geoinference | Geoinference predicts the location from which a piece | Machine Learning library

Note: This is cross-posted on the elasticsearch forum (https://discuss.elastic.co/t/store-size-1-000-times-the-document-byte-size/74258/4).

I am experiencing a roughly 1,000x increase in store.size over the document byte size. I've got a very simple mapping with very small documents (less than 1kb) and I've compared my mapping to Elasticsearch's internal mapping and they are the same, so it does not appear that there is any dynamic mapping going on.

So far, I have ingested 60,437 documents and have a store.size of 19.6Gb (average of 300kb per document), but the average byte size (String.getBytes().length) of the JSON is 300-400 bytes per document. In another run, the documents were averaging about 1MB - 3MB per document.

I'm using Elasticsearch 5.2 on an M4.2xlarge EC2 instance. Elasticsearch was installed with mostly all defaults, except what I needed to do in order to pass the boostrap checks and bind to a non-local IP. I've allocated 16GB (half of my physical memory) to Elasticsearch.

I used to run Elasticsearch 2.x and was ingesting FAR more fields and much larger documents than just these handful of fields and was only experiencing about 20k / document, which was still substantial, though manageable.

If anyone can point out anything that would fix this, I would appreciate it. Or is there an ES 5.x configuration I haven't seen that will resolve this?

Below is my mapping.