Assigning Permanent Numbers Without a Central Registry

Background

Permanent identifiers are assigned at the point a piece is finished, which is a bench, not a desk. Until this trial, every number was issued by one workstation that held the only authoritative sequence. That workstation being unavailable stopped numbering entirely, and numbering stopping stops finishing.

The dependency had been accepted for two years on the assumption that a single sequence was the only way to guarantee uniqueness.

Problem Statement

Determine whether permanent, non-reusable identifiers can be issued by multiple independent workstations with no shared connection at the moment of issue, while preserving three properties: uniqueness across all issuers, no reassignment of a retired number, and an ordering that is meaningful to a human reading a list.

Objectives

1. Zero collisions across the trial period.
2. No workstation blocked by the unavailability of another.
3. Identifiers remain human-readable and sortable by eye.
4. Reconciliation to a single record within twenty-four hours of issue.

Constraints

• No network connection could be assumed at the bench. Two of the three workstations are in areas with no reliable signal.
• Identifiers already issued under the old scheme could not change.
• Any scheme had to be usable by hand if the workstation failed entirely.
• The format could not grow longer than the existing one; it is printed in a fixed space.

Approach

Each workstation was allocated a disjoint block of the sequence in advance, with blocks sized from the highest observed monthly output plus one hundred per cent headroom. Blocks were issued quarterly and recorded on paper at the point of allocation as well as in the system.

The critical decision was to accept gaps as permanent. A workstation that does not exhaust its block leaves that range unused forever. This was judged cheaper than any scheme that reclaims unused ranges, because reclamation requires exactly the coordination the trial was trying to remove.

Implementation

Rolled out over eight weeks. Week one was a single workstation running the new scheme alongside the old one and comparing output. Weeks two to four added the second workstation. Weeks five to eight ran all three.

Paper allocation logs were kept at each bench from day one. This turned out to matter — see Results.

Metrics

Results

Both objectives one and two were met on the second attempt and not the first.

Two collisions occurred, both in week five, and both from the same cause: the third workstation was brought online using a block that had been recorded on paper but not entered into the system, and the system then issued the same range to the second workstation. The paper log is what caught it, four hours after the fact.

No identifier reached a finished piece with a duplicate. Both collisions were resolved by retiring the two duplicated numbers entirely rather than reassigning either — consistent with the constraint that numbers are never reused.

What Worked

The paper log. It was included as a fallback for total workstation failure and instead served as the reconciliation check that caught the only real failure of the trial. It cost nothing and was the single most valuable component.

Permanent gaps. Three hundred and eighty-six unused numbers were created and no operational problem resulted from any of them. The concern that gaps would look like errors did not materialise, because a gap in a list is only confusing if the list is presented as continuous, and it is not.

What Did Not Work

Sizing blocks from observed output plus headroom was too conservative for one workstation and far too generous for another. The workstation that produces least received a block it will not exhaust for several years, which is harmless but means its identifiers cluster in a range that reads as much older than they are.

Reconciliation within twenty-four hours was met on median but not on worst case: one batch took fifty-one hours because the workstation was unattended over a weekend.

Future Improvements

Block allocation should be entered into the system before it is written on paper, not after. Reversing this order removes the exact failure that produced both collisions.

Blocks should be sized per workstation from that workstation's own output rather than from the highest observed across all three.

Reconciliation should be triggered by the workstation on reconnection rather than by a scheduled sweep, which would have reduced the worst case from fifty-one hours to under one.