The arithmetic behind a cheap query

Query cost on a columnar store is mostly one number: bytes scanned. Here is that number worked out on city-air-quality.

Naming 3 of 8 columns cuts the scan to roughly 37.5%. Filtering to one of 1,370 days cuts it to 0.1%. Together — and they multiply — the query reads about 0.03% of what select * would.

-- reads every column, every day
select * from city_air_quality;

-- reads 3 columns, one day
select reading_date, city, pm25
from city_air_quality
where reading_date >= date '2023-06-01'
  and reading_date <  date '2023-06-02';

The second version is not a micro-optimisation. On a real warehouse those two queries differ by three orders of magnitude in cost, and the expensive one is the one that is easier to type.

The trap worth knowing: wrapping the partition column in a function — date(reading_date) = '2023-06-01' — usually defeats pruning, because the engine can no longer reason about the raw column. Same result, full scan, no warning. Compare against a range on the bare column instead.

Read the bytes-scanned line in the plan before optimising anything else. It is usually the entire answer. Why your query costs what it costs has the rest.