Pickle protocol 4 uses framing for reducing the overhead of calling the read() method for small chunks of data. Most read chunks are small -- opcodes, small integers, short strings, etc, and calling read() for every 1 or 4 bytes is too expensive. But using framing itself adds an overhead. It increases the size of pickled data by 9 bytes. A frame itself needs 3 reads -- the opcode, the frame size, and a payload. Thus it doesn't make sense to create a frame containing less than 3 chunks of data. For example after pickling the list [b'a'*70000, b'b'*70000] with the Python implementation produces a data containing 3 frames of sizes 3, 1 and 3. Using frames here is completely redundant.
PR 5127 makes frames be created only when the size of the payload is not less than 4. Since the minimal size of 3 chunks is 3 bytes this is the absolute minimum of frame size. It would be better to count the number of chunks instead of bytes, but this will complicate implementations, especially Python implementation.
In the current form the change is trivial, just an additional check. Actually it fixes a regression introduced in . Currently even empty frames can be produced (when fast=True).
