Missing measurement read as zero

What goes wrong

A value that was never measured is stored as 0 or dropped from the check, with no record that it is actually unknown. A report or alert then treats 'no data' as a measured zero or as healthy silence.

Example

The sales feed is down for a day, the loader writes 0 instead of null, and the dashboard reports 'no sales' while the low-revenue alert stays silent.

Not to be confused with

Not a total that mixes known and unknown parts (Partial total labelled as complete), not a real zero for an empty set, and not a permission check that allows everything when its policy cannot be read (Unreadable policy allows everything).

Discovered by

@sks24git GitHub

Field notes · 23

@sks24git2026-11-06beat 4 review rounds

A report re-ran a costly scan whose detectors were not built yet, so it found nothing and only added load. Its few hits skewed the false-hold rate, and its failures filled the log.

Drop a step that cannot find anything yet. Its cost and failures are real; its signal is zero.
Date a result by when it read the input, not when it was saved. A late save hides a stale read.

@sks24git2026-11-05beat 2 review rounds

A status summary took the max over a list of years and fell back to the all-clear value when the list was empty. When every input was unreadable, the reader got the one status that allows a reassuring message.

Make the empty case the unsafe answer, never the safe one. Nothing known is not all clear.

@sks24git2026-11-04beat 1 review round

A status page reads "nothing billed" from the charge record, but a background job creates that record later. In that window the page states the wrong amount for a result that will be charged.

Derive money text from what will be charged, not from what has been. An async charge leaves a gap.
Test the window before settlement. Final-state tests miss it.

@sks24git2026-11-04beat 1 review round

A page fell back to showing a child's own settings when the parent rule could not be read, and said nothing. One read failure, and the page again showed settings as active while the server turned them off.

Say when a fallback hides a rule. A silent fallback repeats the bug it was built to remove.

@sks24git2026-11-02

A "no changes made" flag was written for every finished job, including ones that crashed or failed to publish. Later retries read the flag as a clean pass and stop for good.

Write the "clean" flag only on a proven clean exit. A default of false turns every failure into a clean pass.

@sks24git2026-10-26beat 2 review rounds

A view merges the history pages that loaded and summarises the union, while its loading and failed flags sit unused beside it. One failed page and the screen states 'nothing happened here' as fact.

A flag nobody reads is not a state. Expose loading and failed where the text is rendered, or drop them.
Test the failed fetch, not only the slow one. Retry-off means one error is final.

@sks24git2026-10-02beat 3 review rounds

An error banner had a close button, and its message also fed a panel's "this source was not read" line. Close the banner and the panel reports all clear about data it never loaded.

Derive "not checked" from the read, not from its message. A dismissible error is view state, not data state.
Let only a successful retry clear a failure. A close button hides it.

@sks24git2026-09-14

An operator switch makes one screen render 'not computed' when a query times out. A new path reading the same query never consults it and renders a bare zero for the same outage.

A new reader of shared state must honour every switch the old reader honoured. Otherwise one outage gets two answers and the operator trusts the wrong one.

@sks24git2026-09-14beat 1 review round

A path stopped failing loudly and began degrading, but unlike its sibling on the identical event it records no counter. One log line is now the entire trace of a silent wrong figure.

When you remove an error, add the signal that replaces it. A degrade with no counter is an outage nobody can measure afterwards.

@sks24git2026-09-12beat 2 review rounds

An empty progress record was read as a fact about the worker. The same emptiness is left by a process that died, by one whose protocol records nothing, and by a reporting channel that broke.

Absence of measurement is not a measurement. Say nothing rather than guess which kind of nothing you have.

@sks24git2026-09-12beat 1 review round

A catch was typed on one library's exception hierarchy while the layer below rethrew another's. The test pinning the degradation threw the type the catch expects, so it stayed green while production kept returning 500.

Throw the type the producer actually throws. A catch and its test can agree with each other and with nothing else.
Check what a cited test really drives before citing it as proof.

@sks24git2026-09-12beat 3 review rounds

A report printed "the snapshot cannot tell these two cases apart" for an empty list. The field that tells them apart was in the same payload, parsed by the same adapter, and printed by the same file three sections below.

Before calling data unavailable, read the payload you already hold. "We need a new API version" is a claim about your own bytes.
A deferral ticket inherits its premise. Nobody re-derives one.

@sks24git2026-09-10

A new counter column defaulted to 0 with no backfill, and 0 meant 'nobody read this'. On first boot every finished review in the history lit up as unread and its next reader was told to stop a heartbeat.

A default on a live table is a claim about the past. Backfill from evidence, or make the column nullable and read null as 'not tracked'.

@sks24git2026-09-10beat 2 review rounds

The deadline was synthesized from an anti-retroactive floor whenever both anchors were missing, which is exactly the scope types whose activity is never read. Such a scope expired and could never revive.

A default computed from a floor is a decision, not a fallback. Ask which inputs are absent by construction, not by accident.
A state nothing can leave is worse than a refusal.

@sks24git2026-09-10

A replay-safety test built its event with a hardcoded scope, so all four cases walked the same branch. The other branch of that replay path, the one touching shared balances, was executed by no test at all.

Check which branch the fixture reaches, not how many cases you wrote. Four cases down one arm is one case.
Vary the field that selects the branch. That field is the test.

@sks24git2026-09-10beat 1 review round

An export gained an unpaid column keyed to rows the same function drops during dedup. Every debit with a real shortfall exported an empty cell, so the finance report read as fully paid.

Key a new column to the row that survives the pipeline, not the one you computed it from. Check the dedup step before the join.

@sks24git2026-09-10

A drift check paged the oldest groups first and stopped at the page limit. The oldest rows are the settled ones, so every run re-inspected the same clean history and never once looked at new drift.

Scan a monitor newest-first. A fixed page over ascending age inspects the same settled rows forever.
A check that has only ever returned zero must prove it can return one.

@sks24git2026-09-10beat 1 review round

A rollback recipe for a table drop named the migration that changed a foreign key but not its second statement, which relaxed the column to nullable. Replay it and the column returns NOT NULL, so deleting a parent fails.

Cite statements, not migration numbers, in a rollback recipe. A version name hides how many things it changed.
Re-read the migration before you summarise it. Memory keeps the headline.

@sks24git2026-09-10beat 3 review rounds

Every timestamp on the screen rendered in the user's chosen display zone, but the date picker parsed its input in the browser zone. Pick another zone and the window you schedule starts hours from the one you read.

One zone per screen, for input and output alike. A picker reading a zone the labels do not use ships a shifted window.
Test the form under a non-browser zone.

@sks24git2026-09-08beat 3 review rounds

A new flag removed a two-meaning null in the renderer, but the check two lines away still read the old value and treated null as "nothing was there before". An action that never happened was reported as done.

Add a state to kill an ambiguity, then grep every reader of the old value. The new state is worthless where it is not read.
Mutation-check every new branch, not the interesting ones.