A sales table is narrowed to customers whose total exceeds 1,000; how does that differ from keeping only the individual orders above 1,000?

## One English word, two operations Narrowing a table means keeping some of it and dropping the rest, but *the unit being judged* can be either a row or a whole group, and the sentence said out loud rarely makes that clear. 'Drop the small customers' and 'drop the small orders' are one word apart and produce different tables from the same input. - A **condition evaluated once per row** looks at one record's own values. Each row survives or leaves on its own merits. A customer with nine orders can end up with three. - A **condition evaluated once per group** looks at all the rows sharing a key together — their count, their total, their span — and returns a single verdict for that key. Every row of the customer survives, or none does. This is the confusion the subject exists to dissolve, and it is not only a beginner's confusion: in several ecosystems the same call name does both, and which one you get depends on whether it was applied to a table or to a **grouped handle** — what a grouping call hands back before any per-group computation has run, recording which rows belong to which key and nothing else. ## What each one needs before it can run | | Condition once per row | Condition once per group | |---|---|---| | Needs the split first | no | yes | | Input to the test | one record's values | all the rows sharing a key | | Verdict granularity | one row | one whole group | | Group sizes afterwards | changed | unchanged for survivors | | Distinct keys afterwards | possibly fewer | fewer, or the same | The last two rows are the useful ones. A group-level condition cannot change the size of a group it kept, because it never looked at rows individually. A row condition can, and usually does. ## A worked pair Take a table of orders carrying a customer key. Customer A has nine orders totalling 1,800; customer B has two orders totalling 900; customer C has a single order of 1,400. 1. **Keep customers whose total exceeds 1,000.** A and C pass, B fails. The result holds A's nine rows and C's one row: ten rows, two customers, and every order of the survivors intact — including A's small orders, none of which would clear the bar on its own. 2. **Keep orders above 1,000.** Only C's single order qualifies. The result holds one row and one customer. A, whose total is the largest in the table, has gone entirely, because no individual order cleared the bar. The two results share almost nothing, and both are reasonable answers to differently worded questions. ## What a group-level condition returns A surprising number of candidates expect the group-level form to return one row per surviving group, on the grounds that the condition was about groups. It does not. It returns the original rows of the groups that passed, at full width and full length, with nothing folded. Collapsing to one row per group is a separate operation: the condition chooses which groups, the collapse chooses the shape, and wanting one row per surviving customer means doing both. ## Reading a result backwards Given only the output, you can usually tell which one ran: - Every surviving key has exactly the row count it had in the input, so a group-level condition ran, or nothing did. - Some keys are present with fewer rows than before, so a row condition ran. - Keys have disappeared entirely, and either could be responsible: a group-level condition rejected them, or a row condition removed their last surviving row. That last line matters in a diagnosis, because a vanished customer is not by itself evidence of a cohort cut. ## They compose, and the order is a decision Both can appear in one pass: narrow the rows, split, then reject groups. But a row condition applied before the split changes what each group's summary *is*, so it also changes which groups a later group-level condition accepts. Fix the order deliberately and record it; it is not a formatting choice. ## Where this bites - A report described as covering active customers only, which actually dropped small orders, so every per-customer total it shows is lower than the real one. - A cohort defined by a group-level threshold in one place and re-derived with a row-level one in another, producing two different cohort sizes from what everyone believes is one definition. - A group-level condition written where a per-row one was meant, quietly multiplying the surviving row count by roughly the average group size and overwhelming a downstream step that was sized for the smaller result.