Total computed over the wrong records

What goes wrong

A sum, average or ratio is computed over a different set of records than its label promises — only the loaded page, a missing date filter, a denominator from another selection. The number is arithmetically right for the wrong population.

Example

'Total spent this month' is summed over the first 50 orders the API returned; the customer has 320.

Not to be confused with

About which records are included; not missing values inside the right set (Partial total labelled as complete), not a different metric under the same name, and not arithmetic errors after selecting correctly.

Discovered by

@sks24git GitHub

Field notes · 28

@sks24git2026-11-04

A summary card shows two counts side by side that use different filters: one counts every severity, the other only the serious ones. Readers compare them and draw the wrong ratio.

Figures shown together must share one filter. Mixed bases make a false comparison.

@sks24git2026-11-04beat 1 review round

A latency percentile counted calls that were cut off at the deadline, using the time they were stopped as their length. A slow run that hit its deadline then looked like one with fast, short calls.

Time only the calls that finished. A cut call's clock shows when you stopped it, not how long it takes.

@sks24git2026-11-04

A timing log stops its clock only when each tool ends, so file checks done between tools get billed to the next tool. The log meant to find the slow step blames the wrong one and sends the search astray.

Time every segment, and check that the parts add up to the total. A gap or overlap hides the real cost.
Log a diagnosis before the slow step it is about. A kill midway loses it otherwise.

@sks24git2026-11-04

A summary check skipped items marked as excluded, while the new page headline claimed every item passed. When an excluded item had failing checks on the same table, the headline contradicted the visible rows.

Match headline scope to its filter. If the count skips excluded rows, name only active rows.
Check new copy against banned words. Rewrites bring back slang unless checked against the glossary.

@sks24git2026-11-04beat 2 review rounds

A new dashboard count skipped the filters its sibling counts apply, so internal jobs and test accounts counted as customers. The one number meant to guide a decision leaned the wrong way.

Copy the scope of the sibling counts. A total without their filters counts traffic no customer made.
Grade a skewed metric by what it decides. A choice may rest on it.

@sks24git2026-11-03beat 3 review rounds

A fix dropped one billing mode from the top of a cost-per-bug ratio, but its confirmed bugs still counted at the bottom. The price per bug drops whenever someone judges a bug in that mode, and the fix's test seeds none.

Filter both sides of a ratio with one predicate. A rule applied to half a fraction is a new bug.
Test a fix with an input that reaches both sides. A seed that touches one passes either way.

@sks24git2026-11-03beat 2 review rounds

A cost-per-bug figure divided all billed money by bugs people confirmed, but one billing mode charges with no confirm step. Its money lands on zero bugs, so the price per bug soars or shows a dash beside a real bill.

Divide money only by the outcomes that money paid for. Mixed billing modes make a ratio lie.
Seed one row of every billing mode in a money test. The mode you skip is the one that breaks.

@sks24git2026-11-02

A sampled second-opinion check counts mismatches only on clear verdicts, yet divides by every compared item, unsure ones too. When the second judge often says unsure, the rate shrinks and the alarm stays quiet.

Divide by what can count. A rate whose numerator skips a class must skip it in the denominator too, or the alarm under-fires.

@sks24git2026-11-01

A retry cap counted every finished job sharing a parent name, not only jobs of the task it guarded. Unrelated successes filled the quota, so healthy work was stopped and sent to a human as stuck.

Tag each attempt with the task it serves. Counting by a shared prefix mixes in jobs of other kinds.
Check freshness before a stop rule. An outdated task should be retried, not escalated.

@sks24git2026-10-28

A cost report counted only work started inside its time window, then labelled the result as each item's own total. Long-lived items lost their early work, so the worst cases looked cheapest.

Fetch per-item history from the item, not from the window. A window cut silently drops the long tail you want to see.

@sks24git2026-10-25

A metric took the earliest row of a shared event table as the first review, but that table also held CI-failure rows. Any change with red CI looked reviewed minutes after opening, and the review-to-merge time stretched.

Filter by row kind before taking a min. A shared table mixes events with different meanings.

@sks24git2026-10-16beat 1 review round

A fallback count summed every row of a work queue, while the list view of the same column skipped work already in progress. One snapshot showed two numbers, and the count overstated the backlog.

Count from the same set the list shows. Two formulas for one column drift apart.
Check what a feed holds before you count it. Rows for in-flight work inflate totals.

@sks24git2026-10-13beat 1 review round

A query kept a row only if one of several amount columns was non-zero, but the total also summed a column the filter skipped. A row that moved only through that column vanished and the totals stopped adding up.

Filter on every column the total sums. A narrower filter drops rows without raising any error.

@sks24git2026-10-13

A value meant to be one per entity was computed inside a report's grouping, which split each entity into several slices. The same entity then showed different values side by side, and an update reached only some of them.

Compute per-entity values over the entity, not the report's grouping. Each slice otherwise gets its own truth.
Test one entity spread across two slices. One row per entity hides it.

@sks24git2026-10-09beat 2 review rounds

A summary counted a run as tested only from a copied log, but read its findings from a separate report. When the copy failed, a run showed findings yet no test, and the rates read as an impossible pair.

Derive linked counts from one source. Two sources drift apart the first time one of them goes missing.
Test the run with a missing input. That is where paired numbers split.

@sks24git2026-10-06beat 5 review rounds

A daily summary keeps one row per key×model even when every request that day failed, with a success count of zero. Read it without a "successes > 0" guard and the card lists models the key never used, at 0/0.

Filter on the reader what the writer counts but does not drop. A FILTER(WHERE ok) measure keeps the row alive at zero.
Diff predicates, not sums. A sum oracle cannot see an extra zero row.

@sks24git2026-09-18beat 12 review rounds

A CLI's list command returns each item's comments as a nested page of 100 and never turns the page. The longest threads lose their closing notes, and the report counts them as never written.

A full page is a ceiling, not a census. At exactly the page size, fetch again or refuse.
Check nested limits, not just the outer one. The cap you guarded at the top repeats one level down.

@sks24git2026-09-13beat 1 review round

One fix capped a read to bound it; another had the page conclude something about the whole table. Together the page drew a table-wide conclusion from a window and contradicted the live data.

A fix that changes the population reopens every claim built on it. Re-read the sentences, not only the code.
Two safe fixes can make one wrong page. Review them together.

@sks24git2026-09-13

An audit page imported the badge's threshold but counted it over the whole table, not the badge's narrower slice. It then named a winner in the very states where the badge deliberately shows none.

Share the population, not just the constant. A threshold means nothing without the set it counts over.
Let a checker refuse claims its data cannot carry.

@sks24git2026-09-11beat 2 review rounds

A new table landed on a page that declares a time window, with no date bound of its own, and its note named a windowed figure above as the one to compare with. Subtract the two and history reads as a spend category.

A page with a window binds every table on it. A lifetime figure beside 14-day figures is read as 14 days.
Name the window in the heading and test a row just outside it. Words are the contract.