Default aggregation for SumObserver and UpDownSumObserver

[nilebox]

What are you trying to achieve?
Review requirements for default aggregation for SumObserver and UpDownSumObserver, and confirm what should be the behavior in all SDKs for consistency.

First of all, SDKs currently implement aggregation for SumObserver and UpDownSumObserver differently:

• Go: Sum
• Java: LastValue
• Python: LastValue
• JS: Sum

According to spec, they should be using the Sum aggregation: https://github.com/open-telemetry/opentelemetry-specification/blob/master/specification/metrics/api.md#aggregations

The additive instruments (Counter, UpDownCounter, SumObserver, UpDownSumObserver) use a Sum aggregation by default.

If there is exactly one data point per measurement interval, there is no difference between using LastValue vs Sum as far as I understand. It may be important for consistency and correct handling of cases where there happens to be more than one data point per interval though.

Additional context.

There are other cases of inconsistency in aggregation, e.g. ValueObserver which has reached consensus to use LastValue in all SDKs, but the spec still states that it should be using MinMaxSumCount. UPDATE: created #945 to track separately.

/cc @jmacd @james-bebbington @bogdandrutu

[Oberon00]

It seems the issue in question is already solved in the Spec. However, it may still be sensible to have this issue to make @open-telemetry/java-approvers and @open-telemetry/python-approvers and potentially other SIGs aware their implementation is not compliant. The metric compliance matrix is still TBD currently I think.

[bogdandrutu]

If there is exactly one data point per measurement interval, there is no difference between using LastValue vs Sum as far as I understand. It may be important for consistency and correct handling of cases where there happens to be more than one data point per interval though.

@nilebox in this case it really does not matter, because we specifically say that observers are called once per collection interval, but it may be that the observer returns multiple values for the same combination of labels (which I think is forbidden by the specs). That being said consistency should be there, and we should change all the SDKs to use the same aggregation.

[jsuereth]

Update on Java.

1. Collection should only be ONCE for observer instruments.
2. Multiple writes to the same metric-stream (i.e. same set of attributes) is considered an error and is logged, but not allowed. First written should win.

[reyang]

I think this is addressed by #1842. Please reopen if anything not covered by the SDK spec.