that noise signals fear, not confidence.

The pivot is that every camp was arguing on the wrong axis. The classic methodologies all answered one question — how to organize information so humans can find, trust, and reason about it — and they're still good answers to that question. But the world now includes machines as first-class consumers: agents query the data, models get grounded in it, pipelines get authored and repaired by systems that don't read the wiki or care which methodology "won." That adds new questions next to the old ones: is this data legible to a machine, trustworthy enough to act on, reasonable-over without a human holding its hand.

His answer is "Mixed Model Arts" — explicitly not a new camp built to beat the others (that would be the exact tribalism he's criticizing). It's the practice of a well-rounded mixed martial artist: keep a base specialty you lead with, then cross-train — dimensional where it wins, normalization where it wins, Vault where change is constant, graph/semantic where meaning carries the load, ML/AI for unstructured data, and increasingly designs that serve machine consumers alongside human ones. Borrow relentlessly, stay loyal to none of it, honor the lineage, and build on the shared foundation (grain, entities, relationships, attributes, time, semantics) that every tradition was really wrestling with.

He closes by redirecting energy to the genuinely open problems: modeling so an agent can reason without hallucinating, encoding enough semantics that a machine grasps meaning not just shape, building trust and lineage where the consumer can't phone a human to ask what a column means. "We are now designing for humans and machines. That's the whole headline."

Mapping against Ray Data Co

Medium-to-strong. Three live RDCO threads connect:

  1. Cross-training is RDCO's operating reality, not just an aspiration. A solo founder plus an always-on COO agent literally cannot afford a single-camp specialty. The "keep your base, round out your game for the problem in front of you" posture is exactly how RDCO has to staff data work — and it validates the audit-model / generate-tests skill stance of being pattern-agnostic rather than dogmatically dimensional or Vault.

  2. "Machines as first-class consumers" is the FDE / agent-native thesis stated from the data-modeling side. Reis is independently arriving at the same conclusion that drives RDCO's COO-agent build: when a probabilistic system is your most demanding consumer, the work shifts to legibility, trust, lineage, and machine-readable semantics. This is useful corroboration for any RDCO positioning that says "the data layer has to be built for agents now," and a potential reframe seed (note: per the no-derivative-Sanity-Check rule, this would need an original angle, not a restatement of Joe's piece).

  3. The anti-tribalism caution applies to RDCO's own tooling opinions. RDCO should resist planting its own methodology flag; the value is in knowing which game a client/problem is in and bringing the right blend — which is also a sharper way to pitch FDE work than "we do dbt/dimensional modeling."

Honest caveat on strength: this is a manifesto, not a how-to. It's directionally aligned and quotable for positioning, but it gives RDCO zero new mechanics — the operational "how do you model for an agent" payload is deferred to "hard, open problems." So it's strong as positioning corroboration, medium as actionable input.

Bias / sponsor notes

No sponsor block; not sponsored. The only self-promotion is the standard Substack reader-supported subscribe line. Worth flagging the reverse signal: Reis explicitly calls out a competitor in the "knowledge"/graph-courses space, reading their data-warehouse-bashing as "more like a sales pitch for the person's knowledge graph courses than something a sensible practitioner would implement." So the bias to watch here is Reis defending the broad modeling tradition (and, implicitly, his own MMA framing/book) against a graph-first sales narrative — he's inside the tent he's describing, which he acknowledges.

Related