range-op data extraction inconsistency

[juliusv]

There is an inconsistency during value extraction for range ops (not range-at-interval ops) when a query crosses the boundary between memory and disk. At the disk/memory boundary, e.g. a rate will drop by a small amount for the duration of the op's range.

To reproduce:

1. monitor a client_golang/examples/random instance
2. wait for the first flush to disk
3. do a rate(rpc_calls_total[5m]) query
4. observe how the rate slightly dips at the mem/disk boundary

[juliusv]

This is caused by the drop in time precision between memory and disk. Disk keeps only seconds, while memory keeps a whole time.Time for each datapoint. Now the question is, do we just want to live with that or also drop the precision in memory?

Pro for keeping higher precision in memory: better fidelity of in-memory data, especially for rates computed over it, which benefits rules, which usually access only in-memory data, so recorded rates are more precise and have less "bumps".

Con: there's a tiny bump (especially in rates) at the border between what's currently in memory and what's on disk. The bump is non-persistent and will be gone (or rather, move to a different position) on the next flush.

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.