Long-term index caching misses chunks on lookups

[gouthamve]

In GetChunksForSeries, we do a RangeValueStart on bucket.from. Now I thought if the bucket.hashKey is same, the bucket.from is also same, turns out that is soo not true.

bucket.from is the relative milliseconds from the actual bucket start time.

Now I did not realise this while building the long-term caching approach, I thought if the hashKey and RangeKey match, everything else matches:

cortex/pkg/chunk/schema_caching.go

Lines 99 to 100 in 78e0607

	// When deduping, the bucket values only influence TableName and HashValue
	// and just checking those is enough.

But when we split the ranges, we split the actual from and to, to "cacheable" and "active" ranges here:

cortex/pkg/chunk/schema_caching.go

Line 66 in 78e0607

cFrom, cThrough, from, through := splitTimesByCacheability(from, through, model.TimeFromUnix(mtime.Now().Add(-s.cacheOlderThan).Unix()))

and end up picking only the "active" range query on merge. This means the bucket.from is higher than it should be and we end up filtering chunks out. This causes gaps in the queries and we're dropping entire chunks to the floor.