[Feature] making p50 not all 0 point when most request time is less than 10ms

[zhyyu]

Search before asking

• I had searched in the issues and found no similar feature requirement.

Description

When most request response time is less than 10ms (the default oal precision is 10), the percentile p50 will be all 0, like this
, and this is obscure. I read the PercentileMetrics code and found the combine function

public final void combine(@SourceFrom int value, @Arg int precision) {
        this.isCalculated = false;
        this.precision = precision;

        String index = String.valueOf(value / precision);
        dataset.valueAccumulation(index, 1L);
    }

So the index will be 0 when response time is less than 10(default precision is 10, oal: service_percentile = from(Service.latency).percentile(10);) and the index key will be multi by precision when calculate funcition called

percentileValues.put(String.valueOf(rankIdx), Long.parseLong(key) * precision);

So, the finally P50 will be 0.

Can we change the combine function in PercentileMetrics, do ceiling operate when calcuate index, so the index will not be 0, like this

public final void combine(@SourceFrom int value, @Arg int precision) {
        this.isCalculated = false;
        this.precision = precision;

        String index = String.valueOf(Math.ceil(value * 1.0 / precision));
        dataset.valueAccumulation(index, 1L);
    }

So, the P50 will be 10ms, when most response time is less than 10ms and easily to understand P50 that 50% req is less than 10ms.

Use case

like above

Related issues

No response

Are you willing to submit a PR?

• Yes I am willing to submit a PR!

Code of Conduct

• I agree to follow this project's Code of Conduct

[wu-sheng]

If 10ms as precision is too long for you, what don't you change it through OAL? 10ms is only a general meaningful latency, but you don't have to follow it.

At the same time, less than 10ms == 10ms is very scary when ppl read metrics. The point of percentile is providing the layout of the latency, rather than an accurate number. We should not scare users.

[zhyyu]

But when user see all P50 metrics is 0, they will think it is weird, their service response time is larger than 0ms.

[wu-sheng]

Again, you could change the precision in the OAL script. If you don't know what this means, read the docs of OALs. This precision is never hard coded.