Materialize/doc/user/layouts/shortcodes/idiomatic-sql/materialize-window-functions.html

For [window functions](/sql/functions/#window-functions), when an input record
in a partition (as determined by the `PARTITION BY` clause of your window
function) is added/removed/changed, Materialize recomputes the results for the
entire window partition. This means that when a new batch of input data arrives
(that is, every second), **the amount of computation performed is proportional
to the total size of the touched partitions**.

For example, assume that in a given second, 20 input records change, and these
records belong to **10** different partitions, where the average size of each
partition is **100**. Then, amount of work to perform is proportional to
computing the window function results for **10\*100=1000** rows.

As a rule of thumb, if the total size of all touched window partitions is at
most 1000000 rows per second, then the system should be able to keep up with the
input data as it arrives. However, if your use case has higher performance
requirements, consider rewriting your query to not use window functions. If your
query cannot be rewritten without the window functions and the performance of
window functions is insufficient for your use case, please [contact our
team](/support/).