sum aggregation in single shard of index for long field type precision problem

[hackerwin7]

Elasticsearch version:
6.3.1, 7.5.1

Plugins installed: []

JVM version (java -version):
12.0.2
OS version (uname -a if on a Unix-like system):
Debian 4.14.81.bm.15

Description of the problem including expected versus actual behavior:

In a single shard of index, when doing a simple sum aggregation for a long mapping type field, if the sum result is huge, it will loose the precision.

Elasticsearch sum aggregation use double to sum the value with kahan summation in SumAggregator, when corrected is very small in contrast to sum, then the double newSum = sum + corrected; will loose the precision, even though the field mapping type of index is long.

[elasticmachine]

Pinging @elastic/es-search (:Search/Search)

[GrayMissing]

I think it's about NumericMetricsAggregation interface design, as it take double as its value, double can't ensure to precisely represent a big integer number greater than 2^53

[hackerwin7]

I think it's about NumericMetricsAggregation interface design, as it take double as its value, double can't ensure to precisely represent a big integer number greater than 2^53

Yes，all numeric metrics for aggregation would use double style to process

[elasticmachine]

Pinging @elastic/es-analytics-geo (:Analytics/Aggregations)

[csoulios]

This ticket is discussed further in PR #50538.

Relates to #9545, #46934 and #60149