HBase vs Cassandra

Apache HBase vs Apache Cassandra

This comparative study was done by me and Larry Thomas in May, 2012. Cassandra stuff was prepared by Larry Thomas.
This information is NOT intended to be a tutorial for either Apache Cassandra orApache HBase. We tried our level best to provide the most accurate information. Please comment or email me if you find any corrections. I would be happy to maintain this list with the most accurate and updated information.

Point	HBase	Cassandra
Foundations	HBase is based on BigTable (Google)	Cassandra is based on DynamoDB (Amazon).   Initially developed at Facebook by former Amazon engineers.  This is one reason why Cassandra supports multi data center.  Rackspace is a big contributor to Cassandra due to multi data center support.
Infrastructure	HBase uses the Hadoop Infrastructure (Zookeeper, NameNode, HDFS).  Organizations that will deploy Hadoop anyway may be comfortable with leveraging Hadoop knowledge by using HBase	Cassandra started and evolved separate from Hadoop and its infrastructure and Operational knowledge requirements are different than Hadoop.  However, for analytics, many Cassandra deployments use Cassandra + Storm (which uses Zookeeper), and/or Cassandra + Hadoop.
Infrastructure Simplicity and SPOF	The HBase-Hadoop Infrastructure has several "moving parts" consisting of Zookeeper,  Name Node,  Hbase Master, and Data Nodes,   Zookeeper is clustered and naturally fault tolerant.  Name Node needs to be clustered to be fault tolerant.	Cassandra uses a a single Node-type.  All nodes are equal and perform all functions.   Any Node can act as a coordinator, ensuring no SPOF.   Adding Storm or Hadoop, of course, adds complexity to the infrastructure.
Read Intensive Use Cases	HBase is optimized for reads, supported by single-write master, and resulting strict consistency model, as well as use of Ordered Partitioning which supports row-scans.  HBase is well suited for doing Range based scans.	Cassandra has excellent single-row read performance as long as eventual consistency semantics are sufficient for the use-case.  Cassandra quorum reads, which are required for strict consistency will naturally be slower than Hbase reads.  Cassandra does not support Range based row-scans which may be limiting in certain use-cases.  Cassandra is well suited for supporting single-row queries,  or selecting multiple rows based on a  Column-Value index.
Multi-Data Center Support and Disaster Recovery	HBase provides for asynchronous replication of an HBase Cluster across a WAN.   HBase clusters cannot be set up to achieve zero RPO, but in steady-state HBase should be roughly failover-equivalent  to any other DBMS that relies on asynchronous replication over a WAN.   Fall-back processes and procedures (e.g. after failover) are TBD.	Cassandra Random Partitioning provides for row-replication of a single row across a WAN, either asynchronous (write.ONE,  write.LOCAL_QUORUM),  or synchronous (write.QUORUM,  write.ALL).  Cassandra clusters can therefore be set up to achieve zero RPO, but each write will require at least one wan-ACK back to the coordinator to achieve this capability.
Write.ONE Durability	Writes are replicated in a pipeline fashion: the first-data-node for the region persists the write, and then sends the write to the next Natural Endpoint, and so-on in a pipeline fashion. HBase’s commit log "acks" a write only after *all* of the nodes in the pipeline have written the data to their OS buffers.  The first Region Server in the pipeline must also have persisted the write to its WAL.	Cassandra's coordinators will send parallel write-requests to all Natural Endpoints, The coordinator will "ack" the write after exactly one Natural Endpoint has "acked" the write, which means that node has also persisted the write to its WAL.   The writes may or may not have committed to any other Natural Endpoint.
Ordered Partitioning	HBase only supports Ordered Partitoning.  This means that Rows for a CF are stored in RowKey order in HFiles, where each Hfile contains a "block" or "shard" of all the rows in a CF.  HFiles are distributed across all data-nodes in the Cluster	Cassandra officially supports Ordered Partitioning, but no production user of Cassandra uses Ordered Partitioning due to the "hot spots" it creates and the operational difficulties such hot-spots cause.  Random Partitioning is the only recommended Cassandra partitioning scheme, and rows are distributed across all nodes in the cluster.
RowKey Range Scans	Because of ordered partitioning,  HBase queries can be formulated with partial start and end row-keys, and can locate rows inclusive-of, or exclusive of these partial-rowkeys.  The start and end row-keys in a range-scan need not even exist in Hbase.	Because of random partitioning,  partial rowkeys cannot be used with Cassandra.  RowKeys must be known exactly.  Counting rows in a CF is complicated.   It is highly recommended that for these types of use-cases,  data should be stored in columns in Cassandra, not in rows.
Linear Scalability for large tables and range scans	Due to Ordered Partitioning, HBase will easily scale horizontally while still supporting rowkey range scans.	If data is stored in columns in Cassandra to support range scans, the practical limitation of a row size in Cassandra is 10's of Megabytes.  Rows larger than that causes problems with compaction overhead and time.
Atomic Compare and Set	HBase supports Atomic Compare and Set. HBase supports supports transaction within a Row.	Cassandra does not support Atomic Compare and Set.   Counters require dedicated counter column-families which because of eventual-consistency requires that all replicas in all natural end-points be read and updated with ACK.  However, hinted-handoff mechanisms can make even these built-in counters suspect for accuracy.   FIFO queues are difficult (if not impossible) to implement with Cassandra.
Read Load Balancing - single Row	Hbase does not support Read Load Balancing against a single row.  A single row is served by exactly one region server at a time.  Other replicas are used ony in case of a node failure.  Scalability is primarily supported by Partitioning which statistically distributes reads of different rows across multiple data nodes.	Cassandra will support Read Load Balancing against a single row.  However,  this is primarily supported by Read.ONE, and eventual consistency must be taken into consideration.  Scalability is primarily supported by Partitioning which distributes reads of different rows across multiple data nodes.
Bloom Filters	Bloom Filters can be used in HBase as another form of Indexing.  They work on the basis of RowKey or RowKey+ColumnName to reduce the number of data-blocks that HBase has to read to satisfy a query.  (Bloom Filters may exhibit false-positives (reading too much data), but never false negatives (reading not enough data).	Cassandra uses bloom filters for key lookup.
Triggers	Triggers are supported by the CoProcessor capability in HBase.  They allow HBase to observe the get/put/delete events on a table (CF), and then execute the trigger-logic.    Triggers are coded as java classes.	Cassandra does not support co-processor-like functionality (as far as we know)
Secondary Indexes	Hbase does not natively support secondary indexes, but one use-case of Triggers is that a trigger on a "put" can automatically keep a secondary index up-to-date, and therefore not put the burden on the application (client).	Cassandra supports secondary indexes on column families where the column name is known.  (Not on dynamic columns).
Simple Aggregation	Hbase CoProcessors support out-of-the-box simple aggregations in HBase.   SUM, MIN, MAX, AVG,  STD.   Other aggregations can be built by defining java-classes to perform the aggregation	Aggregations in Cassandra are not supported by the Cassandra nodes - client must provide aggregations.  When the aggregation requirement spans multiple rows, Random Partitioning makes aggregations very difficult for the client.   Recommendation is to use Storm or Hadoop for aggregations.
HIVE Integration	HIVE can access HBase tables directly (uses de-serialization under the hood that is aware of the HBase file format).	Work in Progress (https://issues.apache.org/jira/browse/CASSANDRA-4131)
PIG Integration	PIG has native support for writing into/reading from HBase.	Cassandra 0.7.4+

