Shared data narrowed for one consumer

What goes wrong

A shared value, label, or default is adjusted to suit one consumer. A second consumer still reads the old form and quietly gets nothing back, or gets a substitute it never asked for. That second reader wrongly relies on the shape and range of shared data staying fixed, because its own code was never touched.

Example

A cleanup step relabels a class of rows so one report stops counting them, and a second report that selects on the old label now finds none and reports nothing left to do.

Not to be confused with

Not a shared helper changed for one caller: there the code every caller runs is edited, while here each consumer runs exactly the code it always ran and only the data underneath it has moved.

Discovered by

@sks24git GitHub

Field notes · 10

@sks24git2026-11-06beat 1 review round

A page script read the live user choice from a button that a new switch could now leave out of the page. With the button gone, links quietly fell back to the old choice from the address bar.

Grep every reader of an element before you hide it. Other scripts may use it as a state store.
Keep shared state off display parts. Put it where no switch can remove it.

@sks24git2026-11-04beat 3 review rounds

A status page dropped a flag together with the progress bar it once fed, while two guards still leaned on it. When one parallel run ends early, the page says setup is starting while another run is mid-review.

List every reader of a field before you delete it. A field added for one widget often guards others.
Keep the test when you drop the feature. The state it pinned still happens.

@sks24git2026-10-30beat 1 review round

We packed many log entries into one message to save writes, and gave each entry a made-up id. Links that point at the message itself then matched no entry, so finished work looked open and got assigned again.

Define what a link to the container means. Packing items changes how every reader resolves them.
Cap the whole message, not each part. One long part can push the rest out.

@sks24git2026-10-30

New code moved to a new version record and stopped updating the old one. Older builds still read only the old record, so a rollback finds its own version there and skips the schema it must restore.

Keep writing every record an older build still reads. A rollback runs the old reader, not your new one.

@sks24git2026-10-28beat 2 review rounds

A fast path accepted an external verdict and skipped writing the record a downstream collector keys on. The item was marked ready, the collector never found its record, and it sat until someone pushed again.

When you add a shortcut, write every record the slow path wrote. Consumers key on records, not on labels.
Test the shortcut in every mode the pipeline has. Batch mode is where shortcuts stall.

@sks24git2026-10-19beat 1 review round

A feature removal deleted the only two code paths that granted two badges, but left the badges listed as earnable. Users keep seeing rewards and steps that point to a page that no longer exists.

When you delete a grantor, retire what it grants. Otherwise the catalog keeps promising rewards nothing can pay out.

@sks24git2026-10-16

A status time is the newest time among the rows a step writes, and a change stops writing rows for items already in progress. When every item is busy no rows remain, and the time reads as missing until one frees up.

Store the step time on its own. A time derived from rows vanishes when the rows do.
Check every reader before you narrow a writer. Fewer rows can break a field built from them.

@sks24git2026-09-12

A sweep skipped running rows because another worker would re-claim them, and one release narrowed that worker to the types it knows. A row of a retired type is now claimed by nobody and swept by nobody.

Grep the reasons your change falsifies, not only its callers. A branch stays right while its reason rots, and the next edit builds on the rotten reason.

@sks24git2026-09-10beat 1 review round

A check read the new zero-value anchor row through the join written for the old shape, computing an outflow of the whole balance against an expected zero. Every deleted account became a permanent phantom imbalance.

A new record shape must be excluded from the query written for the old one. Restrict the legacy join; do not widen the tolerance.
Check both halves of a two-sided check.

@sks24git2026-09-07beat 2 review rounds

A fix relabeled experiment runs so they would not pollute the champion's score. Every challenger now carried that label, so the selector that needed their evidence saw none and could never move again.

Every experiment needs an off switch before it ships. A deploy is not a knob.
Isolating data is also deleting it for someone. Ask who reads the rows you relabel.