Friday, September 20, 2013

Cassandra Vs HBase : Which NoSql store do I need ?

There many NoSql databases out there and it can be confusing to determine which one is suitable for a particular use case. In this blog, we discuss the two more popular ones, Cassandra and HBase. If you are new to NoSql , you may review these earlier posts:

What is NoSql ?
HBase architecture

1. History

To understand the goals and motivation behind any product, it is a good idea to trace its origin.

HBase is based on Google's Bigtable as described in the paper "Bigtable: A Distributed Storage System for Structured Data". You only need to read the first line of the abstract to understand what BigTable attempts to do. It is a distributed storage system for managing structured data that can scale to very large size of the order of petabytes using thousands of commodity servers.

Cassandra derives its motivation from Amazon's Dynamo as described in the paper "Dynamo: Amazon's highly available key value store". Reading the first page of this paper it is clear that the primary goals were reliability at scale and high availability for amazon's online store.

While both papers talk about scale, reliability and availability, the primary problem BigTable (and HBase) is addressing is  random access to data of the scale of  hundreds of terrabytes or petabytes  and the primary problem that Dynamo ( and Cassandra) addresses  is high availability.

2. Stack

HBase is based on Hadoop  and depends on HDFS, namenodes, datanodes, zookeeper etc. If you are already familiar and invested in Hadoop, this is not a big issue. Cassandra has no such dependency. If you are not invested in Hadoop, then Cassandra has less moving parts and could be easier to go with.

3. Data Model

HBase data model has table, rows, column families and columns. Table has rows and column families. Each row has a key. A column family has columns. When a table is created, the column families have to be defined. But the columns in a column family do not have to be defined and can be added ad hoc.

Cassandra data model used to be described in a similar manner. More recently Cassandra has adopted CQL which is like SQL, as the client interface to data. The data model consists of  tables, rows and columns. You have to define Tables and their columns. Additional columns can be added ad hoc. The data is partitioned based on row key.

Neither supports joins on tables. This is all about denormalization.

HBase does not support secondary indexes. Cassandra supports indexes on columns by denormalization and creating additional tables under the hood.

4. Consistency

HBase supports strong consistency for reads and writes at a row level. Out of the box Cassandra is eventually consistent, though it can be configured for strong consistency. Eventual consistency means that a read may not see the latest update by a write. When that happens , the application is expected to deal with it in an appropriate manner.

In Cassandra, client might connect and write to any node. The write is also replicated to multiple nodes. At any time, not all nodes might have that latest value replicated. Multiple writes that conflict can happen at the same. These might need to be reconciled.

5. Range queries on row keys  

Partioning of rows in HBase is ordered. This implies one can do range queries by row key in HBase. In Cassandra the default partitioner does not keep ordering and you cannot do range queries. There is an ordered partitioner but the recommendation is to never use it in production because it can cause the cluster to become unbalanced. This is a pretty important difference. If your use cases require range queries as would be the case in data warehousing type applications, then you will need to go with HBase.

5. High Availability and Partition tolerance

Cassandra is built ground up for availability and partition tolerance while sacrificing on consistency. All nodes are equal. Data is replicated across nodes. Client can connect to any node and read or write any key.

Early versions of  HBase had very obvious single point of failures in the HDFS name node and the HBASE master. Early versions required external HA solutions (Eg. Linux HA project)  in an Active - Passive configuration. For example you would have a stand by Name node , shared storage and monitoring of the active name node. When monitoring detects that the active node is down, the HA solution would fail over to the stand by node. Recently Hadoop 2 has added native support for hot failover (active active) for the namenode. HBASE has added hot failover support for HBASE master as well. But clearly these came after the fact.

If high availability is a primary use case, then Cassandra clearly wins.

6. Storage

HBase stores data on HDFS. Cassandra stores data locally on each node.

The primary storage data structure for both is based on Google's Bigtable and is the Log Structured merge tree. Your typical RDBMS uses the B+ tree as the primary data structure for creating and storing indexes. In a B+ tree, each node is a disk block and to find a key or to write a key, the code has to read and write  a fews blocks from disk. The log structured merge tree has lesser disk access than a B+tree. LSM has 2 tree like structures one is memory and one on disk. Initially reads and writes happen to the tree in memory. When the in memory tree reaches a certain size, it is written to disk, where it is merged with the larger tree.

In both HBase and Cassandra, users have complained that the merge process mentioned above ( also known as compaction) slows things down.

7. Summary

HBase and Cassandra are suitable for different use cases.

If you have data warehousing type of use cases and large amounts the data that will continue to grow, HBase is a more suitable choice. If you are thinking data from large application log files, several hundred terrabytes or petabytes of time series data or monitoring type data that needs to be aggregated or mined, then HBase might be appropriate. Recently I spoke to a cloud based application monitoring company that was storing large amounts of monitoring data generated every minute in MySql and were complaining about not being able to scale. They need HBase.

Cassandra on the other hand is more suitable if you needs to serve traffic from a highly available data store. If you are already on an RDBMS but are have availability or scale issues, then Cassandra might be something to look at. With CQL, which is like SQL, Cassandra is positioning itself as an alternative to your typical RDBMS. A company that I know uses a key value store that is fast, but does not scale nor does it have built support for high availability. Very often, their shards go down and they lose data. They should consider Cassandra.

Related Blogs:

Cassandra data model
Choosing Cassandra 
Cassandra Compaction