I'm not entirely sure if this is being considered or not, but it does seem that this is less of a static configuration if I'm understanding it correctly.

ES seems to allow you to write a Date value in a variety of formats. What I'm hoping for is that the interfaces between ES and processing frameworks like Hadoop and Spark will dynamically deserialize into the correct data type given the format it is stored in.

So if I have a mapping like:

"created_at": {
    "type": "date"
},
"updated_at": {
    "type": "date",
    "format": "yyyy-MM-dd"
}

and I index a long UTC millis value to created_at and a human readable date string to updated_at I would expect to get the same formats back during the deserialization process.

The issue has been addressed in master, the docs updated and the nightly builds published.
In short, by default, date objects are returned as rich objects (they vary depending on the library used). To get the fields as primitives (long or String depending on how they are stored), one should use es.mapping.date.rich field and configure it to false.

Closing the issue; if the problem persists please create a new one.