Outlier detection: unset failure_percent_ejection results in 100% instead of 0%

success_rate_ejection.enforcement_percentage defaults to 100% per Envoy and our spec:

https://github.com/grpc/proposal/blob/4125fbd155353d35386d8fbd6b23454f6ab0e345/A50-xds-outlier-detection.md?plain=1#L72-L75

This is implemented (correctly) here:

grpc-java/xds/src/main/java/io/grpc/xds/EnvoyServerProtoData.java

Lines 379 to 396 in ed3bf6b

    
           SuccessRateEjection successRateEjection; 
        
           // If success rate enforcement has been turned completely off, don't configure this ejection. 
        
           if (envoyOutlierDetection.hasEnforcingSuccessRate() 
        
               && envoyOutlierDetection.getEnforcingSuccessRate().getValue() == 0) { 
        
             successRateEjection = null; 
        
           } else { 
        
             Integer stdevFactor = envoyOutlierDetection.hasSuccessRateStdevFactor() 
        
                 ? envoyOutlierDetection.getSuccessRateStdevFactor().getValue() : null; 
        
             Integer enforcementPercentage = envoyOutlierDetection.hasEnforcingSuccessRate() 
        
                 ? envoyOutlierDetection.getEnforcingSuccessRate().getValue() : null; 
        
             Integer minimumHosts = envoyOutlierDetection.hasSuccessRateMinimumHosts() 
        
                 ? envoyOutlierDetection.getSuccessRateMinimumHosts().getValue() : null; 
        
             Integer requestVolume = envoyOutlierDetection.hasSuccessRateRequestVolume() 
        
                 ? envoyOutlierDetection.getSuccessRateRequestVolume().getValue() : null; 
        
             successRateEjection = SuccessRateEjection.create(stdevFactor, enforcementPercentage, 
        
                 minimumHosts, requestVolume); 
        
           }

But failure_percent_ejection should default to zero instead (i.e. disable checking when unset):

https://github.com/grpc/proposal/blob/4125fbd155353d35386d8fbd6b23454f6ab0e345/A50-xds-outlier-detection.md?plain=1#L148

This is implemented (incorrectly) here:

grpc-java/xds/src/main/java/io/grpc/xds/EnvoyServerProtoData.java

Lines 398 to 414 in ed3bf6b

    
           FailurePercentageEjection failurePercentageEjection; 
        
           if (envoyOutlierDetection.hasEnforcingFailurePercentage() 
        
               && envoyOutlierDetection.getEnforcingFailurePercentage().getValue() == 0) { 
        
             failurePercentageEjection = null; 
        
           } else { 
        
             Integer threshold = envoyOutlierDetection.hasFailurePercentageThreshold() 
        
                 ? envoyOutlierDetection.getFailurePercentageThreshold().getValue() : null; 
        
             Integer enforcementPercentage = envoyOutlierDetection.hasEnforcingFailurePercentage() 
        
                 ? envoyOutlierDetection.getEnforcingFailurePercentage().getValue() : null; 
        
             Integer minimumHosts = envoyOutlierDetection.hasFailurePercentageMinimumHosts() 
        
                 ? envoyOutlierDetection.getFailurePercentageMinimumHosts().getValue() : null; 
        
             Integer requestVolume = envoyOutlierDetection.hasFailurePercentageRequestVolume() 
        
                 ? envoyOutlierDetection.getFailurePercentageRequestVolume().getValue() : null; 
        
             failurePercentageEjection = FailurePercentageEjection.create(threshold, 
        
                 enforcementPercentage, minimumHosts, requestVolume); 
        
           }

Note that it's a copy/paste of the enforcement_percentage above it, meaning it is only disabled if it is explicitly set to zero.

When we go to actually build the enforcer, failurePercentage is non-null:

grpc-java/xds/src/main/java/io/grpc/xds/CdsLoadBalancer2.java

Lines 632 to 652 in ed3bf6b

    
           FailurePercentageEjection failurePercentage = outlierDetection.failurePercentageEjection(); 
        
           if (failurePercentage != null) { 
        
             OutlierDetectionLoadBalancerConfig.FailurePercentageEjection.Builder 
        
                 failurePercentageConfigBuilder = new OutlierDetectionLoadBalancerConfig 
        
                 .FailurePercentageEjection.Builder(); 
        
             if (failurePercentage.threshold() != null) { 
        
               failurePercentageConfigBuilder.setThreshold(failurePercentage.threshold()); 
        
             } 
        
             if (failurePercentage.enforcementPercentage() != null) { 
        
               failurePercentageConfigBuilder.setEnforcementPercentage( 
        
                   failurePercentage.enforcementPercentage()); 
        
             } 
        
             if (failurePercentage.minimumHosts() != null) { 
        
               failurePercentageConfigBuilder.setMinimumHosts(failurePercentage.minimumHosts()); 
        
             } 
        
             if (failurePercentage.requestVolume() != null) { 
        
               failurePercentageConfigBuilder.setRequestVolume(failurePercentage.requestVolume()); 
        
