Should timeout be read_timeout?

[jesseforrest]

We are trying to understand the difference between the timeout and read_timeout. We noticed that c->waitms is set to timeout instead of read_timeout on this line:

phpredis/cluster_library.c

Line 818 in 31b0531

/* Set up our waitms based on timeout */

We recently reduced timeout from 5 seconds to 1 second. Now we are seeing a significant number of "Timed out attempting to find data in the correct node!" errors and would have assumed it was tied to read_timeout, but it's actually tied to timeout. Is timeout the correct value to be set for c->waitms?

Our current production settings have:

timeout = 1 sec
read_timeout = 6 sec

Any advice on the appropriate next steps would be appreciated.

[michael-grunder]

Hey,

Right so the "timeout" setting in the context of cluster refers to how long you are willing to let phpredis attempt to get your data before it returns execution to you.

Timeouts are pretty simple on a single instance as you can just use something like a connection timeout (how long to wait for a connection), and a read timeout (how long to wait to read a response).

If you're using a cluster, however, it gets more complicated. The cluster could be resharding, or failing over, etc. A way to describe it is "how long should phpredis let the cluster bounce us around before failing". It's different than a standard socket timeout, in that there may never be any TCP/IP communication error, but we still want to time out. The reason for this is so it doesn't just wait an infinite amount of time while it could be getting MOVED, ASKING, requests, etc.

Does that make sense?

Cheers,
Mike

[jesseforrest]

Thanks for the fast response Michael. We are using Redis cluster, so I do understand the logic is more complicated. Your description does clearly define the differences.

However, I'm still confused why the code is using the timeout variable instead of the read_timeout. I figured timeout would only be for establishing a connection, while read_timeout would be how long to wait to read a response. A little more clarity would be helpful. I just want to make sure the engineers here understand the difference between the two variables.

[CatKang]

Hey
I encounter the same confusion with @jesseforrest

What's more, I found the condition set on the line :

phpredis/cluster_library.c

Line 1436 in 31b0531

timedout = resp && c->waitms ? mstime() - msstart >= c->waitms : 0;

makes the client has no chance to retry after it received MOVED, when the connect timeout happened in CLUSTER_LAZY_CONNECT.

Would you please give us more clarity information?

[ppanphper]

redis_cluster_init {
...
/* Set our timeout and read_timeout which we'll pass through to the
* socket type operations */
c->timeout = timeout;
c->read_timeout = read_timeout;

/* Set our option to use or not use persistent connections */
c->persistent = persistent;

/* Calculate the number of miliseconds we will wait when bouncing around,
 * (e.g. a node goes down), which is not the same as a standard timeout.*/

**c->waitms = (long)(timeout * 1000);**

...
}

cluster_send_command {
....
/* Figure out if we've timed out trying to read or write the data */
timedout = resp && c->waitms ? mstime() - msstart >= c->waitms : 0;
...
}
the redis cluster not use read_timeout?

[yatsukhnenko]

Old issue without any activity for a long time