Five minutes with salary_min_usd in data-job-postings

Someone asked what is in salary_min_usd on data-job-postings, and the honest answer took one query.

3,500 rows, 762 nulls (21.8%), 436 distinct values. Bottom of the disclosed band. Blank when undisclosed.

The five-number version: min 17,000, p10 41,500, median 98,000, p90 171,150, max 307,000. The mean is 104,106.

The p10 to p90 band — 41,500 to 171,150 — is where ordinary rows live, and it is the pair worth quoting when somebody asks what to expect. Min and max describe the two strangest rows in the table and nothing else; they are useful for spotting impossible values and misleading for everything else.

762 rows have no value at all here, which is 21.8% of the table. That is enough to move an aggregate and small enough that nobody notices it doing so — every average above is computed over 2,738 rows, not 3,500, and the two denominators produce different answers.

select
  count(*)                         as rows,
  count(salary_min_usd)            as present,
  count(*) - count(salary_min_usd) as nulls,
  count(distinct salary_min_usd)   as distinct_values
from data_job_postings;

Nothing dramatic: the mean and median are within 6.2% of each other, so an average is a fair summary. That is worth confirming rather than assuming — it is not true of most money columns.

Where this matters: a symmetric column is one you can average, threshold and chart without hedging, which makes it unusually cheap to work with. Knowing which of your columns are like this and which are not is most of knowing when to be careful.

The grain is one row per posting, which is the context every one of those numbers depends on. None of them survive a change of grain, which is why "profile the column" and "profile the table" are the same job. Full schema, and the CSV, on the dataset page.