In PerDestinationBundler, if we have {use_single_sender_thread=false}}, we have 1 thread per destination sending messages. In this case, the bundler disables TCP.use_lock_to_send.

However, this can lead to corrupted data at the receiver:

packet from 192.168.1.110:7802 has different version (0.0.0) than ours (5.5.0); packet is discarded

The cause is this:

• When we send a multicast, the SendBuffer thread responsible for destination NULL sends the message to all members
• When - at the same time - we send a unicast message to one of the members, the multicast and the unicast are concurrent, and might therefore lead to stream corruption at the receiver

Solution

• Set TCP_use_lock_to_send to true when use_single_sender_thread is false
• Alternatively:
  • Handle a multicast M to {A,B,C,D} by adding M to the 4 queues. A unicast and a multicast message will therefore not be sent concurrently
  • Note: only do this for TCP / TCP_NIO2; UDP benefits from sending multicast messages as an IP multicast (single packet)