Thanks Alex for your swift reply!
I think I should have given more details about the use case.
The idea is to store a lot of small documents for down the line NLP applications. Each document would have a few fields like "client", "language", "origin", etc. as well as the core text data.
These fields would allow to retrieve data and build corpora based on different filters on the different fields.
One of the fields is then related to the "domains", which can take a relatively fixed number of values (let's say two dozens).
There may be some new ones along the way but it should really not explode.
Finally, these domains are not mutually exclusive, hence the need to count for each doc where it stands.
So, as I would have few domains, but tens of millions of documents, I don't really see the benefit of inverting the structure. Not sure if any other structure would be interesting.
Anyways, as the domains are supposed to be pretty fixed, I can always do the aggregation per domain. Just thought it would be nice to have an existing entry point for such cases!
( + it would be probably easier to integrate in some Kibana visualisations)