RFC: Tools to block and/or drop data which is bad

[bboreham]

Suppose a user suddenly starts sending thousands of samples where the labels change each time: the ingester will create new series for each of them and quickly grow in memory. (cf #665)

Many installations of Cortex will have an upstream component to validate users, so we perhaps don't need to block inside Cortex, but we may want to have a button to erase all series from the flush queue for a particular user.

[bboreham]

Another variant: drop all series for a particular metric as they arrive, because it has been observed to be very high-cardinality.

[gouthamve]

I think this is fixed with the -ingester.min-chunk-length

[bboreham]

Some aspects are improved, but this feature is to block before the data gets into ingesters.

[gouthamve]

Closing this in favour of #1507