I have a few propositions regarding socket usage in send_pdu() and read_pdu() methods.

Here, instead of manually tracking how much is sent in a loop, using socket.sendall() method should make the code more concise and readable.

In read_pdu, we don't check len(raw_len) == 4. For less than 4 bytes we raise PDUError. But we could possibly get raw_len in multiple with recv, just like when we are reading the rest of the PDU.

I propose we refactor the "keep recving until we got all bytes" into a separate method (i.e. _recv_exact(n)), and then read raw_len and raw_pdu with this method.

If you are interested, I will send pull requests for these changes.