Best practices
Never run blocking code in OnTraffic(), OnOpen() and OnClose()
The above three event handlers (callbacks) are executed in event-loops, therefore, running blocking code in them blocks event-loops, which means that the subsequent tasks will have to wait for the preceding blocking event handlers to complete before they get executed.
To avoid blocking event-loops, asynchronize your blocking code, for example by starting a goroutine with your blocking code and invoking c.AsyncWrite() or c.AsyncWritev() to send response data to the peer endpoint.
If you're not familiar with how gnet works, go back and read this.
Either loop read data in OnTraffic() or invoke c.Wake() regularly
Despite the fact that gnet leverages epoll/kqueue with level-triggered mode under the hook, OnTraffic() won't be invoked constantly given there is data left in the inbound buffer of a connection, OnTraffic() is only invoked when the connection has new data arriving, which is like edge-triggered mode from the user's point of view.
Thus, you should loop call c.Read()/c.Peek()/c.Next() for a connection in OnTraffic() to drain the inbound buffer of incoming data, but if you don't, then make sure you call c.Wake() periodically, otherwise you may never get a chance to read the rest of the data sent by the peer endpoint (client or server) unless the peer endpoint sends new data over.
Leverage Conn.Context() to monopolize data instead of sharing it across connections
It's recommended to use Conn.Context() to store necessary resource for each connection, so that each connection can take advantage of its exclusive resource, avoiding the contention of single resource across connections.
Enable poll_opt mode to boost performance
By default, gnet utilizes the standard package golang.org/x/sys/unix to implement pollers with epoll or kqueue, where a HASH map of fd->conn is introduced to help retrieve connections by file descriptors returned from pollers, but now the user can run go build with build tags poll_opt, like this: go build -tags=poll_opt, and gnet then switch to the optimized implementations of pollers that invoke the system calls of epoll or kqueue directly and add file descriptors to the interest list along with storing the corresponding connection pointers into epoll_data or kevent, in which case gnet can get rid of the HASH MAP of fd->conn and regain each connection pointer by the conversion of void* pointer in the I/O event-looping. In theory, it ought to achieve a higher performance with this optimization.
See #230 for code details.