You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In #17354 we added prioritization to client allocation update frequency. Updates are sent on a 1000ms interval unless there's a high-priority update (like a failed allocation) which shrinks the interval to 200ms for that batch.
We didn't want to expose this behavior as a required tunable because users should mostly not have to care about it. The user-facing behavior didn't meaningfully change for most users because the speed of failure detection was unchanged. But for users who don't care about liveness so much (mostly folks running huge batch workloads), it could be valuable to have a configuration knob that lets them increase the interval even further. For example, if you're not serving live requests from your workloads, you might not care about the difference between 200ms and 10s in terms of failure recovery.
The text was updated successfully, but these errors were encountered:
In #17354 we added prioritization to client allocation update frequency. Updates are sent on a 1000ms interval unless there's a high-priority update (like a failed allocation) which shrinks the interval to 200ms for that batch.
We didn't want to expose this behavior as a required tunable because users should mostly not have to care about it. The user-facing behavior didn't meaningfully change for most users because the speed of failure detection was unchanged. But for users who don't care about liveness so much (mostly folks running huge batch workloads), it could be valuable to have a configuration knob that lets them increase the interval even further. For example, if you're not serving live requests from your workloads, you might not care about the difference between 200ms and 10s in terms of failure recovery.
The text was updated successfully, but these errors were encountered: