Is performance important enough for the MetricDescriptor Type to encode it as a unit64?

[bogdandrutu]

Currently there is a TODO in the metrics proto that needs a decision:

// TODO: Determine if encoding all possible values in a uint64 bitset improves performance significantly and propose that if that is the case.

[jmacd]

We have discussed that the performance implications of this are more important in the collector. Clients can cache descriptor objects and/or use alternate encoders fairly easily.

[tigrannajaryan]

It is very likely that there will be significant performance gains for single datapoint metrics. Benchmarking using https://github.com/tigrannajaryan/exp-otelproto should be easy.

[tigrannajaryan]

Clients can cache descriptor objects and/or use alternate encoders fairly easily.

True, but I suspect most clients won't bother, so would be good not to rely on this. Using bitmasks is trivial in most (?) languages, so I don't think usability-wise it is any worse than oneof with a bunch of messages (in fact my preference would be a single bitfield with masks regardless of performance).

[jmacd]

(In #168 I wrote a program to generate the bits of all valid combinations...)

[bogdandrutu]

@tigrannajaryan are you most worried about bytes on the wire or number of allocations?

The biggest problem with PR #168 was that it makes things harder to use since not every possible combination in a bitmap is valid.

[bogdandrutu]

As an example the current Gauge does not have temporality and we tried to add a new temporality just for that, then found hard to reason about all the corner cases

[tigrannajaryan]

@tigrannajaryan are you most worried about bytes on the wire or number of allocations?

The later.

The biggest problem with PR #168 was that it makes things harder to use since not every possible combination in a bitmap is valid.

Can we provide valid combinations of bitmasks as enums/consts and clearly document that other combinations are not valid?

[bogdandrutu]

The later.

Let me give a try to use gogo proto to avoid allocations. If that works we should not worry and complicate things.