Add LabelNamesForMetricName for the series store

[cyriltovena]

This adds LabelNamesForMetricName for the series store only, I wanted to first get your feedback if this seems correct.

• I'm getting all series ids for a given metric name
• lookupSingleChunkBySeries which takes the first index query and lookup index entries then take the first index entry and parse the chunk key.
• fetch a single chunk for each series
• merge all label names possible except metric name.

I would like a special attention to that lookupSingleChunkBySeries and its usage, not sure if this is the right way to approach the problem. (May be I should loop though queries and entries until I get a result ?)

I still need to implement this for the chunk store but I'm not sure what would be the most efficient way of doing this.

The PR is rebased on my other work to include the split of the time range validation. So don't mind the two other commit they won't be part of this PR.

Thank you !

[cyriltovena]

/cc @gouthamve @jtlisi

[bboreham]

Seems kinda gross 😬
I can't think of a better way to do it in the current schema; maybe we should think again about writing this data to the index (e.g. a metricname->labelnames row).

Back with the code as you have it, it may be worthwhile to pass a hint down to the DB interface that you only want one chunk, and/or avoid reading the same series across multiple day buckets.

[cyriltovena]

So I tried another implementation, I'm fetching all index but only one chunk per metric fingerprint.

I had to create two implementation to make it works (series+chunk store)

[cyriltovena]

I've rebased my branch, I think this is good to go.

@gouthamve @tomwilkie @jtlisi @cstyan : I would really love to get that fix for Loki, been sitting here for a month already.

[cstyan]

just a few comments, don't know enough about the chunk storage to 👍 though

[cyriltovena]

please @tomwilkie @bboreham would you be able to take a look ? From my point of view this is good to go.

[bboreham]

Seems ok. I'm a bit disappointed with the amount of repetition, but it isn't obviously easy to refactor.

Did worry about the efficiency of dedupe.

[bboreham]

This bit seems likely to be inefficient - most chunks will have the same label names, but potentially thousands of unique fingerprints, so you will add thousands of copies of the same string then sort them then dedupe them.
Suggest using a map instead.

[cyriltovena]

actuallyI'm only fetching a single chunk per unique fingerprints 257:

	filtered, keys := filterChunksByUniqueFingerprint(filtered)

but you're right unique fingerprint can have many label name duplicate. I'll improve this.

[cyriltovena]

Voila !

[tomwilkie]

LGTM from me!

[bboreham]

Thanks!