Better estimate for aggregated events count

[lahsivjar]

Motivation

The current algorithm calculates events total for the whole CombinedMetrics. This approach works well as long as we need to merge forward. However, when we need to split or partition CombinedMetrics then we can no longer estimate the events accurately.

Solution

Track events total for each aggregate (including overflow) rather than for the whole CombinedMetrics. This will require more storage but we will be able to track accurate events total for splits or partitions. Extra storage cost based on current limits will be:

(max_span_groups (40_000) + max_txn_groups (40_000) + max_svc_txn_groups (8_000) + overflows_per_svc (8_000) + overflow_global (1)) * (events_total_size_tracked_by_float64) 8 bytes = ~770KB

The extra storage will be used by pebble.