Polyglot persistence which takes a hybrid approach to persistence is getting a lot of traction these days. Currently I am reading NoSQL Distilled: A Brief Guide to the Emerging World of Polyglot Persistence by Pramod J. Sadalage and Martin Fowler. Basic idea behind polyglot persistence is to use specialized databases (both NoSQL and Relational) for different purposes within the same web application. One-database-fits-all is no more the de facto choice for web applications. In fact polyglot persistence is about choice – choice to select best data store for a given data type or purpose. Polyglot persistence leverages the strength of multiple data stores.
Multiple data storesIn a polyglot persistence pattern, your web application can use one or more relational databases or combination of relational and NoSQL data stores. Most of modern web frameworks such as Django, Rails etc already support multiple databases. As an example, a polyglot persistence enabled eCommerce application can store
- Session data in Redis or Memcached. Session data requires faster read and write but not durability. For better durability we can always use write-through.
- Transactional data (order, payment and account) in a ACID compliant traditional RDBMS stores such as MySQL or Oracle. In addition to that, HBase/Hive with Hadoop can be used to process transaction level data such as order history for market basket analysis.
- Shopping cart data in a high availability and fault-tolerance data store such as Riak or Cassandra. Riak will be a better choice because it is a key/value store with excellent query API with primary key operations such GET, PUT, DELETE, UPDATE.
- Log level data (audit and activity) in a very high write throughput data store such as Cassandra. This is also good for analytic and real-time data mining such as product ranking,
- Data for product recommendations, related products and similar products in a graph database such as Neo4j. Also possible to use Apache Giraph as distributed graph processing infrastructure which runs on Hadoop
- Product catalogue in a document oriented data store such as CouchDB or MongoDB. Requirement include high read throughput, frequent data change (stock level information) which makes MongoDB a better choice.
- Customer profile data including purchase history, shipping & billing address – again in a document oriented data store. CouchDB is better option because this data change occasionally.