Avoid sorted source for time_series aggs without rates

[dnhatn]

With this change, TS index | STATS ... will be translated to FROM index METADATA _tsid | STATS ... to avoid emitting docs in _tsid and timestamp order, which is expensive. For example, this reduces the execution time of the below query with tsdb track from 50 seconds to 4.5 seconds.

TS tsdb 
| STATS sum(max_over_time(kubernetes.container.memory.usage.bytes)) BY bucket(@timestamp, 5minute)

[elasticsearchmachine]

Pinging @elastic/es-storage-engine (Team:StorageEngine)

[kkrik-es]

To make sure I understand this, with this change the inner agg is still evaluated per tsid, but we can push this down to segments since we don't need global ordering per shard?

[dnhatn]

Yes, that's correct. We only need the global ordering per shard for rate aggregations to avoid holding all data points.

[martijnvg]

👍

[dnhatn]

@kkrik-es @martijnvg Thanks!