We implemented KISS state machine based on TCP timeouts and server response.

Pseudocode:

  Send msg to server
    if socket timeout:
      reinit socket and send again
  Recv from server
   if recv.msg != 'ok':
     send again
   else
     remove msg from storage

One message has from few to few hundreds Kilobytes. Messages are stored and sent as custom binary data for better compression (better that protobuffs). ZMQ takes complexity of raw tcp sockets away and this is what we needed it for. As much and as little.