GeezerVison

I think it's recommended to manually release a memoryview. Implementations other than CPython may use a garbage collector that doesn't immediately finalize unreferenced objects. Maybe it's simpler to always use a memoryview:

        # Accept any data that supports the buffer protocol
        with memoryview(data) as data:
            nbytes = data.nbytes
            self._file_size += nbytes
            self._crc = crc32(data, self._crc)
            if self._compressor:
                data = self._compressor.compress(data)
                self._compress_size += len(data)
            self._fileobj.write(data)
            return nbytes

Good point.
There are some sites in stdlib have this problem, maybe they can be solved in another issue together.
In addition, I suspect if other Python implementations don't release the underlying buffer of memoryview, the program will run abnormally.

Let @serhiy-storchaka decide.

update: I ran a benchmark, the performances are no significant different.

@eryksun is right. But it may be simpler to merge the current solution and fix issues with releasing memoryviews in all places uniformly later.

_file_size is a private attribute. It is better to use a public API in tests. For example read the content of the file.

It would be better to move this test to AbstractWriterTests and test with different compressions.

@eryksun is right. But it may be simpler to merge the current solution and fix issues with releasing memoryviews in all places uniformly later.

_ZipWriteFile is a private class. It would be better to reword the NEWS entry in terms of the public API.