Several weeks ago I read two interesting blog posts about the EBS failure. Jeff Darcy gave the developer’s view of failures in distributed systems, while Robin Harris provided a great overall summary of what happened.
Any time I hear about random failures leading to timeouts (leading to retries), I think about staged event-drive architectures (SEDA). I heard about SEDA when Matt Welsh (currently of Harvard) came into EMC to talk to the Centera team. His thoughts about queues and feedback mechanisms were influential to Centera (which in itself is a distributed system).
No matter how you slice it, replication is a hard problem. Replication of online social networks (OSNs) are presenting new challenges that researchers are trying to tackle.
OSNs, like Facebook and Twitter, often use Cassandra for the ingest and organization of content. Cassandra uses a distributed hash table (DHT) as part of its operation. DHT’s do not preserve social locality of reference (keep my data close to my friend’s data). Social locality of reference is a key performance enabler during OSN query operations (e.g. refresh my home page feed). Research is already underway on the development of new algorithms that leverage social graphs. These algorithms attempt to co-locate data in an OSN (e.g. see the SPAR presentation based on the SIGCOMM 2010 paper).
This brings up an interesting question: can social locality be preserved at a disaster recovery site? Are there new replication algorithms that can leverage social graphs during replication?
This topic has been researched as part of a collaboration between EMC Research Cambridge and UMASS Boston. The resulting paper, Preserving Social Locality in Data Replication for Online Social Networks, will be presented next month at SIMPLEX 2011. I won’t steal too much of their thunder ahead of the conference, but there are some new algorithms coming. I will say that the algorithms are replication-friendly.
Why is EMC expanding its knowledge of distributed algorithms in Cambridge? The research is essentially an outgrowth of the first place award winner in EMC’s 2008 Innovation Showcase. Winning ideas go through an incubation process within EMC. After the incubation period, a course of action is recommended. Given EMC’s emphasis on research partnerships with area universities, UMASS Boston was chosen based on the quality of their PhD candidates.
If you have any questions about the paper or upcoming conference, please contact me by leaving a comment below.
Steve
Twitter: @SteveTodd
