Apache Hadoop turned ten this year. To celebrate, Karthik and I gave a talk at USENIX ATC '16 about open problems to solve in Hadoop's second decade. This was an opportunity to revisit our academic roots and get a new crop of graduate students interested in the real distributed systems problems we're trying to solve in industry.

This is a huge topic and we only had a 25 minute talk slot, so we were pitching problems rather than solutions. However, we did have some ideas in our back pocket, and the hallway track and birds-of-a-feather we hosted afterwards led to a lot of good discussion.

Karthik and I split up the content thematically, which worked really well. I covered scalability, meaning sharded filesystems and federated resource management. Karthik addressed scheduling (unifying batch jobs and long-running services) and utilization (overprovisioning, preemption, isolation).

I'm hoping to give this talk again in longer form, since I'm proud of the content.

Slides: pptx

USENIX site with PDF slides and audio

Read More    |    Comments

I gave a presentation titled Happier Developers and Happier Software through Distributed Testing at Apache Big Data 2016, which detailed how our distributed unit testing framework has decreased the runtime of Apache Hadoop's unit test suite by 60x from 8.5 hours to about 8 minutes, and the substantial productivity improvements that are possible when developers can easily run and interact with the test suite.

The infrastructure is general enough to accommodate any software project. We wrote frontends for both C++/gtest and Java/Maven.

This effort started as a Cloudera hackathon project that Todd and I worked on two years ago, and I'm very glad we got it across the line. Furthermore, it's also open-source, and we'd love to see it rolled out to more projects.

Slides: pptx

Source-code: cloudera/dist_test

Read More    |    Comments

What makes this paper special is that it is one of the only published papers about a production cloud blobstore. The 800-pound gorilla in this space is Amazon S3, but I find Windows Azure Storage (WAS) the more interesting system since it provides strong consistency, additional features like append, and serves as the backend for not just WAS Blobs, but also WAS Tables (structured data access) and WAS Queues (message delivery). It also occupies a different design point than hash-partitioned blobstores like Swift and Rados.

This paper, "Windows Azure Storage: A Highly Available Cloud Storage Service with Strong Consistency" by Calder et al., was published at SOSP '11.

Read More    |    Comments

Google recently unveiled one of their crown jewels of system infrastructure: Borg, their cluster scheduler. This prompted me to re-read the Mesos and Omega papers, which deal with the same topic. I thought it'd be interested to do a compare and contrast of these systems. Mesos gets credit for the groundbreaking idea of two-level scheduling, Omega improved upon this with an analogy from databases, and Borg can sort of be seen as the culmination of all these ideas.

Read More    |    Comments