Hi! I recently joined an ML startup (mainly Python) and there are an order of magnitude more repos than people. Code sharing is ridiculously complex and costly here. I've previously worked with and on build systems (make and various permutation, in-house, Bazel w/ RBE, Gradle, etc.). As I know well the steep learning curve of Skylark and Bazel, I'm evaluating options to bring the polyrepo into a mono while; keeping it approachable to engineers and data science, deterministic and reproducible, and something that lessens the gitops/release concerns away of moving to monorepo (ie. something easily able to determine the dirty graph and present only those targets -- so we only test code and output containers that are changed -- thinking this might do it),.

"order of magnitude more repos than people" - been there, done that, did not enjoy

How about "order of magnitude more branches than engineers", where each branch is a "product". That's what we're hoping to change with Pantsbuild, and some strategic refactors.