Bryan Boreham

Complicated fix would be to send invalidation messages from ingesters to query-frontend when older data arrives. Simpler fix would be to expire recent cache results after a few minutes. We...

> Query catching already doesn't cache anything within the last 10 minutes And this is inadequate to avoid problems, as I noted in the description. Following your line of reasoning...

1. Currently, instant queries are not cached. But this could change in future. Splitting is not relevant. 2. I think what you wrote is correct. As I wrote above, "this...

I'm a "no thanks" on this pending some specific justification why we need them. The metrics were only added to Mimir via #1715, 4 months ago, and were removed upstream...

Alternative: have the forgetting done by an operator outside of the compactor ring. This way is easier to coordinate (operator can run a leader-election to ensure there is only one...

Sorry, I'm using "operator" as a term of art meaning a (Kubernetes) program that automates some functions that historically a person would have done. https://kubernetes.io/docs/concepts/extend-kubernetes/operator/

The test that is failing requires the old behaviour. IDK; I think double-counting if you manage to register the same hook twice is better than unregistering something that someone else...

@marccarre in the case where I hit this, the other person was deliberately trying to match the name used here, to avoid disruption. But they didn't anticipate both libraries being...

Feels a bit borderline - I understand you are projecting a 5% improvement when using an optional Cortex feature - but in line with the original point of adding `GetRef()`....

I disagree this is "very niche"; my guess is every Prometheus admin runs such queries. I edited the description to add that info. As a data point, on a v2.31.1...