Brokers to catch repetitive queries

[leventov]

Description

Broker registers queries being executed and associates them with Future objects. When it notices that a query has arrived that is already being executed, it just puts the HTTP thread to await to the same (already existing) Future, instead of sparking a duplicate execution.

This can be seen as a form of results "pre-cache".

Motivation

When a query executes for a long time, users get impatient and reload a web interface, initiating the same queries to Druid. Currently, by doing so they only make things worse, because brokers (and historicals), still crunching results for the previous query (if no form of cancellation protocol is used), start another instance of the query.

[gianm]

A wrinkle to consider for anyone implementing this: some queries involve streaming results back, and the Broker does not buffer them. Meaning that it could be in the middle of running a query but unable to allow new requests to jump on board (since some portion of the results have already been returned and forgotten).

[leventov]

The industry term for this idea appears to be "request coalescing" or "collapsing" (posting so that somebody could find this issue via search).