Warehouse Operations

A warehouse supervisor asks four questions in a shift, and they are four different measurements. What has to move before the dispatch cutoff, and what is already late? How much of what arrived at the dock is actually available to pick? What shipped, as opposed to what merely got picked? And when a count disagrees with the book, which number is the count and which is the book?

The pattern keeps them apart deliberately. Receiving is not putaway availability. Picking is not shipping. A physical count is not an inventory position. Each is its own fact at its own grain, and each measurement names the accountable role that owns it.

Two instants are pinned separately: the business as-of decides what has happened, and the capture cutoff decides what is knowable. A report keyed wrongly at dispatch and corrected the next afternoon gives two honest readings — the one an earlier reader legitimately saw, and the corrected one — and the pattern can show both rather than pretending the first never existed.

1. 1Pins — the business as-of and the capture cutoff are separate controls, and the history window is separate from both.
2. 2Measure row — counts, per-unit quantities and a pooled age stay four numbers; none of them is averaged into a score.
3. 3State breakdown — overdue, at risk, due before cutoff and due after cutoff are one state per task, so the states partition the queue.
4. 4Worklist — each row keeps its own due instant, its revision count and the role accountable for finishing it.
5. 5Evidence — partials, cancellations, short picks, corrections and recounts stay visible rather than being netted away.

The accumulating fact is the task: it gathers a versioned commitment and an append-only stream of activity events, and it is never edited in place. A due instant that moves is a new revision; a cancellation is a revision; a mis-keyed completion is a correcting event that supersedes the report it replaces. Both a business instant and a capture instant are kept on every row, which is what makes a reading reproducible at either pin.

Around it sit the records the tasks work against: receipt lines and the putaway movements that make their quantity available, allocations and picks and shipments on the outbound line, and cycle-count plans whose lines carry the book quantity read at the plan’s own boundary snapshot with every count attempt kept beside them.

Every one of those base facts carries its own capture instant, not just the events reported against them. A receipt, an order line, a task or a count plan that reaches the system late is absent from a reading pinned before it arrived, and the reading says how many facts it could not see — a smaller population, disclosed, rather than a quantity of zero.

Current state and historical performance are different populations and are controlled separately. The queue is read at the pinned instant; completion, duration, receiving and outbound populations are read over a trailing window and always publish how much eligible work is still open inside it.

A duration here is elapsed wall-clock time between two business instants. No labour time is captured in this fact model, so none of these measures is labour productivity and none may be labelled as such.

Outbound line measurement is warehouse-wide. These facts carry no allocation of an ordered quantity to a zone — a line is typically picked in one zone and shipped from another — so a zone filter over order lines would report a denominator nothing supports. The queue is at task grain, where the zone is a fact, and remains scoped by it.

• Pin the business as-of and the capture cutoff separately, and compare timestamps as instants — two rows written at different UTC offsets are not ordered by their strings.
• Reject ambiguous commitment revisions (two revisions effective at the same instant on one task) and correction chains that fork, loop or point at another task.
• Declare one terminal-state profile and let every view read it from one kernel. Completion is a physical fact: a cancellation recorded afterwards cannot un-report it, and a cancellation recorded before it does not make the work disappear either — it makes the work something that happened outside a live commitment, which is counted on its own rather than scored or dropped. A reinstatement is a later active revision in the data, never an assumption in the code. Two views walking the same ladder in different orders is how a task ends up completed on one screen and cancelled on the next.
• Keep receiving and putaway as separate records, and never let the sum of a line’s putaways exceed what it received. A partial stays partial.
• Keep allocation, pick and shipment as separate quantities. A split shipment is several rows; a correction supersedes a row and both rows are retained.
• Hold the cumulative quantity to the obligation, not just each row to its own allocation. Ten picks of ten units each are individually consistent and collectively an invention: sum every effective pick and shipment on a line and compare the total with what the line committed, under a declared fulfilment profile. If that profile permits a surplus, publish the surplus as its own measure and keep it out of the completion numerator.
• Scope only what an allocation supports. An ordered quantity is a commitment on the warehouse; unless the model carries an allocation of it to a zone, outbound line measurement is warehouse-wide and the filter is not offered rather than being offered and quietly ignored.
• Give every base fact a capture instant, not only the events reported against it, and apply the knowledge cutoff to all of them. A late-arriving receipt, order line, task or count plan is absent from an earlier reading and its absence is published.
• Pair a physical count with the book snapshot named by its own plan boundary, declare which attempt a recounted line is scored on, and never average attempts. Hold that declared sequence to the chronology: a recount counted before — or at the same instant as — the attempt it recounts is a data fault, not a scorable result.
• Keep signed variance, absolute variance, within-tolerance accuracy, sample size and coverage as five separate numbers. No counts is unavailable, not 100%, and a plan is a sample rather than a census.
• Treat the dispatch cutoff, the at-risk horizon and the count tolerances as versioned configuration with an owner and a stated provenance, and score a plan under the version it pinned.
• Partition every quantity by unit of measure, and reject a row whose unit disagrees with the record it belongs to rather than converting it silently.