| Re: [jetty-users] Jetty 9 websocket ping/pong messages |
|
Hi,
So it is appears not worthwhile for the application layer to try to use ping/pong.
SetMaxIdleTimeout can be used to keep connections alive for the specified time.
But connections can get closed for other reasons like loss of network connection. It appears that the best strategy for both
The server and the client is to be aware of connection close ( OnClose called by jetty) and handle it gracefully. In the
Case of the client , the client can reconnect when required. In the case of the server , it has to discard the current connection
And wait for the client to connect again.
If the connection fails while the server is trying to write to the client, I assume an IOException or some Exception will be thrown,
The server can reattempt the write when the client makes a connection the next time.
If there are other better ways to do connection management, please let me know.
regards
From: Joakim Erdfelt <joakim@xxxxxxxxxxx>
Reply-To: JETTY user mailing list <jetty-users@xxxxxxxxxxx> Date: Monday, July 28, 2014 at 5:51 PM To: JETTY user mailing list <jetty-users@xxxxxxxxxxx> Subject: Re: [jetty-users] Jetty 9 websocket ping/pong messages PING / PONG on websocket is considered a back-channel. A form of health check.
Its supposed to be automatically responded to (similar to ICMP echo, SPDY ping, HTTP/2 ping, etc...).
Many implementations do not support sending a pong.
This is because sending a pong with a different payload (something that becomes possible when you expose .sendPong()) is often interpreted as a failed PING/PONG pair.
The jetty exposure is because of JSR-356, aka javax.websocket. (This API has many small issues. the next iteration will hopefully soften the numerous papercuts present)
Know this, Jetty will *ALWAYS* respond to a PING it receives with a PONG of the same payload (as soon as it has parsed the PING, it shoves the PONG at the front of the frame queue for outgoing reply)
Therefore, using javax.websocket.RemoteEndpoint.sendPong() will result in a naked pong for a unsent ping.
The PING/PONG is also out-of-band, meaning it can exist in the middle of other messages.
Since the APIs for notifications are supposed to be thread-safe, this out of band behavior of websocket control messages means that if you were handling a large message, consisting of many websocket frames (fragments), we couldn't notify you of the PING
or PONG until your endpoint is done handling the message. (this is especially true of Decoded or Stream based message handling).
As for keeping a connection alive, that's a facility of idle timeout.
If you want to track latency, send a PING with a sequence indicator + timestamp. in your onPong, compare the sequence indicator (if desired), and the timestamp for latency. Also know that the latency you see here can be skewed by a large frame. (if a
frame is in progress, nothing else can be sent until that frame is completed).
If you want to discover network issues faster (such as a mobile device going out of coverage), then use a PING/PONG, the result will be a non-response, or have ridiculous latency, or even cause an IO error with an abnormal close.
--
Joakim Erdfelt <joakim@xxxxxxxxxxx>
Expert advice, services and support from from the Jetty & CometD experts
On Mon, Jul 28, 2014 at 4:31 PM, Manoj Khangaonkar
<Manoj.Khangaonkar@xxxxxxxxxx> wrote:
|