The change history the system does not keep

Delete a payroll input and it is not marked deleted; it is gone. Users report that removed inputs disappear entirely from the Payroll Inputs data source, leaving no sign they existed. That is the sharpest case of a general property: Workday has no subscribable change feed for most objects, so nothing downstream is told when a record changes or dies.

The integration vendors confirm it in writing: a major connector’s public documentation states that deleted rows are never captured during incremental ingestion. Rescinds and corrections go further, rewriting effective-dated history in place, so the timeline you read today is not the timeline that produced last quarter’s numbers.

The recurring forum questions follow directly: who deleted this, what changed in this business process definition, which delegations were active last year. The delivered answer to each is silence.

The fix does not need the source to cooperate. Snapshot each object daily into your warehouse, then diff consecutive days: a row present yesterday and missing today is a delete, new today is an insert, present in both with different values is an update. The diffs accumulate into a change ledger, one row per detected change, retained forever. The headcount history snapshot is the same foundation; the ledger falls out of it.

-- deletions between two consecutive snapshots
SELECT y.worker_id, y.input_code
FROM snap_payroll_input y
LEFT JOIN snap_payroll_input t
  ON  t.worker_id = y.worker_id
  AND t.input_code = y.input_code
  AND t.snapshot_date = DATE '2026-08-06'
WHERE y.snapshot_date = DATE '2026-08-05'
  AND t.worker_id IS NULL

The pipeline that feeds it is the ordinary one: extraction under a locked-down integration account, snapshots landed on schedule, and the payroll extraction for the inputs Workday itself forgets first.

Auditors ask for exactly what the ledger stores: what changed, when, and the values before and after. A SOX control that depends on knowing about deletions cannot rest on a data source that forgets them. Three days of one diffed object look like this.

Sample values are illustrative, never client data. Each day’s changes are the sum of its inserts, updates, and deletes, and the three-day totals crossfoot exactly.