MongoDB Schema Design (Richard Kreuter's Mongo Berlin preso)

Schema Design — MongoBerlin

Richard M Kreuter
10gen Inc.
richard@10gen.com

March 25, 2011


Observations about Relational Database Schemas

Relational schema design is often presented and thought of as
an exercise in normalization. While academics debate how
many normal forms can ﬁt on the head of a pin, practitioners
tend to employ just one or two.
However, all nontrivial real-world applications employ a variety
of strategic denormalizations: materialized views in the
RDBMS, caching layers outside the RDBMS. These
denormalizations tend to be vital to real-world performance.
Finally, application programmers seldom code in relations, but
rather in object graphs; the RDBMS’s model, the set of
tuples, isn’t a great ﬁt for modern programming languages or
developers’ minds.


MongoDB Documents, Queries, Features

MongoDB documents are deeply nestable sequences key-value
pairs, thus permitting “rich” structure.
The MongoDB query language is relatively SQL-like in its
capacity to find documents satisfying complicated, dynamic
criteria.
MongoDB documents can be updated atomically, with
special eﬃciency at updates that don’t alter a document’s size
or shape.


MongoDB Schema Design Generalities

When designing for MongoDB, do...
... let the application direct the schema.
... denormalize judiciously.
... design your schema for indexing.
... resort to application-level JOINs when needed
And don’t ...
... treat collections as heaps.
... frequently resize documents.


Letting the application direct the schema

Most applications mostly view their data in a small number of,
distinguished “shape”, generally congruent to graphs of
inter-object has-a relationships among instance classes in the
applications’ models. MongoDB lets you store your data more or
less directly according to the shape of your model.


Letting the application direct the schema, continued

db.blog_posts.findOne()
{ _id : Object(...)
text : "A blazingly clever blog post.",
by : "A. U. Thor",
date : "Mon Mar 21 2011 03:54:51 GMT-0400 (EDT)",
tags : [ "funny", "ironic" ]
}


Denormalizing Judiciously

Most application entities turn out to have some ﬁelds that are very
frequently altered, and other ﬁelds that are exceedingly seldom
altered. Embedding infrequently altered attributes around the
database is a reasonable strategy to improve performance.


Denormalizing Judiciously, continued

db.product_reviews.findOne()
{ _id : Object(...)
comment : "The best thing ever!"
date : "Mon Mar 21 2011 03:54:51 GMT-0400 (EDT)",
reviewer : { uid : ObjectId("987654abcxyz"),
name : "Khan Sumer",
thumbnail : "thumb-123456.jpg",
url : "http://blahblah.com/" } }
db.users.find({ _id : ObjectId("987654abcxyz")})
{ uid : ObjectId("987654abcxyz"),
name : "Khan Sumer",
thumbnail : ..., url : ...
last_post : "Mon Mar 21 2011 03:54:51 GMT-0400 (EDT)",
favorites : [ ... ], friends : [ ... ] }
}

Design your schema for indexing

There’s a subtle relationship between schemas and indexes.
Consider this query:

db.boxes.find({$where : "this.height > this.width"})

This query doesn’t take advantage of MongoDB indexes, both
because of the JavaScript and also because this predicate isn’t
something MongoDB knows how to index. If this sort of query is
important, maintaining a separate boolean attribute in the
document is the right thing; and the separate value can be indexed.


Application-level JOINs

Because most MongoDB documents are “richer” than RDBMS
rows, they tend to represent “pre-JOINed” data; and so
application-level JOIN operations should be few. However,
sometimes you do need relational-style normalization and
application-level JOINS. This comes up in some many-to-many
relationships, and may not cost much in practice.


Don’t treat collections as heaps

Although MongoDB permits quite a bit of freedom in document
structure, documents in a collection ought to share a common
subset of attributes, for programmatic processing eﬀective
indexing, and developer comprehension. If you have documents
with very diﬀerent sets of attributes, consider storing them in
separate collections.


Don’t frequently resize documents

Resizing a document (e.g. by adding/removing attributes or
adding/removing elements of lists) is generally costly. (In-place
updates are quite eﬃcient, however.) In general, a schema whose
documents’ sizes are highly volatile should be considered suspect;
such data might best be stored as separate documents.


Don’t frequently resize documents, continued
So, instead of this
db.urlhits.findOne()
{ _id : ..., url : "http://10gen.com",
// this is counting with granularity of 1 day
counts : { "2011-03-01" :
{ firefox : 12345, chrome : 23456 },
"2011-03-02" :
{ firefox : 15678, chrome : 24567 }
... } }
consider this:
db.urlhits2.findOne()
{ _id : ..., url : "http://10gen.com",
date : "2011-03-01",
counts : { "firefox : 12345, chrome : 23456 } }

Don’t frequently resize documents, continued

So, instead of this

db.user_events.findOne()
{ _id : ..., user : "kreuter"
clicks : [ { url : <url1>, time : <time1> },
{ url : <url2>, time : <time2> },
... ] }

consider this:

db.user_events.findOne()
{ _id : ..., user : "kreuter", url: <url1>, time: <time1> }


Going forward

www.mongodb.org — downloads, docs, community
mongodb-user@googlegroups.com — mailing list
#mongodb on irc.freenode.net
try.mongodb.org — web-based shell
10gen is hiring. Email jobs@10gen.com.
10gen oﬀers support, training, and advising services for
mongodb


MongoDB Schema Design (Richard Kreuter's Mongo Berlin preso)

More Related Content

What's hot (20)

Similar to MongoDB Schema Design (Richard Kreuter's Mongo Berlin preso) (20)

More from MongoDB (20)

Recently uploaded (20)

MongoDB Schema Design (Richard Kreuter's Mongo Berlin preso)