There is a 10x increase in fit duration as reported here:

https://scikit-learn.org/scikit-learn-benchmarks/#linear_model.SGDRegressorBenchmark.time_fit?commits=e6b46675-b4afbeee&p-representation='sparse'

This happened between between e6b4667 (still fast) b4afbee (slow).

Note that the same estimator with dense data was not impacted.

I have not investigated the cause myself. Just spotted it when reviewing the benchmarks.

Looking at the commit messages of the output of git log e6b46675..b4afbeee, it could be the case that 1a669d8 introduced the regression but this needs confirmation:

• ENH Support float32 in SGDClassifier and SGDRegressor #25587

EDIT: this is actually a bug in the design of the benchmark rather than a performance regression. See the discussion below.