Samples can be rejected for various reasons:

• Timestamp too old
• Out of order timestamps
• Repeated timestamps
• Invalid labels or metric names
• Too many series
• ...

Our approach to these is rather ad-hoc at the moment - some are silently dropped, some return errors, and some cause (voluminous) logging. Instead we could record them as a set of synthetic metrics in cortex itself on a per-instance basis under an appropriate namespace (e.g. weaveworks_cortex_timestamp_too_old_count) so that they're a) visible to those who're sending them to us and b) available for the generation of alerts or notifications.