Micrometer: Azure Registry publishing metrics with count equal to zero

[uvalashchuk]

We're using micrometer to collect metrics into Azure application insights and faced an issue when data from timer being published even when timer was not called in this period. This results into invalid aggregation, because Azure app insights replaces valueCount = 0 to valueCount = 1.

For example, we're using 1 minute collecting interval and a method which has Thread.sleep(300) in it's body. This method produces metrics described below.

For entries which have value = 0 there where no method calls, but it was pushed into app insights. This results in wrong aggregation when we're building dashboards using default metrics visualizer from Azure.

(Method executes 300ms, but on time chart we're observing average time of 150ms)

[uvalashchuk]

Also, I've noticed that there is a comment on com.microsoft.applicationinsights.telemetry.MetricTelemetry#setCount that count should be greater or equal to 1. But io.micrometer.azuremonitor.AzureMonitorMeterRegistry#trackTimer (and other methods) sets count from io.micrometer.core.instrument.Timer#count without any zero-checks despite Timer#count can return 0.

```
/**
* Sets the number of samples for this metric.
* @param count Number of samples greater than or equal to 1
*/
public void setCount(Integer count)

[vagpap]

I am also facing the issue.

In a try to overcome this, I "recreated" AzureMonitorMeterRegistry class but replaced the publish method as by adding a filter in the stream before forEach as follow:

protected void publish() {
    logger.info("Publishing information");
    for (Meter meter : getMeters()) {
        meter.match(
                this::trackGauge,
                this::trackCounter,
                this::trackTimer,
                this::trackDistributionSummary,
                this::trackLongTaskTimer,
                this::trackTimeGauge,
                this::trackFunctionCounter,
                this::trackFunctionTimer,
                this::trackMeter
        )
        .filter(m -> m.getCount().intValue() != 0) // Do not track the meter in case count == 0
        .forEach(telemetry -> {
            try {
                client.track(telemetry);
            } catch (Throwable e) {
                logger.warn("failed to track metric {} in azure monitor", meter.getId());
                TraceTelemetry traceTelemetry = new TraceTelemetry("failed to track metric " + meter.getId());
                traceTelemetry.getContext().getOperation().setSyntheticSource(SDK_TELEMETRY_SYNTHETIC_SOURCE_NAME);
                traceTelemetry.setSeverityLevel(SeverityLevel.Warning);
                client.trackTrace(traceTelemetry);
                client.flush();
            }
        });
    }
}

Most probably not the best solution, but it works, at least for Timer metrics, which we use

[uvalashchuk]

We found other way to solve this which may be useful.

Microsoft's library provides interface com.microsoft.applicationinsights.extensibility.TelemetryProcessor which allows to filter out telemetries. As we're using applicationinsights-spring-boot-starter-2.6.2, it autowires TelemetryProcessor into telemetry client. So this filtering can be done just by registering bean

Kotlin code:

@Component
class ZeroCountMetricsFilter : TelemetryProcessor {

    override fun process(telemetry: Telemetry): Boolean {
        if (telemetry !is MetricTelemetry) {
            return true
        }

        return telemetry.count != 0
    }
}

[izeye]

@uvalashchuk @vagpap Thanks for sharing! I created https://github.com/micrometer-metrics/micrometer/pull/2472 to try to resolve this.

[jonatan-ivanov]

I understand that Azure wants count to be greater than zero but I might create an issue for them because this is super weird:

Azure app insights replaces valueCount = 0 to valueCount = 1.