Couchbase views - just how far can I push this thing :) ... 10's or 100's of millions of index entries?

I'm wonder just how far I can push Couchbase views... what are the limits of the technology with regards to a realistic maximum number of design documents and views within a design document, and what is a realistic maximum number of elements and overall aggregate amount of data I can store reliably within a view index and still have reliable performance?

Are we talking millions, tens of millions or hundreds of millions of possible elements/nodes in a view index? How many views containing large numbers of elements are possible in a single bucket?

How far have Couchbase views been stress tested? What is the largest amount of data and number of elements seen in the field so far?

How far should I realistically push Couchbase views/indexes - am I pushing it too far to have a dozens of views in a single bucket with 50-100 million indexed elements/nodes each?

My use case:

I'm primarily interested in grouping related data through the use of collated views. I'm mostly only creating views that have null values, as I'm only interested in the collated keys and the doc._ids.

I'm currently prototyping a database that contains around 2 million entities and around 8 million 'facts' relating to these entities. The database is around 4GB on disk. I'm using collated views to relate/aggregate the entities and their facts. (See any recent presentations by Rich Hickey about value-oriented and fact-oriented programming and databases for the rationale). Essentially I'm creating a precalculated query dataset implemented via views, using map functions to crawl the database and generate and maintain the query dataset.

For example, a typical view is:
function (doc) {
var key;
if(doc.schemaType == "person"){
emit([doc._id, 0], null);
} else if(doc.schemaType == "personLink"){
key = "person:" + doc.fromPersonId;
emit([key, 1], null);
key = "person:" + doc.toPersonId;
emit([key, 2], null);
} else if(doc.schemaType == "personIdentityLink"){
key = "person:" + doc.personId;
emit([key, 3], null);
} else if(doc.schemaType == "personAddressLink"){
key = "person:" + doc.personId;
emit([key, 4], null);
} else if(doc.schemaType == "personSystemLink"){
key = "person:" + doc.personId;
emit([key, 5], null);
}
}

... a pretty standard collated view. There is a 'person' entity and lots of 'facts' about what this person is linked to. I have similar views, eg. collating on addresses and the entities linked to them.

On a single node 8 core computer this view intiallly took around 1.5 hours to index.

Querying this particular view reports 7 million 'total_rows', and so far pulling down results including documents using range queries only averages around few tens of milliseconds. The longest query for a person with a lot of facts and documents is around 70 milliseconds.

So far so good - scaling up and out is probably only going to improve this

My question is, once I increase my dataset, with a couple of dozen similar views, across maybe 20 million entites and 50-100 million associated facts, will Couchbase hack it?

BTW: Any tips to speed indexing up?

Cheers,

Tim Pedersen