Point

HBase

Cassandra

CAP Theorem Focus	Consistency, Availability	Availability, Partition-Tolerance
Consistency	Strong	Eventual (Strong is Optional)
Single Write Master	Yes	No (R+W+1 to get Strong Consistency)
Optimized For	Reads	Writes
Main Data Structure	CF, RowKey,  Name Value Pair Set	CF, RowKey, Name Value Pair Set
Dynamic Columns	Yes	Yes
Column Names as Data	Yes	Yes
Static Columns	No	Yes
RowKey Slices	Yes	No
Static Column Value Indexes	No	Yes
Sorted Column Names	Yes	Yes
Cell Versioning Support	Yes	No
Bloom Filters	Yes	Yes(only on Key)
CoProcessors	Yes	No
Triggers	Yes(Part of Coprocessor)	No
Push Down Predicates	Yes(Part of Coprocessor)	No
Atomic Compare and Set	Yes	No
Explicit Row Locks	Yes	No
Row Key Caching	Yes	Yes
Partitioning Strategy	Ordered Partitioning	Random Partitioning recommended
Rebalancing	Automatic	Not Needed with Random Partitioning
Availability	N-Replicas across Nodes	N-Replicas across Nodes
Data Node Failure	Graceful Degredation	Graceful Degredation
Data Node Failure - Replication	N-Replicas Preserved	(N-1) Replicas Preserved + Hinted Handoff
Data Node Restoration	Same as Node Addition	Requires Node Repair Admin-action
Data Node Addition	Rebalancing Automatic	Rebalancing Requires Token-Assignment Adjustment
Data Node Management	Simple (Roll In, Role Out)	Human Admin Action Required
Cluster Admin Nodes	Zookeeper, NameNode, HMaster	All Nodes are Equal
SPOF	Now, all the Admin Nodes are Fault Tolerant	All Nodes are Equal
Write.ANY	No, but Replicas are Node Agnostic	Yes (Writes Never Fail if this option is used)
Write.ONE	Standard, HA, Strong Consistency	Yes (often used), HA,  Weak Consistency
Write.QUORUM	No (not required)	Yes (often used with Read.QUORUM for Strong Consistency
Write.ALL	Yes (performance penalty)	Yes (performance penalty, not HA)
Asynchronous WAN Replication	Yes, but it needs testing on corner cases.	Yes (Replica's can span data centers)
Synchronous WAN Replication	No	Yes with Write.QUORUM or Write.EACH-QUORUM
Compression Support	Yes	Yes