Performance & scale

There are only four reasons spatial code is slow, and the fixes are not interchangeable. Python-level iteration, a missing spatial index, I/O and parsing, and memory pressure each look like the others from the outside, so the first move is always a measurement rather than a guess.

One of the four is different in kind. Vectorising, parallelising and chunking all reduce the constant: the work still grows with the data, just less steeply. An index changes the complexity class — a spatial join goes from comparing every pair to comparing each feature with a handful of candidates — and that is the only fix that turns an impossible job into a possible one rather than a slow one into a faster one.

The pages below cover the four bottlenecks, how to tell which you have, and the techniques for each: indexes, vectorisation, parallelism, chunking and out-of-core formats.

Learning paths run through this topic too — see the end of this page.