This blog is a follow up on our previous blog “Snapshots for HDFS”

In June we had posted an early prototype of snapshots that allowed us to experiment with a few ideas in HDFS-2802. Since then we have added more details to the design document and made significant progress on a brand new implementation (over 40 subtasks in HDFS-2802).

Some of the highlights of this new design include:

• Read-Only Copy-on-Write (COW) snapshots (but can be extended RW later)
• Snapshots for entire namespace or sub directories
• Snapshots are managed by Admin, but users are allowed to take snapshots
• Snapshots are efficient
• Creation is instantaneous with O(1) cost.
• Additional memory is used only when modifications are made relative to  a snapshots (memory usage is O(M), where M is the number of modified files/directories)
• Snapshots do not adversely affect regular HDFS operations

An initial implementation of snapshots with some tests is already completed.  We are now working on the improvements, some new tools for snapshots and adding more tests.

The major work-in-progress items are:

• Persistent data structure based solution for efficient creation and memory usage
• Snapshot diff tool
• Restore/rollback snapshots

Meetup at Hortonworks: rallying the community
Recently we also held a Meetup at Hortonworks office, where over 30 folks attended to discuss the design and some of the features in great detail. A wide range of topics were discussed from Snapshot usage by HBase, administration aspects of snapshots, overhead of creating and maintaining snapshots, and lower level details such as length of files that are open-for-writing. We had  representation from HDFS developers, HBase developers and engineers with deep experience in managing hadoop and other storage systems.   We thank the community for the valuable discussion and feedback on the feature requirements and the open questions.

Motivation

Apache Hadoop provides a high performance native protocol for accessing HDFS. While this is great for Hadoop applications running inside a Hadoop cluster, users often want to connect to HDFS from the outside. For examples, some applications have to load data in and out of the cluster, or to interact with the data stored in HDFS from the outside. Of course they can do this using the native HDFS protocol but that means installing Hadoop and a Java binding with those applications. To address this we have developed an additional protocol to access HDFS using an industry standard RESTful mechanism, called WebHDFS. As part of this, WebHDFS takes advantages of the parallelism that a Hadoop cluster offers. Further, WebHDFS retains the security that the native Hadoop protocol offers. It also fits well into the overall strategy of providing web services access to all Hadoop components.

Read More