My org went from a monorepo where every project had to obey the same CI model and you could not introduce entirely new CI tools for new prototypes over to a polyrepo with separated semver library repos for shared dependencies, and it simplified everything so much.
Adding additional PRs across different repos is functionally no different than the same PR with scattered dependencies in a monorepo, except that separating the PRs makes each isolated set of changes more atomic and focused, which has led to fewer bugs and better quality code review and, the hugest win, each repo is free to use whatever CI & deployment tooling it needs, with absolutely no constraints based on whatever CI or deployment tool another chunk of code in some other repo uses.
The last point is not trivial. Lots of people glibly assume you can create monorepo solutions where arbitrary new projects inside the monorepo can be free to use whatever resource provisioning strategy or language or tooling or whatever, but in reality this not true, both because there is implicit bias to rely on the existing tooling (even if it’s not right for the job) and monorepos beget monopolicies where experimentation that violates some monorepo decision can be wholly prevented due to political blockers in the name of the monorepo.
One example that has frustrated me personally is when working on machine learning projects that require complex runtime environments with custom compiled dependencies, GPU settings, etc.
The clear choice for us was to use Docker containers to deliver the built artifacts to the necessary runtime machines, but the whole project was killed when someone from our central IT monorepo tooling team said no. His reasoning was that all the existing model training jobs in our monorepo worked as luigi tasks executed in hadoop.
We tried explaining that our model training was not amenable to a map reduce style calculation, and our plan was for a luigi task to invoke the entrypoint command of the container to initiate a single, non-distributed training process (I have specific expertise in this type of model training, so I know from experience this is an effective solution and that map reduce would not be appropriate).
But it didn’t matter. The monorepo was set up to assume model training compute jobs had to work one way and only one way, and so it set us back months from training a simple model directly relevant to urgent customer product requests.
Had we been able to set this up as a separate repo where there were no global rules over how all compute jobs must be organized, and used our own choice of deployment (containers) with no concern over whatever other projects were using / doing, we could have solved it in a matter of a few days.
In my experience, this type of policy blocker is uniquely common to monorepos, and easily avoided in polyrepo situations. It’s just a whole class of problem that rarely applies in a polyrepo setting, but almost always causes huge issues with monorepo policies and fixed tooling choices that end up being a poor fit for necessary experiments or innovative projects that happen later.