Scans and aggregations over billions of rows

An analytical query touches a small number of columns but a huge number of rows: a GROUP BY over a year of orders, a funnel over a billion page views, a join of a fact table with a few dimension tables. A system built for this stores each column on its own, so a query reads only the columns it names, and it compresses those columns well because the values in one column look alike. It splits the scan across every core, and often every machine, and skips whole blocks of data whose minimum and maximum values can’t match the filter.

To tell whether a system really handles it, look at how it lays data out (columnar files or row pages), whether it can skip data by time or sort key without reading it, and whether it adds more machines to one query or only to more queries. Test with your own data at full size: a query that is fast on ten million rows can still be slow on ten billion if the engine has to read every row or shuffle the whole table between nodes for a join.