A Data Engineer’s Checklist for Auditing the People Record

If your team already audits pipelines for freshness and schema drift, there is one dataset in the company that rarely gets the same attention: the people record. Who is employed, in which role, on which dates and under whose management feeds payroll, access control, headcount reports and compliance. This checklist walks a data engineer through auditing it, using ideas from PeopleOps, the habit of treating hiring, moves, pay and exits as work that gets timed, owned and counted.

Before you start

Grab a read-only extract of the HR system and one downstream consumer, such as the identity directory or the payroll feed. You are not here to fix anything yet. You are here to answer a handful of yes-or-no questions and write the answers down. The framing comes from the Ops Atlas page on PeopleOps, which treats the HR record as a ledger of effective-dated facts and the lifecycle as a set of owned transitions.

Section A: identity of the record

  • Is there a permanent worker ID? The Atlas says never to use email for this purpose, because addresses change at marriage and when someone is rehired. Check that the ID survives those events.
  • Do all joins use the ID? Search your SQL and your transformation models for joins on email or display name. Each one is a latent duplicate or a silent drop.
  • What happens on rehire? The Atlas wants the identity reactivated with no prior access carried over. Confirm that your pipeline does not resurrect old entitlements.

Section B: time

  • Are records effective-dated? You should be able to ask what a person’s role was last March without guessing.
  • Can future-dated changes exist? The Atlas recommends entering a termination ahead of time with its effective moment, so downstream systems can plan. See whether your models treat a future date as already true.
  • Are time zones explicit? A departure at the end of the working day in one region can land on the next date in another.

Section C: ownership of attributes

For every field in the record, write down who is allowed to change it and which system is the origin. The Atlas calls this an attribute-ownership matrix. Typical gaps are a manager field edited by both HR and the directory, or a job title overwritten nightly by a sync from a different tool. In a short sheet, mark each field as owned, shared or unknown, and treat every “unknown” as a defect.

You may find it useful to compare this exercise with a data-quality tooling walkthrough you already trust. A hands-on article about validation suites for pipelines describes expectations written as tests, and the same idea applies here: assert that every active worker has a manager, a start date and exactly one primary role.

Section D: quality sampling

  • Do you sample on a schedule? The Atlas lists quarterly data-quality samples in its governance rhythm. Draw a random set of records and compare them with source documents.
  • Is there a written standard? The Atlas mentions an international standard on workforce data quality, ISO 30435:2023, as a place to look for vocabulary.
  • Who reviews the findings? A sample nobody reads is theatre.

Section E: reconciliation between systems

This is where most audits find something. Compare three lists on the same day: people marked as terminated in the HR record, people whose accounts are disabled in the directory, and people paid in the latest payroll run. The Atlas wants a daily reconciliation, run as a check that raises alerts and not as a monthly report. A person who is terminated in HR but active in identity must page someone.

  1. Export all separations for the last thirty days.
  2. Export all disabled accounts for the same window.
  3. Compute the difference in both directions.
  4. For each mismatch, record how long it persisted.

Section F: how data moves

  • Is the transport fragile? Nightly flat files over SFTP are common and brittle. A silent failure means yesterday’s leavers remain active.
  • Is it scheduled pull or event push? Pull intervals add delay to every termination. Measure the actual lag.
  • Is provisioning standard? The Atlas names SCIM 2.0 (the 2015 pair of RFCs numbered 7643 and 7644) as the usual wire format for provisioning, and it includes a patch operation suited to role changes. Find out whether your connector speaks it or relies on a hand-rolled script.

Section G: metrics you can compute today

Two quick ones make good baselines. The separation rate is separations divided by average headcount, and in the Atlas worked example five leavers against 119.5 average heads gives 4.18 percent for the quarter. The second is revocation latency: what share of leavers lost access inside the written window? Its sample reconciliation shows two of four within the one-hour rule, which is fifty percent, with one orphan account and one breach of twenty-six hours. Small as those numbers are, they turn an argument into a measurement.

For context on how analysts talk about transformation layers, a primer on building models with a popular transformation tool is a good companion, since many of these checks fit naturally as model tests.

Section H: privacy floor

  • What is the smallest group any report can show? The Atlas gives five as an example minimum reporting group.
  • Is complementary suppression applied? If one team is hidden, a neighbouring one must be too, otherwise subtraction reveals it.
  • Is retention enforced? Deletions belong on the annual calendar, not in a backlog.

Section I: the awkward edge cases

  • No-shows: does a candidate who never arrives still end up with provisioned accounts?
  • Leave of absence: if a separation lands while someone is on leave, which system learns first?
  • Movers: when a person changes team, is there a prompt to re-review access, as the Atlas suggests citing a NIST personnel-security control?

Write a test row for each of these, even a synthetic one, and run it through the pipeline end to end. Edge cases are where a tidy-looking dataset usually breaks, and they are cheap to rehearse.

Scoring your audit

Tally every line where your honest answer was a clear yes. Anything below roughly two thirds of the list means start with sections A and E, because stable identifiers and daily reconciliation unlock most of the rest. Write the results in a one-page note, name an owner for each gap and schedule a re-run in a quarter. This is informational guidance for data teams, not advice on employment law.

Related Posts

Priority-First vs Inbox-First: Two Ways to Run a Data Day

A friendly comparison of two ways to start a data engineering day, reacting to the inbox or naming one must-ship outcome first, with a verdict on when each style fits.

Read More

Data Pipeline Failed Again? A Blameless Review FAQ for Data Teams

Questions data engineers ask after a broken pipeline: how to run a blameless review, what to write down, and how to finish with one concrete fix.

Read More

The Weekly DataOps Review: A Beginner’s Primer and Glossary

New to data engineering? This primer explains the weekly DataOps review, defines the key terms and walks through a four-step routine you can start this week.

Read More

Memory, Automation or a Card? Guarding Risky Data Changes

Three ways to protect a backfill or table rewrite are compared head to head, with a seven-line pre-flight list that fills the gaps tests leave behind.

Read More

Pipeline Fixes: One by One or in a Weekly Batch?

A human comparison of two DataOps habits, fixing each pipeline problem as it lands versus grouping similar fixes into scheduled review sessions.

Read More

Six DataOps Myths Corrected for Working Data Engineers

Six persistent DataOps myths, from the DevOps-with-SQL shortcut to the green-job fallacy, corrected against the Ops Atlas entry.

Read More