fix: pod tracker logic for calculating processing rate #838
Conversation
Signed-off-by: Yashash H L <[email protected]>
yhl25
changed the title
| if pt.isActive(v.Name, podName) { | ||
| pt.activePods.PushBack(podKey) | ||
| } else { | ||
| pt.activePods.Remove(podKey) | ||
| // if a pod is not active, we can assume all the following pods are not active as well. |
There was a problem hiding this comment.
Why? A pod could be inaccessible because of node upgrade.
There was a problem hiding this comment.
that is the current behaviour we break if we find an inactive pod, to decrease the number of HTTP calls. But there is a bug in that logic, assume we have a vertex with 10 active pods and it scales down to 5, but the active pod list will still have 9 pods (only the pod with index 5 will be removed from the list which is wrong we should remove all the pods from index 5 to 9).
There was a problem hiding this comment.
I agree with a pod going down because of node upgrade and we should not delete all the pods after that index from the active list, but in the next tracker refresh the active pod list will be updated with the correct count.
There was a problem hiding this comment.
I was expecting there's a logic at the place of getting metrics call, if there's an error, remove that pod from the active pod list, but it looks like it's not there.
There was a problem hiding this comment.
+1 even I thought it was handled during the metrics call.
I thought of adding the remove logic there, but the whole purpose of the pod tracker is to track the pods, so added it here.
There was a problem hiding this comment.
I understand a pod going down for some reason will result in the deletion of all the pods with an index greater than the pod's index from the active pods list. But in the next refresh it will be updated, only for a duration of 30s we might end up with a low processing rate.
There was a problem hiding this comment.
doesn't look right, it is not a likely path to be optimized in the first place. let's do the below?
every 30s
for vertex in vertices:
for replica range vertex.maxReplica:
ping the pod
if pod is active:
add to active pod list
else
- remove all the pods from the list with index > replica index
+ remove the pod from the list
- breakThere was a problem hiding this comment.
+1
yeah, I thought about it, optimisation is not very useful here and also in case if pod is in a crash loop state our processing rate calculation will not be accurate.
There was a problem hiding this comment.
This is better.
My original suggestion to @KeranYang was, if there are consecutive x pods not active, then we break. Otherwise the default value of GetMaxReplicas()) is 50, which would take a while.
There was a problem hiding this comment.
since we refresh every 30 seconds, it should be ok
Signed-off-by: Yashash H L <[email protected]>
Signed-off-by: Yashash H L <[email protected]>
Signed-off-by: Yashash H L <[email protected]>
Signed-off-by: Yashash H L <[email protected]>
What did we fix?
Bug in pod tracking logic, assume we have a vertex with 10 active pods and it scales down to 5, but the active pod list will still have 9 pods (only the pod with index 5 will be removed from the list which is wrong we should remove all the pods from index 5 to 9).
How did we fix?
We delete all the pods one by one if it's inactive. We never skip deletion.