MQE improve memory consumption estimate with Label information

[lamida]

What this PR does

Add memory consumption estimate with Label information. The estimation is based on the following:

• The size of labels.Label struct (32 bytes).
• Average length of string name of value of the label (8 bytes each).
• Average number of labels per series (guesstimating to 10).

I will try to figure out to improve the guesstimate of the two averages above by looking at the ops data.

Which issue(s) this PR fixes or relates to

Fixes #

Checklist

• Tests updated.
• Documentation added.
• CHANGELOG.md updated - the order of entries should be [CHANGE], [FEATURE], [ENHANCEMENT], [BUGFIX].
• about-versioning.md updated with experimental features.

[charleskorn]

This isn't quite what I had in mind: I think we need to accurately account for the size of labels given they can vary wildly, and we need to track this for anything that contains labels (SeriesMetadata slices being the main missing source not yet accounted for).

[lamida]

@charleskorn

I think we need to accurately account for the size of labels given they can vary wildly

Yes I am thinking about this too, but I started only with this estimate, because to track the size labels accurately, it means we need to move the call to types.VectorPool.Get(len(series), q.memoryConsumptionTracker) from its current place in populateVectorFromInstantVectorOperator. Now we just get the pool based on the count of series. But to account the exact size of labels, we need to move this or having additional place to calculate it at the point we load the series. Is this inline with what you have in mind for handling labels size variation precisely? I have a small worry on the performance impact if we want to measure the labels counts and the exact string size for each labels and label names though.

SeriesMetadata slices being the main missing source not yet accounted for

To be honest, I didn't think about this. Will add this. Do you think there is additional place where labels are used apart of SeriesMetadata and the time when we load sample?