Reconsider the cap on chunk age.

[tomwilkie]

Perhaps instead we should have a cap on total number of in memory chunks?

Problem is: dynamodb provisioned capacity has to be > #series / chunk age * number of writes per chunk.

For example, 800k series / 1 hour * 8 writes per chunk = 1.7k writes / s to dynamodb.

We should probably only flush chunks:

• when they are full
• when they haven't been written to in some time period
• some (much higher - 24hrs) max age
• we're running out of memory (maybe limit to 1million chunks per ingester?)

This will allow the very efficient chunks (up) to live for much longer,

[tomwilkie]

This is basically done, other than some static limit on total number of chunks.