Remove usage of cluster level setting for circuit breaker #2567
Conversation
jmazanec15
added
Refactoring
jmazanec15
requested review from
heemin32,
navneet1v,
VijayanB,
vamshin,
naveentatikonda,
junqiu-lei,
martin-gaievski,
ryanbogan,
luyuncheng,
shatejas,
0ctopus13prime and
Vikasht34
as code owners
jmazanec15
force-pushed
the
issue-1607
branch
from
2318611 to
3eaf8b3
Compare
Simplification of circuit breaker management. Before, we were periodically updating the circuit breaker cluster setting to have a global circuit breaker. However, this is inconvenient and not actually useful. This change removes that logic and only trips based on node level circuit breaker. For knn stats, in order to fetch the cluster level circuit breaker, a transport call is made to check all of the nodes. This isnt super efficient but its made for stats calls so its not on the critical path. Signed-off-by: John Mazanec <[email protected]>
jmazanec15
force-pushed
the
issue-1607
branch
from
3eaf8b3 to
6365da6
Compare
| } | ||
| } | ||
| }; | ||
| this.threadPool.scheduleWithFixedDelay(runnable, TimeValue.timeValueSeconds(CB_TIME_INTERVAL), ThreadPool.Names.GENERIC); | ||
| threadPool.scheduleWithFixedDelay(runnable, TimeValue.timeValueSeconds(CB_TIME_INTERVAL), ThreadPool.Names.GENERIC); |
There was a problem hiding this comment.
I could go even a step further an move this to the native cache manager to simplfiy things more.
There was a problem hiding this comment.
I made this change - I think its fairly straightforward and completely removes the potential circular dependency.
| @Override | ||
| public Boolean get() { | ||
| try { | ||
| KNNCircuitBreakerTrippedResponse knnCircuitBreakerTrippedResponse = client.execute( |
There was a problem hiding this comment.
TODO: This is going to have to somehow be made to be async, which will complicate things.
There was a problem hiding this comment.
I made this change. Its a bit messy - will clean up a bit.
Signed-off-by: John Mazanec <[email protected]>
Signed-off-by: John Mazanec <[email protected]>
| import java.util.Map; | ||
| import java.util.function.Function; | ||
|
|
||
| public class CircuitBreakerStat extends KNNStat<Boolean> { |
There was a problem hiding this comment.
TODO: In mixed cluster state, this should fall back to reading from the (deprecated) index setting.
There was a problem hiding this comment.
I fixed this in latest revision
Signed-off-by: John Mazanec <[email protected]>
Signed-off-by: John Mazanec <[email protected]>
jmazanec15
force-pushed
the
issue-1607
branch
2 times, most recently
from
b910f38 to
df687f7
Compare
Signed-off-by: John Mazanec <[email protected]>
jmazanec15
force-pushed
the
issue-1607
branch
from
df687f7 to
ecd1318
Compare
|
@jmazanec15 - I like the overall idea of getting rid of the cluster level breaker and relying on the node level breakers. It will get us closer to the backpressure constructs: https://opensearch.org/blog/shard-indexing-backpressure-in-opensearch/ |
| @@ -500,4 +486,21 @@ private void startMaintenance(Cache<String, NativeMemoryAllocation> cacheInstanc | |||
|
|
|||
| maintenanceTask = threadPool.scheduleWithFixedDelay(cleanUp, interval, ThreadPool.Names.MANAGEMENT); | |||
| } | |||
|
|
|||
| private void circuitBreakerUpdater() { | |||
| Runnable runnable = () -> { | |||
There was a problem hiding this comment.
I think the previous approach was directly in the codepath. Are we moving to runnable tasks because we want a mechanism to unset the CB?
I would rather have it as a side effect than a constant monitor.
There was a problem hiding this comment.
This code was actually in the KNNCircuitBreaker class: https://github.com/opensearch-project/k-NN/blob/main/src/main/java/org/opensearch/knn/index/KNNCircuitBreaker.java#L60-L109. I moved it to the cache manager because it removes the circular dependency between the 2. However, the logic remains consistent
| @Override | ||
| public ActionListener<Void> setupContext(KNNStatFetchContext knnStatFetchContext, ActionListener<Void> actionListener) { | ||
| // If there are any nodes in the cluster before 3.0, then we need to fall back to checking the CB | ||
| if (minVersionSupplier.get().compareTo(Version.V_3_0_0) < 0) { |
There was a problem hiding this comment.
nit: can we use the compare from Version class directly? minVersionSupplier.get().orOrBefore(Version.V_3_0_0)
@kotwanikunal Why do you say that? We only trip up per node. |
Signed-off-by: John Mazanec <[email protected]>
|
|
@kotwanikunal Oh I see. Yeah I figured thatd be a breaking change and so couldnt remove it without deprecation. Hence, the stats update. I might be mistaken |
Description
Simplification of circuit breaker management. Before, we were periodically updating the circuit breaker cluster setting to have a global circuit breaker. However, this is inconvenient and not actually useful.
This change removes that logic and only trips based on node level circuit breaker. For knn stats, in order to fetch the cluster level circuit breaker, a transport call is made to check all of the nodes. This isnt super efficient but its made for stats calls so its not on the critical path.
Overall, this greatly simplfiies the code. However, it does mean that the cluster setting "knn.circuit_breaker.triggered" will no longer have an effect. But, I think for 3.0 this is okay. For users interested in this info, they can get it from the stats API.
Still fixing up tests pending, but I wanted to get opinions early on the change before investing that time to fix.
Related Issues
Resolves #1607
Check List
--signoff.By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.
For more information on following Developer Certificate of Origin and signing off your commits, please check here.