The req is approved. You post it, and within a week you have forty applications and none of them are right.
That's the expected outcome, and it has a specific cause. The title barely exists on resumes yet, so keyword matching returns people who have written the words metadata and engineer in the same document. Almost none of them have run the practice.
Everything below assumes you already have the job description. This is the part that comes after.
Look inside first
The strongest candidate is usually already on the payroll, and they're easy to identify.
Ask your analytics and data engineering leads one question: when two reports disagree about a number, who gets pinged? The name that comes back is a person already doing three of the five stages without a mandate, without a title, and without it appearing in their review.
They will have the estate knowledge, which takes eighteen months to rebuild in an external hire. What they usually lack is the operating discipline: environments for definition changes, promotion, rollback, verification. That gap is teachable in a quarter. Estate knowledge takes far longer to build.
Promote them and backfill their old role. That's a faster path than a six-month external search, and it shows the rest of the team that the career path is real.
If you hire externally, search for the adjacent work
The title is unsearchable, so search for what the work leaves behind.
BI migrations. Anyone who has moved an organization between BI platforms has confronted metric definitions that exist in three places and agree in none. That experience produces the right instincts.
Internally built lineage or metadata tooling. People who have written a parser, built a dependency graph, or maintained a homegrown catalog understand why this is hard. They also understand why buying is usually correct, which makes them easier to work with.
Data engineers who talk about definitions. Scan for semantic models, metric layers, and metric governance in their own words. The vocabulary signals whether someone who builds pipelines has started thinking about meaning as well.
Anyone with regulated reporting exposure. SOX, BCBS 239, and their equivalents force people to learn evidence and provenance the hard way.
Referrals outperform search here, because the reconciliation person at another company is known to your reconciliation person.
What to test, in four exercises
Skip the take-home. The role is judgment under ambiguity and a take-home tests neither.
One: the trace. Give them a sanitized diagram of your estate and one metric. Ask how they would determine which column produces the number and whether it's the right column.
Listen for the questions they ask. Strong candidates ask what transformation logic sits between the hops, which systems have harvested metadata and which are dark, and who owns the definition. Weak candidates open the catalog and read what it says.
Two: the conflict. Finance defines active customer one way, product defines it another, and both definitions are live in production reports. What do you do?
The failure is picking a winner. Strong candidates ask who consumes each one and why, propose that both survive under distinct names with named owners, and raise deprecation as a separate exercise with a timeline. Anyone who treats this as a data quality problem has the wrong model of the job.
Three: the change. A source schema change is proposed for next sprint. Walk me through what happens before it ships.
You're testing whether they think in blast radius. Strong answers cover impact analysis ahead of the change, which consumers are affected at what granularity, a lower environment, and a way back. Weak answers describe finding out when something breaks.
Four: the stakeholder. Explain to a CFO, in plain language, why two reports show different revenue figures and what you would do about it.
This is the scarce half of the role and the one most technical loops leave out. The answer should contain no architecture, no tool names, and a clear path to resolution. Plenty of strong engineers cannot do this, and in this role it's load-bearing, because the person holding it spends real time in front of finance, risk, and audit.
Signals that mislead
Vendor certifications on a catalog product. These test tool operation. The job is judgment about meaning, and the tool changes every few years.
Governance framework fluency with no code. Someone who can discuss stewardship models at length and has never written a transformation will run the practice as a documentation exercise. That's the specific failure the role exists to prevent.
A deep data quality background. Adjacent discipline, different job. Metadata quality describes the estate; data quality describes the values inside it. Candidates who conflate the two will spend their first quarter building the wrong thing.
Long tenure in a single estate. Sometimes it means deep expertise. Sometimes it means they know one estate very well and have never had to generalize. Probe which.
Leveling and compensation
The market has no comp data for a title this new, which is a practical problem and a temporary one.
Anchor to your senior analytics engineer or staff data engineer band, at the top of it. The role requires SQL, transformation code, at least one BI semantic model, and enough governance vocabulary to hold a conversation with risk. That combination is scarcer than any of its parts, and the premium belongs there. Creating a new band would mean defending it to finance for very little gain.
Use the three-level skills matrix from the job description to place them, and level on work products. Someone who can state coverage as a number with a denominator is operating a level above someone who reports that coverage improved.
Set the first quarter before they start
Decide two things before the offer goes out.
Which three business metrics they will own end to end in the first ninety days, and who supplies the definitional requirements they will work from. The second one is the interface to governance, and leaving it undefined is the most common way a strong hire stalls.
Then give them one number to move: the metadata completeness score across those three metrics, with the denominator stated. A new function that reports a measurable delta in its first quarter gets a second headcount. One that reports activity does not.
