Presentation on how to use document databases in software architecture, revised for code mash 2016

1. Got Documents?
AN EXPLORATION OF DOCUMENT DATABASES IN SOFTWARE
ARCHITECTURE

2. About Me
www.maggiepint.com
maggiepint@gmail.com
@maggiepinthttps://www.tempworks.com

4. Flavors
MONGODB, COUCHDB, RAVENDB, AND MORE

5. MongoDB
•Dominant player in document databases
•Runs on nearly all platforms
•Strongly Consistent in default configuration
•Indexes are similar to traditional SQL indexes in nature
•Stores data in customized Binary JSON (BSON) format that allows typing
•Limit support for cross-collection querying in latest release
•Client API’s available in tons of languages
•Must use a third party provider like SOLR for advanced search capabilities

6. CouchDB
•Stores documents in plain JSON format
•Eventually consistent
•Indexes are map-reduce and defined in Javascript
•Clients in many languages
•Runs on Linux, OSX and Windows
•CouchDB-Lucene provides a Lucene integration for search

7. RavenDB
•Stores documents in plain JSON format
•Eventually consistent
•Indexes are built on Lucene. Lucene search is native to RavenDB.
•Server only runs on Windows
•.NET, Java, and HTTP Clients
•Limited support for cross-collection querying

8. Other Players
•Azure DocumentDB
• Very new product from Microsoft
•ReactDB
• Open source project that integrates push notifications into the database
•Cloudant
• IBM proprietary implementation of CouchDB
•DynamoDB
• Mixed model key value and document database

9. Architectural
Considerations

10. How do document databases work?
•Stores related data in a single document
•Usually uses JSON format for documents
•Enables the storage of complex object graphs together, instead of normalizing data out into
tables
•Stores documents in collections of the same type
•Allows querying within collections
•Does not typically allow querying across collections
•Offers high availability at the cost of consistency

11. Consideration: Schema Free
PROS
Easy to add properties
Simple migrations
Tolerant of differing data
CONS
Have to account for properties being missing

12. ACID
Atomicity
◦ Each transaction is all or nothing
Consistency
◦ Any transaction brings the database from one valid state to another
Isolation
◦ System ensures that transactions operated concurrently bring the database to the same state as if they
had been operated serially
Durability
◦ Once a transaction is committed, it remains so even in the event of power loss, etc

13. ACID in Document Databases
•Traditional transaction support is not available in any document database
•Document databases do support something like transactions within the scope of a document
•This makes document databases generally inappropriate for a wide variety of applications

14. Consideration: Non-Acid
PROS
Performance Gain
CONS
No way to guarantee that operations across
succeed or fail together
No isolation when sharded
Various implementation specific issues

15. Case Study: Survey
System

16. Requirements
•An administration area is used to define ‘Surveys’.
• Surveys have Questions
• Questions have answers
•Surveys can be administrated in sets called workflows
•When a survey changes, this change can only apply to surveys moving forward
• Because of this, each user must receive a survey ‘instance’ to track the version of the survey he/she got

17. A Traditional SQL Schema
•With various other requirements not described here, this schema came out to 83 tables
•For one of our heaviest usage clients, the average user would have 119 answers in the ‘Saved
Answer’ table
•With over 200,000 users after two years of use, the ‘Saved Answer’ table had 24,014,330 rows
•This table was both read and write heavy, so it was extremely difficult to define effective SQL
indexes
•The hardware cost for these SQL servers was astronomical
•This sucked

18. Designing Documents
•An aggregate is a collection of objects that can be treated as one
•An aggregate root is the object that contains all other objects inside of it
•When designing document schema, find your aggregates and create documents around them
•If you have an entity, it should be persisted as it’s own document because you will likely have to
store references to it

19. Survey System Design
•A combination SQL and Document DB design was used
•Survey Templates (one type of entity) were put into the SQL Database
•When a survey was assigned to a user as part of a workflow (another entity, and also an
aggregate), it’s data at that time was put into the document database
•The user’s responses were saved as part of the workflow document
•Reading a user’s application data became as simple as making one request for her workflow
document

