Currently, dump_npz either destroys all existing arrays in an npz file or append arrays with names that exist in the npz file as duplicated entries. This is a rather strange semantics, especially when loading individual arrays with load_npz doesn't follow those semantics.

A reasonable semantics would be replacing existing arrays. This requires a context object, with a role similar to HighFive::File.

NumPy doesn't support this. They only overwrite all arrays at once.

I checked out libzippp and libzip++, neither work with streams. It's primarily because dump_npy_stream and libzip are both "push" interfaces (thus, both need to run the main loop), so they cannot work together without writing a special stream class that serves as a pipe so that libzip can pull data from it...

I think there is a rather simple way to support replacing semantics given a context object. When an npz_file is opened for update, append as usual while keeping the central directory as a data structure in memory. When closing, write a temporary file with only the up-to-date arrays, finish the file with the central directory entry, do an atomic move to replace the old file. (It's possible to shrink a file in-place, but I guess that would invalidate too much I/O buffer).