A solution to the issue fixed by 6d0f659 could be to use the client flow id also for the server flow id.
This would require additional step after socket_accept() functions where the server would read an information from the client about flow id (and any other information suitable here) and use that information when initializiing new flow id on the server.
With that the tcp_crr stats on the server would fall under correct flow id and would not produce such big volume of data.