20. Consideration: Models Aggregates Well
PROS
Improves performance by reducing lookups
Allows for easy persistence of object oriented
designs
CONS
none

21. Sharding
•Sharding is the practice of distributing data across multiple servers
•All major document database providers support sharding natively
•Document Databases are ideal for sharding because document data is self contained (less need
to worry about a query having to run on two servers)
•Sharding is usually accomplished by selecting a shard key for a collection, and allowing the
collection to be distributed to different nodes based on that key
•Tenant Id and geographic regions are typical choices for shard keys

22. Replication
•All major document database providers support replication
•In most replication setups, a primary node takes all write operations, and a secondary node
asynchronously replicates these write operations
•In the event of a failure of the primary, the secondary begins to take write operations
•MongoDB can be configured to allow reads from secondaries as a performance optimization,
resulting in eventual instead of strong consistency

23. Consideration: Scaling Out
PROS
Allows hardware to be scaled horizontally
Ensures very high availability
CONS
Consistency is sacrificed

24. Survey System: End Result
•Each user is associated with about 20 documents
•Documents are distributed across multiple databases using sharding
•Master/Master replication is used to ensure extremely high availability
•There have been no database performance issues in the year and a half the app has been in
production
•Because there is no schema migration concern, deploying updates has been drastically
simplified
•Hardware cost is reasonable (but not cheap)

26. Indexes
•All document databases support some form of indexing to improve query performance
•Some document databases do not allow querying without an index
•In general, you shouldn’t query without an index anyways

27. Consideration: Indexes
PROS
Improve performance of queries
CONS
Queries cannot reasonably be issued without
an index so indexes must frequently be
defined and deployed

29. Consideration: Eventual Consistency
PROS
Optimizes performance by allowing data
transfer to be a background process
CONS
Requires entire team to be aware of eventual
consistency implications

30. Case Study 2: CRM

31. CRM Requirements
•Track customers and basic information about them
•Track contacts and basic information about them
•Track sales deals and where they are in the pipeline
•Track orders generated from sales deals
•Track user tasks

32. Customers and Their Deals
•Customers and Deals are both entities, which is to say that they have distinct identity
•For this reason, Deals and Customer should be two separate collections
•There is no native support for cross-collection querying in most Document Databases
• The cross-collection querying support in RavenDB doesn’t perform well

33. Consideration: One document per
interaction
PROS
Improves performance
Encourages modeling aggregates well
CONS
Not actually achievable in most cases

34. Searching Deals by Customer Name
•The deal document must contain a denormalized customer object with the customer’s ID and
name
•We have a choice to make with this denormalization
• Allow the denormalization to just be wrong in the event the customer name is changed
• Maintain the denormalization when the customer name is changed

35. Denormalization Considerations
•Is stale data acceptable? This is the best option in all cases where it is possible.
•If stale data is unacceptable, how many documents are likely to need update when a change is
made? How often are changes going to be made?
•Using an event bus to move denormalization updates to a background process can be very
beneficial if failure of an update isn’t critical for the user to know

36. Consideration: Models Relationships
Poorly
PROS
None
CONS
Stale (out of date) data must be accepted in
the system
Large amounts of boilerplate code must be
written to maintain denormalizations
In certain circumstances a queuing/eventing
system is unavoidable

38. Consideration: Administration
PROS
Generally less involved than SQL
CONS
Server performance must be monitored
Hardware must be maintained
Index processes must be tuned
Settings must be tweaked

39. Consideration Recap
•Schema Free
•Non-Acid
•Models Aggregates Well
•Scales out well
•All queries must be indexed
•Eventual Consistency
•One document per interaction
•Models relationships poorly
•Requires administration

40. …nerds like us are allowed to be unironically
enthusiastic about stuff… Nerds are allowed to
love stuff, like jump-up-and-down-in-the-chair-
can’t-control-yourself love it.
-John Green