Out-of-order

As of 1.6.0, dnsdist supports accepting and processing queries out-of-order as long as the maxInFlight parameter has been set on the frontend, via addLocal() and/or addTLSLocal(). Note that it is always enabled on DoH frontends. As many as maxInFlight queries will then be read from a TCP connection, processed and forwarded to a backend simultaneously. If there is more queries pending, they will be processed once a response has been sent for one of the already processed queries.

Backends are assumed not to support out-of-order by default, so only one query at a time will be sent over a TCP connection to a backend, meaning that up to maxInFlight connections to a backend might be needed to be able to process all accepted queries. Setting maxInFlight to a value greater than zero on newServer() changes that, and up to maxInFlight queries can be sent to a backend simultaneously over the same TCP connection. This of course requires the backend to actually process incoming queries out-of-order, otherwise the latency will be considerably increased, leading to timeouts and degraded service.

As of 1.6.0, only queries from the same incoming client connection will be sent to a server over a single outgoing TCP connections. This will likely change in 1.7.0, once we have had time to check that it has no adverse effects.

Backends for which Proxy Protocol support has been enabled will never be able to reuse the same outgoing TCP connections for different clients, given that the payload indicating the source IP of the client, as seen by dnsdist, is sent once at the beginning of the TCP connection. For the same reason, it might not even be possible to reuse a TCP connection for the same client if any Type-Length-Value data has been sent over that connection.