we recently started compiling against spark 3.0.0-SNAPSHOT (build inhouse from master branch) to uncover any breaking changes that might be an issue for us.
we ran into some of our tests breaking where we use mllib. most of it is immaterial: we had some magic numbers hard-coded and the results are slightly different because spark changed its random number generation or because spark fixed a genuine bug in a classifier, etc. however we see somewhat significant changes in ALS factors and also in resulting recommendations. all this while there seems to be no changes in the ALS code between spark 2.4.1 and current master. we cannot come up with a good explanation so far. any idea what is going on? thanks!