The Default That Moved Your Denominator

## Two counts, and what each is a property of A table's **row count** is a property of its shape: how many records it holds. It is computed without looking inside any cell, and nothing about the contents changes it. A column's **count of present values** is a property of that one column's contents: how many of those records actually recorded something for that field. The tool derives it from its own record of which cells hold a value — whatever mechanism it uses to mark absence — and reports how many do. For a table of 10,000 rows whose `amount` column reports 8,400 present values, the arithmetic is the whole story: **1,600 records have a hole in `amount`**. Those rows have not gone anywhere. They are still in the table, they still take part in a filter, a sort or a match on a key, and they still show up in the row count. They simply recorded nothing for that one field. The **absence test** — the predicate that asks, cell by cell, whether a value is absent — is the mechanism underneath. Counting the cells for which it answers false is the count of present values; counting the rows ignores it entirely. ## Why the pair is the cheapest audit you have - It costs **one pass** over the data and needs no configuration. - Run per column, the shortfalls are a **hole census**: which fields the producer actually fills and which are aspirational. - The census is comparable **across runs**. A column that was complete last month and is three percent short today is a change in an upstream system, not in your code. - It pre-empts the denominator problem, which is the next thing an interviewer asks about. - It tells you where a **shared population** will be needed later: two columns with holes on different rows cannot be folded into one statistic without somebody choosing the rows. The pairing matters because tools name these two surfaces almost identically — a length or size of the table against a count of a column's values — and the two sit one keystroke apart. Reading a review diff you frequently cannot tell which of the two numbers a report is quoting, and they agree exactly on every test fixture that happens to have no holes in it. ## What the difference proves, and what it does not | You observe | You may conclude | You may not conclude | |---|---|---| | present count below the row count | that column has holes, and exactly how many | anything about why they are missing | | present count equal to the row count | nothing in the column is *encoded* as absent | that the column is fully measured | | two columns short by the same amount | each has the same number of holes | that the holes fall on the same rows | The middle row is the one that separates someone who has read about absence from someone who has cleaned another team's extract. A **stand-in code** is a present value. Placeholder text is a present value. A measured zero is a present value and should be. Counting present values detects only the absence the tool itself encoded, so the count is a **lower bound** on the column's incompleteness, and the gap between that bound and the truth is precisely the part no automated check will find for you. The last row matters for anything that reads two columns at once. Absence is recorded per cell, not per record: two columns can each be 1,600 short and overlap on none of those rows, on all of them, or anywhere between. ## Worked On a 10,000-row table: 1. `spend` reports 9,100 present values, so it has 900 holes. 2. `visits` reports 8,700 present values, so it has 1,300 holes. 3. Rows that recorded **both** lie somewhere between 7,800 and 8,700 — the two column counts alone do not pin it down, and only a per-row check finds the real number. Step 3 is the habit worth forming. The per-column counts are not enough to tell you how many complete records you have, which is why a completeness report generally carries both the per-column counts and the count of rows complete across the fields that actually matter to the question. ## In an interview Name both numbers, say which is the larger and why, and say what the difference is. Then add the caveat before you are asked — equality proves only the absence of *encoded* holes — and finish with the practical use: this is the pair you check before you believe any average computed from data you did not produce.