Availability
Availability in the context of HBase can be defined as the ability of the system to handle failures. The most common failures cause one or more nodes in the HBase cluster to fall off the cluster and stop serving requests. This could be because of hardware on the node failing or the software acting up for some reason. Any such failure can be considered a network partition between that node and the rest of the cluster.
When a RegionServer becomes unreachable for some reason, the data it was serving needs to instead be served by some other RegionServer. HBase can do that and keep its availability high. But if there is a network partition and the HBase masters are separated from the cluster or the ZooKeepers are separated from the cluster, the slaves can’t do much on their own. This goes back to what we said earlier: availability is best defined by the kind of failures a system can handle and the kind it can’t. It isn’t a binary property, but instead one with various degrees.
Higher availability can be achieved through defensive deployment schemes. For instance, if you have multiple masters, keep them in different racks.
Reliability and durability
Reliability is a general term used in the context of a database system and can be thought of as a combination of data durability and performance guarantees in most cases. Data durability, as you can imagine, is important when you’re building applications atop a database. HBase has certain guarantees in terms of data durability by virtue of the system architecture.
HBase assumes two properties of the underlying storage that help it achieve the availability and reliability it offers to its clients.
Single namespace
HBase stores its data on a single file system. It assumes all the RegionServers have access to that file system across the entire cluster. The file system exposes a single namespace to all the RegionServers in the cluster. The data visible to and written by one RegionServer is available to all other RegionServers. This allows HBase to make availability guarantees. If a RegionServer goes down, any other RegionServer can read the data from the underlying file system and start serving the regions that the first RegionServer was serving.
At this point, you may be thinking that you could have a network-attached storage (NAS) that was mounted on all the servers and store the data on that. That’s theoretically doable, but there are implications to every design and implementation choice. Having a NAS that all the servers read/write to means your disk I/O will be bottlenecked by the interlink between the cluster and the NAS. You can have fat interlinks, but they will still limit the amount you can scale to. HBase made a design choice to use distributed file systems instead and was built tightly coupled with HDFS. HDFS provides HBase with a single namespace, and the DataNodes and RegionServers are collocated in most clusters. Collocating these two processes helps in that RegionServers can read and write to the local DataNode, thereby saving network I/O whenever possible. There is still network I/O, but this optimization reduces the costs.
Reliability and failure resistance
HBase assumes that the data it persists on the underlying storage system will be accessible even in the face of failures. If a server running the RegionServer goes down, other RegionServers should be able to take up the regions that were assigned to that server and begin serving requests. The assumption is that the server going down won’t cause data loss on the underlying storage. A distributed file system like HDFS achieves this property by replicating the data and keeping multiple copies of it. At the same time, the performance of the underlying storage should not be impacted greatly by the loss of a small percentage of its member servers.
Theoretically, HBase could run on top of any file system that provides these properties. But HBase is tightly coupled with HDFS and has been during the course of its development. Apart from being able to withstand failures, HDFS provides certain write semantics that HBase uses to provide durability guarantees for every byte you write to it.