xds: ORCA to LRS propagation changes #12203
Conversation
shivaspeaks
changed the title
| @@ -25,6 +25,7 @@ | |||
| import com.google.common.collect.Sets; | |||
| import io.grpc.Internal; | |||
| import io.grpc.Status; | |||
| import io.grpc.xds.BackendMetricPropagation; | |||
There was a problem hiding this comment.
io.grpc.xds.client can't depend on io.grpc.xds. We moved client into its own package so it could be used without the rest of grpc.
| @@ -420,6 +421,29 @@ public void run() { | |||
| return loadCounter; | |||
| } | |||
|
|
|||
| @Override | |||
| public LoadStatsManager2.ClusterLocalityStats addClusterLocalityStats( | |||
There was a problem hiding this comment.
Shouldn't the old method just call this one with backendMetricPropagation set to null? (Feel free to do that in XdsClient.java)
| Map<String, Struct> filterMetadata, @Nullable BackendMetricPropagation backendMetricPropagation, | ||
| @Nullable OutlierDetection outlierDetection, Object endpointLbConfig, | ||
| LoadBalancerRegistry lbRegistry, Map<String, | ||
| Map<Locality, Integer>> prioritizedLocalityWeights, List<DropOverload> dropOverloads) { |
There was a problem hiding this comment.
nit: Can we add backendMetricPropagation param to the end of the methods for better consistency ?
There was a problem hiding this comment.
I will push it after outlierDetection since all the arguments taken from ClusterState are together and then others.
| if (memUtilization > 0) { | ||
| boolean shouldPropagate = true; | ||
| if (backendMetricPropagation != null) { | ||
| shouldPropagate = backendMetricPropagation.propagateMemUtilization; | ||
| } | ||
|
|
||
| if (shouldPropagate) { | ||
| String metricName = "mem_utilization"; | ||
| if (!loadMetricStatsMap.containsKey(metricName)) { | ||
| loadMetricStatsMap.put(metricName, new BackendLoadMetricStats(1, memUtilization)); | ||
| } else { | ||
| loadMetricStatsMap.get(metricName).addMetricValueAndIncrementRequestsFinished(memUtilization); | ||
| } | ||
| } | ||
| } |
There was a problem hiding this comment.
Can this be extracted out to a separate function ?
|
|
||
| public synchronized ClusterLocalityStats getClusterLocalityStats( | ||
| String cluster, @Nullable String edsServiceName, Locality locality, | ||
| @Nullable BackendMetricPropagation backendMetricPropagation) { |
There was a problem hiding this comment.
We should always pass the parsed object from the cluster resource. If the feature is not enabled then pass an empty instance.
| */ | ||
| public synchronized void recordBackendLoadMetricStats(Map<String, Double> namedMetrics) { | ||
| // If no propagation configuration is set, use the old behavior (propagate everything) |
There was a problem hiding this comment.
This should be done only when the feature is not enabled. If the feature is enabled, only when the * is specified for named_metrics we should propate everything.
Prefixing "named_metrics" should also happen only when the feature is enabled.
Also in recordTopLevelMetrics.
There was a problem hiding this comment.
This will be handled by the current code. The methods in BackendMetricPropagation are implemented in such a way to take care of these cases.
However I can now see the case where feature is enabled but no backendMetricPropagation config is available then it creates problem. Best is to check if the feature is enabled or not instead of checking null on backendMetricPropagation. I'll refactor and make it more clear paths for normal path and feature-enabled path.
I think recordTopLevelMetrics works fine I believe.
| */ | ||
| public LoadStatsManager2.ClusterLocalityStats addClusterLocalityStats( | ||
| Bootstrapper.ServerInfo serverInfo, String clusterName, @Nullable String edsServiceName, | ||
| Locality locality, @Nullable BackendMetricPropagation backendMetricPropagation) { |
There was a problem hiding this comment.
Make it non nullable and no need to create a new overloaded method. Parsed cluster resource will always have it.
| @@ -227,6 +231,12 @@ private static StructOrError<CdsUpdate.Builder> parseNonAggregateCluster( | |||
| UpstreamTlsContext upstreamTlsContext = null; | |||
| OutlierDetection outlierDetection = null; | |||
| boolean isHttp11ProxyAvailable = false; | |||
| BackendMetricPropagation backendMetricPropagation = null; | |||
|
|
|||
| if (isEnabledOrcaLrsPropagation) { | |||
There was a problem hiding this comment.
Set it to empty proto if not enabled so the null check is not required anywhere. c-core does the same https://github.com/markdroth/grpc/blob/0c1e889bb70b422899bda04166524c678d628c70/src/core/xds/grpc/xds_cluster_parser.cc#L486
There was a problem hiding this comment.
You mean something like:
else {
backendMetricPropagation = BackendMetricPropagation.fromMetricSpecs(null);
}
This doesn't seem to be a good fit. If we do this then we are unnecessarily touching the other pieces of code and dealing with empty (or default) BackendMetricPropagation.
Implements gRFC A85 (grpc/proposal#454).