Call.cancel() doesn’t cancel slow DNS lookups

[danvinokour]

I simulated network problems (100% packet loss), called Call.cancel() after 5 seconds and it actually took about 42 seconds for the execute method to throw IOException.

Sometimes i get exception after connect timeout (set on 3 seconds) but if connection was successful it takes very long time for cancel to work (if it really works and this is not the read timeout kicking in which is set to 25 seconds)

776 is the thread that suppose to cancel the call.
ModernAsyncTask #5 is the thread from which call is executed.

logs:
E/Censored ( 4870): [Thread-776] sleeping for 5000 ms
E/Censored ( 4870): [ModernAsyncTask #5] before execute
E/Censored ( 4870): [Thread-776] after cancel - took 5005 ms
E/Censored ( 4870): [ModernAsyncTask #5] IOException 42680ms | error - {}java.net.UnknownHostException: Unable to resolve host "Censored": No address associated with hostname

[swankjesse]

We have no mechanism to cancel during DNS.

[danvinokour]

You already have this task: #2191
which will be a great help in this field.

Do you think that once we have access to deadline we will be covered and will be able to count on timeout?

Can you insert a mechanism to cancel during DNS?

[swankjesse]

At the moment the APIs Java gives us for DNS are quite limited. You could implement your own Dns class, possibly one that uses two threads so that the waiting thread can always break prematurely.

[danvinokour]

Im no expert in this field but currently what im going to do is call.execute() in a different thread which will return an answer to Response[0] and if execute takes more than needed i will just kill this thread.

Maybe you could make something similar but more safe socket wise etc...?

Also, do you think that when okio's deadline will be reachable, will it definitely make the call exit immediately after the deadline expires?

[swankjesse]

Unfortunately OkHttp doesn’t use okio for DNS. That call doesn’t do I/O; it’s just a magical system call. It’d be pretty awesome to do a standalone DNS API, but that’d be potentially tricky, especially since you’d want to honor the local operating system’s cache & configuration.

[swankjesse]

(Your approach to using 2 threads is right. But at the moment OkHttp doesn’t signal the thread that’s doing DNS to break out prematurely. We’ll need to address this.)

[JakeWharton]

Can this be closed as a dupe of #95?

[JakeWharton]

Or rather, should we think of #95 as expanded to encapsulate explicit cancel vs. timeout-based cancelation since they likely have similar fixes?

[swankjesse]

Sure.