My approach to this is to take the complicated bits of the mapreduce and put them in a separate class. Then I do a combination of two things as appropriate:
1. Hook it up to a debug server that fetches from the same datastore as the mapreduce, then test it on some keys that I'm interested in.
2. Test it like any other class.
The only awkward part of this is abstracting out the output calls, which I usually do by passing in a "handle some data" callback that outputs in the mapreduce and dumps some pretty html in the debug server.
The great part about this is that if the mapreduce ends up being something important, you already have the tools to introspect its internals on data you are interested in.
1. Hook it up to a debug server that fetches from the same datastore as the mapreduce, then test it on some keys that I'm interested in.
2. Test it like any other class.
The only awkward part of this is abstracting out the output calls, which I usually do by passing in a "handle some data" callback that outputs in the mapreduce and dumps some pretty html in the debug server.
The great part about this is that if the mapreduce ends up being something important, you already have the tools to introspect its internals on data you are interested in.