As part of coming to grips with CouchDB (and a particularly odious graph-theory problem) I've been looking around for design guidelines, hints and tips.
This MongoDB Schema Design document is quite helpful. The Link vs. Embed section clarifies the essential tradeoff here. In SQL world, link is the only tool. In this document-database world of CouchDB and MongoDB (as well as XML schema design) we have a link vs. embed decision.
Here is a presentation on trees (a specific kind of graph) in a document database: Trees in MongoDB. It enumerates a number of alternatives that are part of this new, larger design space for databases.
I found this because it was referenced in the myNoSQL blog, which seems to be a collection of sometimes useful links.
A September 2011 DAMA-NY meeting included a presentation on NoSQL Data Stores. It's findable on Google if you search for "dama nisql data stores" [sic; it is misspelled]. However, it's hard to link to directly because of the way Google obscures the target of their search. What's important in this presentation is the slightly defensive posture it takes about data modeling. It seems to describe ways that relational database modelers can cling to relevance in spite of threats represented by "NoSQL" databases.
The Transit System Problem
For a particularly gnarly problem, look at the Google Transit Feed Specification.
Then, look at Hampton Roads Transit on the GTFS Data Exchange.
How do we build a CouchDB document-centric view of this highly-normalized graph?
Route-centric? Each route has multiple trips. Each trip has a sequence of stop times. Do we repeat the stop definition over and over again? Seems silly, so perhaps it's Route - Trip - Stop-Time as a single document with links to Stop definitions.
Stop-centric? Each stop has multiple stop-times, and each stop has a parent route (based on trips along a route.) While this allows us to have a Stop document with a list of stop times and a (generally) single Route definition, it's not too useful.
We generally use transit based on the routes, not based on a single stop. So we need to query the stops based on a Route as well as based on a Stop Time. We may be able to use the CouchDB map definitions to provide some of these alternative views of a stop (i.e., by stop time, by route).
Some No SQL Lessons
What's really important here is that NoSQL schema design is not precisely the same as RDBMS schema design. In the RDMS world, with a single, fixed schema, proper up-front design is life-or-death. A great deal of design hand-wringing is required to get the relational model correct. In a good organization, this design effort involves prototyping, modeling and experimentation. In a bad organization, this design effort follows trivialized rules of thumb without too many second thoughts.
On the other hand, the No SQL schema design is essentially the same as RDBMS schema design.
In the NoSQL world, we still have to do prototyping, modeling and experimentation. We still have the three-tier separation between conceptual, logical and physical. Unlike the relational database, however, these tiers are more closely aligned in a document-oriented database. The conceptual tier is usually very, very close to the logical tier document structure. The conceptual gaps are filled by map-reduce views. The physical tier is just the logical tier document structure with some description of the sharding policies.
We do have to be more circumspect about committing to a design. In SQL world, DDL is a formal commitment to a design. DDL changes lead to breakage; making the dependencies more clear. In NoSQL world, there isn't the same depth of commitment. A technical spike which looks promising can lead to a gradual path of progressive dependence on the model.
The breakage that comes from schema change is more manageable but can spin out of control. It's more manageable because we can design our application around optional, missing and variant definitions of a document. It can become less manageable if we introduce too layers of useless abstraction to handle schema evolution.
The discipline of an ORM-like mapping between documents and Python classes is somewhat helpful for keeping the design focused around documents that have first-class meaning in the problem space. For that reason, couchdbkit seems useful.