When many people talk about object storage implementation, they focus on the software pieces like consistent hash, partitioning, replication, scale horizontally etc. But another critical component of a total object storage solution is the hardware. The operative word for object storage is still data and data lives on hard disk drives, but the hardware layer is also important for data availability in a total object storage solution.
There are two choices for the object store hardware layer: commodity hardware or storage array with built-in reliability, availability and serviceability (RAS). Commodity hardware is attractive for economic reasons, but storage array based solutions still matter.
Hard Disk Failures
Take hard disk drive failure as an example for challenges we face in the hardware layer. A typical object storage use case may have hundreds of petabytes (PB) of object data and hundreds of thousands of hard disk drives. Monitoring the health of such a great number of hard disk drives is a big challenge. Especially when the hard disk drive is the most error-prone component in modern computers. According to a hard drive failure analysis report from Google’s own data center, Annual Failure Rate (AFR) is ranging from 2% to 10% depending on the age of disk drives (see figure below). For more information, please see Google’s research paper titled “Failure Trends in a Large Disk Drive Population.”
Figure 1. Google Research: Annualized failure rates broken down by disk age groups
Take for example a scenario where you are building a 100PB object store with three replicas for each object. This would mean you would have at least 300,000 1T SATA drives. If AFR is 2%, it means you’ll have 6,000 failed disk drives per year and 16 failed drives per day. The number increases four to five times after disk age is greater than two years old!
So disk health monitoring and failed disk replacement will be a daily operation in any cloud service provider’s data center creating demand for a centralized management GUI and notification/alerts for failure disks.
Opportunity for Traditional Storage Arrays
Disk failure prediction is very important for data availability in enterprise class object store. Without disk failure prediction, think about putting all three object replicas to three individual drives that are all going to fail in next two days, but it is the weekend and your data center maintenance engineer could not arrive. You are putting yourself at the risk for data loss.
Mature storage arrays like EMC VNX also have other value-add features like FAST (a.k.a storage tiering) which could put hot data in Solid State Disk to improve I/O throughput.
EMC ViPR object storage is designed for enterprise class object store and runs on top of either commodity hardware, or traditional storage arrays like EMC VNX/Isilon, or a mixture of both.
We leave the hardware choice to our customers.
If you prefer a more cost effective solution, you could run ViPR object on commodity hardware. Or if you prioritize serviceability and availability, the combination of ViPR and an EMC VNX could be best choice for your business. Cloud service providers that want to compete on more than just price understand that the market will demand enterprise class object storage solutions that provide variable data durability and governance and compliance features.