             } 
        
             configBuilder.setFailurePercentageEjection(failurePercentageConfigBuilder.build());

But enforcement percentage is null, so the default percentage of 100 is applied:

grpc-java/util/src/main/java/io/grpc/util/OutlierDetectionLoadBalancer.java

Lines 1116 to 1121 in ed3bf6b

    
           public static class Builder { 
        
             int threshold = 85; 
        
             int enforcementPercentage = 100; 
        
             int minimumHosts = 5; 
        
             int requestVolume = 50;

The fix appears to be trivial. In EnvoyServerProtoData.java, change the check to:

if (!envoyOutlierDetection.hasEnforcingFailurePercentage()
    || envoyOutlierDetection.getEnforcingFailurePercentage().getValue() == 0) {
  failurePercentageEjection = null;
} else {
  // ...

Hi!
I would like to work on this if no one on the team has picked it up yet.
I will go with the fix suggested above, and add tests for the unset, explicit-zero, and non-zero cases.
Happy to have a PR up within a few days.

Navigation Menu

Repository navigation

Outlier detection: unset `failure_percent_ejection` results in 100% instead of 0% #13165

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	SuccessRateEjection successRateEjection;
	// If success rate enforcement has been turned completely off, don't configure this ejection.
	if (envoyOutlierDetection.hasEnforcingSuccessRate()
	&& envoyOutlierDetection.getEnforcingSuccessRate().getValue() == 0) {
	successRateEjection = null;
	} else {
	Integer stdevFactor = envoyOutlierDetection.hasSuccessRateStdevFactor()
	? envoyOutlierDetection.getSuccessRateStdevFactor().getValue() : null;
	Integer enforcementPercentage = envoyOutlierDetection.hasEnforcingSuccessRate()
	? envoyOutlierDetection.getEnforcingSuccessRate().getValue() : null;
	Integer minimumHosts = envoyOutlierDetection.hasSuccessRateMinimumHosts()
	? envoyOutlierDetection.getSuccessRateMinimumHosts().getValue() : null;
	Integer requestVolume = envoyOutlierDetection.hasSuccessRateRequestVolume()
	? envoyOutlierDetection.getSuccessRateRequestVolume().getValue() : null;

	successRateEjection = SuccessRateEjection.create(stdevFactor, enforcementPercentage,
	minimumHosts, requestVolume);
	}

	FailurePercentageEjection failurePercentageEjection;
	if (envoyOutlierDetection.hasEnforcingFailurePercentage()
	&& envoyOutlierDetection.getEnforcingFailurePercentage().getValue() == 0) {
	failurePercentageEjection = null;
	} else {
	Integer threshold = envoyOutlierDetection.hasFailurePercentageThreshold()
	? envoyOutlierDetection.getFailurePercentageThreshold().getValue() : null;
	Integer enforcementPercentage = envoyOutlierDetection.hasEnforcingFailurePercentage()
	? envoyOutlierDetection.getEnforcingFailurePercentage().getValue() : null;
	Integer minimumHosts = envoyOutlierDetection.hasFailurePercentageMinimumHosts()
	? envoyOutlierDetection.getFailurePercentageMinimumHosts().getValue() : null;
	Integer requestVolume = envoyOutlierDetection.hasFailurePercentageRequestVolume()
	? envoyOutlierDetection.getFailurePercentageRequestVolume().getValue() : null;

	failurePercentageEjection = FailurePercentageEjection.create(threshold,
	enforcementPercentage, minimumHosts, requestVolume);
	}

	FailurePercentageEjection failurePercentage = outlierDetection.failurePercentageEjection();
	if (failurePercentage != null) {
	OutlierDetectionLoadBalancerConfig.FailurePercentageEjection.Builder
	failurePercentageConfigBuilder = new OutlierDetectionLoadBalancerConfig
	.FailurePercentageEjection.Builder();

	if (failurePercentage.threshold() != null) {
	failurePercentageConfigBuilder.setThreshold(failurePercentage.threshold());
	}
	if (failurePercentage.enforcementPercentage() != null) {
	failurePercentageConfigBuilder.setEnforcementPercentage(
	failurePercentage.enforcementPercentage());
	}
	if (failurePercentage.minimumHosts() != null) {
	failurePercentageConfigBuilder.setMinimumHosts(failurePercentage.minimumHosts());
	}
	if (failurePercentage.requestVolume() != null) {
	failurePercentageConfigBuilder.setRequestVolume(failurePercentage.requestVolume());
	}

	configBuilder.setFailurePercentageEjection(failurePercentageConfigBuilder.build());

	public static class Builder {
	int threshold = 85;
	int enforcementPercentage = 100;
	int minimumHosts = 5;
	int requestVolume = 50;

Outlier detection: unset failure_percent_ejection results in 100% instead of 0% #13165

Description

Activity

ahmadmaaz commented on Nov 6, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Outlier detection: unset `failure_percent_ejection` results in 100% instead of 0% #13165