Golden Data

Issue #15 · July 24, 2026

The Metric Definition Is Becoming a File. Most Portfolios Are Still a Dashboard.

Paul BrownWrites Golden Data from Ogden, Utah5 min read

Every Friday I read job descriptions in the industry I’m in. This week the document worth reading wasn’t a JD. It was a hiring guide telling managers to stop asking the question most candidates spent the spring practicing.

KORE1’s 2026 analytics engineer guide is blunt about trivia screens — asking a candidate to define a slowly changing dimension from memory establishes only that they studied. The screen it recommends instead is a walkthrough of a dbt project the candidate actually built: whether tests exist, when they were added, who reviewed the code.

That is a hiring-practice change, and it has its own causes. But it points at the same object two standards bodies moved toward this month. On July 28 the Model Context Protocol publishes the final version of its largest revision since launch. Underneath it, the Open Semantic Interchange is turning metric definitions into portable YAML — and in June handed itself to the Apache Software Foundation.

Line those three up and the shape is legible. The durable unit of analytics work is becoming a file. A file has an author, a review history, and a test. A dashboard has none of those.

SQL still appears in more than 80% of data analyst postings, and the recommended loop still contains a live modeling session. SQL is the gate. It stopped being the round that decides.

What’s actually moving in the market

MCP publishes its final 2026-07-28 specification on July 28. The release candidate locked on May 21, giving implementers a ten-week validation window; the current finalized version is still 2025-11-25, so this is the change landing next week rather than one already in production. The protocol goes stateless, which matters here for exactly one reason: remote MCP servers stop needing sticky sessions and start scaling on ordinary infrastructure. The plumbing is somebody else’s job. The consequence is yours. The spec also adds a lifecycle policy guaranteeing at least twelve months between deprecation and removal. Read that as the tell. Twelve-month deprecation guarantees are what maintainers publish when customers are about to put production systems on top. Agent access to warehouse data stops being a pilot.

OSI moved into the Apache Incubator in June 2026, as Apache Ossie. OSI is an Apache 2.0, YAML-based spec for encoding semantic-layer constructs — datasets, metrics, dimensions, relationships, contexts — announced at Coalesce in September 2025 by Snowflake, Salesforce, dbt Labs and others, with the specification published on GitHub in January 2026 and Databricks joining at release. The roster has grown from seventeen launch organizations past fifty, including direct competitors of the founders. That is the usual tell that a standard is real. The move to a neutral foundation is a stronger one: vendor-led specs that stay with their vendors tend to become press releases.

The generalist skill list hasn’t moved. The experience band has. SQL is in over 80% of data analyst postings, Excel or Sheets in over 60%, Python or R in roughly half. The sharpest demand increase is for candidates with four to six years of experience — the band where someone has shipped enough to have opinions and not enough to have delegated them. U.S. senior data analyst base sits at $92K–$135K, with analytics engineers in tech metros adding 15–30%. The tools got cheap. The premium moved to whoever decides what the tool is pointed at.

What I’d do this week

Move one metric definition out of a dashboard and into a file, and write down why it needed to move.

One action, two artifacts, and they are the same argument from opposite ends. The file is the fix. The write-up is the evidence you knew it was broken.

The user moment. The next time two dashboards disagree in a meeting and someone offers to look into the discrepancy. That’s the incident. Capture it while it’s live, before it resolves into a story about how you fixed something.

The shape. One committed YAML file. One page of prose.

The time budget. Sixty minutes for the file, forty-five for the page.

The file. Take your most-cited metric. Write it as a dbt semantic model or an OSI-shaped YAML file: the grain, the filters, the exclusions everyone forgets, and one test that fails if it drifts. The exclusions are the valuable part. Those are the tacit rules that currently live in your head and in nobody’s documentation, and they are the reason the three dashboards disagreed.

The page. Name the metric, the definitions that competed, which one was correct and why, who made the call, and what enforces it today. If nothing enforces it today, write that. An honest gap reads as governance instinct. A vague success story reads as a story.

What success looks like. A month of Fridays gives you four of these. Four written incidents plus four committed definitions is a governance record, and it is the exact evidence the dbt project tour is built to surface. You cannot cram for that screen the week before, which is the entire reason it replaced the one you could. Most candidates arrive with adjectives.

The interchange tooling that consumes these files is still mostly roadmap — native support lands across the back half of 2026. That is the reason to write the file now rather than after. When the agent conversation reaches your org, you will already have a definition that is reviewable, testable, and attributable to you. The person who owns the definition is not competing with the tool that writes the query.

Sources

Crafting runs weekly. The job is to be useful, not to sell. If a line in here is close to a move you're working on, the email reaches a person.

Hiring rather than hunting? Signal is the column for your side of the desk.

Email meRead more Crafting

← All Crafting issues · Drafted with Claude · Edited by Paul Brown