Currently flushing is driven off a priority queue in order of the start time of the first unflushed chunk in a series.

This is especially bad when we’ve scaled down write capacity on the previous table and someone sends some samples from that time.

It’s also bad if there are a lot of timeseries that map to the same index hash (i.e. same metric name for same user on same day), because that creates a hot-spot in the data store.

It would be better to have separate queues for each table, and to send a spread of data which is known to have different hash keys. Even better if you can batch it up to reduce network overhead to the back end.