Output directories are unique by the user and source project workspace directory. Output artifacts don't clobber, even though they are all in ~/.cache.
> The net result is a Bazel (and Blaze) that are less burdened by the baggage of legacy, but the cost is a faster treadmill to keep pace with changes.
(I work on Bazel.)
This is accurate. A few of the biggest breaking change themes are:
1) Converting functionality linked within the Bazel binary into the extensibility mechanism implemented in Starlark. An example includes converting the native Java, C++, Android, Python, Protobuf, Obj-C and packaging rules into rules_java, rules_cc, etc. Many languages now are already implemented exclusively in Starlark. See rules_scala, rules_rust, rules_go and rules_haskell.
2) Starlark and Build API cleanups that accumulated over organic growth and development within Google for the past decade.
3) New build system features to support seamless integration with other build systems and package managers.
There are many layers of caching within Bazel (remote/local, inmemory/disk), but the central functional incremental engine is called Skyframe [1]. Almost every computation within Bazel that can be incrementally executed is managed in this engine.
That's right. However, there are distinct implications on how this dependency graph is constructed, analyzed, and evaluated. "Build Systems à la Carte" (Mokhov, Mitchell, Peyton Jones) [1] goes deeper into formalizing these differences.
Check out Bazel's remote execution feature and Google RBE (remote build execution) [1]. Bazel itself is language agnostic (language support comes in the form of rules).
Bazel supports persistent compilers with its worker protocol [0]. It's already supported for javac. It may be worthwhile to investigate how well zapcc can integrate with Bazel.