Druid doesn't emit any metric to tell time taken before query fanout to Data nodes

[ravisharm]

Currently, at Broker, while Druid does emit ‘query/time’ metric that captures overall time taken to complete the query, it doesn’t emit any Broker metric that captures time taken to plan the query before the query gets sent to the Data nodes. This is essentially the time spent to determine segments to query, grouping segments by server etc. When query/time shows a high value, OnCall needs to understand the reason for the high latency. The only relevant metric is query/node/time. If it shows normal values, the natural question becomes how much time did broker take to process the query. Broker spends some CPU cycles before fanning out the query (the subject of the proposed metric) and later spends some CPU cycles to do the final aggregation. Using the proposed metric we can pinpoint if fanout took longer and further investigate that part of the query lifecycle if needed.