The dominant advantages are the ability of rapid split and merge of ranges, the linear scalability and the performance benefits from C++.

An essential function of stardard library is to wrapper the syscalls. Besides that, you can make a live without the library. But why would you do that?

first of all, impl in C++ (JVM/GC is pain in the ass) - clear arch (only master and dataserver) - very concise config file and easy to deploy - most important, 10k nodes scalability without federation design of namespace

Lack of good documentation, no tests and possibly undefined behaviour in a few places. The code also doesn't look any cleaner than HDFS and uses some weird mix of C (*printf, error codes) and C++ (vectors, smart pointers, RAII etc).

> weird mix of C (*printf, error codes) and C++ (vectors, smart pointers, RAII etc).

Haven't looked at any code, but what you describe is very common usage.