TraceQL: Grouping

[joe-elliott]

What this PR does:

• Adds by() and coalesce() to the language. Purposefully only documents by() b/c I can't think of a use case for coalesce() and I want to keep docs focused.
• Reduces the tempodb search tests from 250 -> 50 traces for test sanity.
• Adds a rebatchIterator to do the work of persisting the SecondPass callback spansets through the iterators.
• Correctly persist matched and other spanset attributes to the frontend!

TODO:

• Correctly handle recombination at the query frontend
• Make sure that the order of returned spanset attributes is pinned

Which issue(s) this PR fixes:
Fixes #2136
Fixes #2307

Checklist

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

[knylander-grafana]

Thank you for adding doc! Updates look good.

[joe-elliott]

[joe-elliott]

After much hand-wringing I decided to deprecate the current "SpanSet" field on the return in favor of "SpanSets".

For now the SpanSet field is a "random" spanset from the slice. For queries that do not create multiple spansets per trace (i.e. everything but by()) there will be one SpanSet in the slice and it will match the old field.

[mdisibio]

LGTM. Very nice solution to get spans back in order for the second pass after grouping. A few q's but just surface-level stuff.

[mdisibio]

What is the reason for the slice over a map? Also map[string]Static wasn't right, it probably should be map[Attribute]Static

[joe-elliott]

the order is meaningful and needs to be preserved b/c it's established by the query, but the map can't preserve the order.

[mdisibio]

Initial thought is that the whole Attributes slice could uniquely define the spanset. Does it contain attributes we don't want?

[joe-elliott]

yeah, if the attributes are:

by(.namespace) = "prod"
by(.service.name) = "app"
avg(duration) = 1.32ms

we only want to use the first 2 to identify the spanset

[mdisibio]

Instead of a bool on the last span, we could probably rebatch them based on equal ss pointers.

[joe-elliott]

we could do that, but then the rebatchIterator would have to call Next() until it received a nil span to guarantee its batches were complete. Then it would be able to dump all batches. I did this to preserve the "streaming" nature of the iterators