Pull scheduling into separate process

[savagematt]

That is:

Our current approach to scheduling jobs relies on agents polling the database at sufficient frequency for the job to be triggered near its due date. This doesn't really scale.

We could move the polling on to the server, but that's also less than ideal.

A better solution:

• A process on each server keeps track of what the next due job is, initially from a snapshot
• The server notices a scheduled job being created and updates the next due job
• The process then sleeps until either the next due date arrives, or is awoken when a bew scheduled job arrives
• The process generates events into the same stream as newly created jobs with no due date on, which are easier to process because there is a natural stream of updates from the database