A couple of years back I wrote a Java microservice that talks to Jenkins to find the state of a series of builds. The code was structured to be “observationally functional” — there were plenty of side effects, but the main data abstractions behaved as if they were immutable as far as the abstract API was concerned. This allows us to treat code that uses these objects as if it were pure functional code.
If a data structure is observationally functional, then regardless of what the implementation does, there is no way to observe side effects at the abstract level. Primarily, this means that if you call a function twice with the same arguments, you always get the same answer. (This implies, but it isn't obvious, that calling a function should not mutate anything that would cause a different function to change.) This restriction has a lot of wiggle room. You can certainly side effect anything local to the abstraction that doesn't get returned to the caller. You can side effect data until the point it is returned to the caller.
The main data abstraction my microservice works with is a representation of the build metadata tree on the Jenkins server. The higher level code walks this tree looking for builds and metadata. The code maintains the illusion that the tree is a local data structure, but the implementation of the tree contains URL references to data that is stored on the Jenkins server. As the higher level code walks the tree, the lower level code fetches the data from the Jenkins server on demand and caches it.
Writing the code this way allows me to separate the data transfer and marshaling parts from the data traversal and analysis part. The tree, though it is mutated as it is traversed, is immutable in the parts that have already been visited. The caching code, which actually mutates the tree, needs to be synchronized across multiple threads, but the traversal code does not. Nodes in the tree that have already been visited are never mutated, so no synchronization is needed.
Once the caching tree abstraction was written, the higher level code simply walks the tree, selecting and filtering nodes, then reading the field values in the nodes. But the higher level code can be treated as if it were pure functional because there are no observable side effects. An advantage of pure functional code is that it is trivially thread safe, so my microservice can run hundreds of threads in parallel, each walking separate parts of the Jenkins tree and none interfering with the other. The only part of the code that uses synchronization is the tree caching code.
This implementation approach was quite fruitful. Once the code was tested with a single thread, it was obvious that multiple threads ought to work (because they couldn't observe each other's side effects) and when I turned the thread count up, no debugging was necessary. The code has been running continuously with dozens of threads for the past couple of years with no timing, synchronization, or race condition bugs.