Proposal to reduce memory usage of metric aggregation

[alangy98]

Hello,

We are currently working on reducing the redundant storage of MetricAttribute. In the current implementation, the entire MetricAttributes is stored in an unordered_map. Labels with lower cardinality are stored multiple times alongside labels with higher cardinality. Additionally, attributes cannot be shared across different metrics, even though using the same or similar MetricAttributeSet to update multiple metrics is common in real-world scenarios.

Our preliminary work is to adopt the idea of HPACK, using a global table to minimize redundant storage of label keys and values. Would the community be receptive to these changes?

[lalitb]

@alangy98 Yes indeed, any such change to reduce memory usage would be welcome, as long as any such global structure won't bring further mutex locks in the hot path of instrumentation.

ps. There are also few places of heap allocations which can be optimized, but that could be different topic :)