Use a state machine to implement AsyncTimeout #1397
Conversation
| head = AsyncTimeout() | ||
| Watchdog().start() | ||
| } | ||
| private fun insertIntoQueue(node: AsyncTimeout, timeoutNanos: Long, hasDeadline: Boolean) { |
There was a problem hiding this comment.
Most of the changes in this function are just whitespace. I moved withLock {} to the caller.
| } | ||
| prev = prev.next | ||
| } | ||
|
|
||
| // The node wasn't found in the linked list: it must have timed out! | ||
| return true |
There was a problem hiding this comment.
Now that we keep a state machine we don’t need to traverse the linked list to learn this
Previously we had 3 boolean state variables to track 4 possible states. That was more complex than necessary.
swankjesse
force-pushed
the
jwilson.1216.state_machine
branch
from
d3d1ad6 to
ab9e0df
Compare
| @@ -57,11 +56,6 @@ public class okio/AsyncTimeout : okio/Timeout { | |||
| public final fun withTimeout (Lkotlin/jvm/functions/Function0;)Ljava/lang/Object; | |||
| } | |||
|
|
|||
| public final class okio/AsyncTimeout$Companion { | |||
There was a problem hiding this comment.
Ugh, I think these snuck in by accident. I don’t believe they belong in the public API, and I don’t think anyone ever used them.
|
Hello, I'm not sure if this is the right place to ask, but I hope @swankjesse, as the author of several timeout-related PRs, might be able to help. Please know that I ask this question with the utmost respect and curiosity, and it is not meant to be critical or blaming in any way. Is there a specific reason why I'm asking because I've observed that list manipulations, such as linear time insertions into the linked list, have become a bottleneck in a back-end service that heavily relies on Thank you for any insights you can provide. Best regards, Yury |
|
@yyaroshevich that’s a good point! Wanna open a tracking bug for us to consider switching? At the time of creation we optimized for fewest allocations, which were particularly slow on Android. One other option that we should investigate – could we accomplish better performance by searching backwards instead of forwards in the list? If all of the timeouts are symmetric (they often are), then inserts will usually be at the end. We could reduce contention without increasing allocation? |
Sure, I've opened ticket where we can proceed discussion #1488. Thanks a lot! |
Previously we had 3 boolean state variables to track 4 possible states. That was more complex than necessary.