Percentiles Bucket Aggregation relies on ordered `percents` array

[cbuescher]

An interesting edge case that I found when working on #24330: It seems that percentiles_bucket relies on the percents parameter beeing specifies in increasing order or magnitude. If the list of percentiles to calculate is provided in a different order, e.g.

"percentiles_monthly_sales": {
            "percentiles_bucket": {
                "buckets_path": "sales_per_month>sales", 
                "percents": [ 50.0, 25.0, 75.0 ] 
            }
        }

the following error is thrown:

"type": "illegal_argument_exception",
    "reason": "Percent requested [50.0] was not one of the computed percentiles.  Available keys are: [50.0, 25.0, 75.0]"

The reason seems to be that in InternalPercentileBucket#percentile(double percent) (

elasticsearch/core/src/main/java/org/elasticsearch/search/aggregations/pipeline/bucketmetrics/percentile/InternalPercentilesBucket.java

Line 75 in f217eb8

int index = Arrays.binarySearch(percents, percent);

) the internal array is assumed to be sorted, but we don't enforce this, e.g. in the PercentilesBucketPipelineAggregator or even before when parsing the request.

I think we could either throw an error early if we think the percents need to be provided in order, re-sort the array if the user provided it unsorted or change the implementation of InternalPercentileBucket so it doesn't rely on an ordered array.

[cbuescher]

@colings86 do you have any preferences whether we should throw an error here or should try to make InternalPercentileBucket work without relying on an ordered array?

[colings86]

I think we should not rely on the order of the array the user provides

[tlrx]

Good catch @cbuescher