Today I drove up to Boston for the day to attend NoSQL Live. My experience so far within the NoSQL community has been limited to what we've built in-house at Disney and ESPN over the past decade to solve our scaling issues, more recently has been ESPN's use of Websphere eXtreme Scale, and the very latest has been my own experimentation with HBase which hasn't gotten much further than setting up a four node cluster. I've read a little about Cassandra, memcached, Tokyo Cabinet and that's about it. So before the sandman wipes away most of my first impressions of the technologies discussed today, I wanted to record my thoughts for posterity or, at the very least, tomorrow.
Cassandra seems to be the hottest NoSQL solution this month with press about both Twitter and Digg running implementations. My impression, I'm wary of "eventual consistency". I don't feel I understand the risk and ramifications well enough to design a system properly. When Jonathan Ellis of Rackspace Cloud mentioned that Digg needed to implement Zookeeper-based locking on top of Cassandra so that diggs get recorded correctly, I realized how poorly I understand eventual consistency and how risky it could be. But my impression of Cassandra isn't all negative, it definitely seems to have less baggage than HBase by not being built on top of HDFS. I'll get into what that means a little later.
Unfortunately the speaker that 'represented' memcached gave off a vibe that really turned me off to the product. I know that's incredibly shallow, but this is first impressions after all and not perfectly-evaluated impressions. Mark Atwood sat on the first panel of the day "Scaling with NoSQL" and his whole attitude seemed to say "memcached is all you'll ever need and these guys next to me are just overdesigning hacks". His answers were short and his tone was quite condescending even when addressing audience questions. Not a very good first impression of him. But luckily today wasn't my first impression of memcached as I was pointed in its direction just last week by a Disney colleague. My research before today has me intrigued about using it as a replacement for ehcache as a second-level cache provider to Hibernate which we use as an ORM in one system at ESPN right now.
Document Oriented Databases (Riak, CouchDB, MongoDB)
Wow, this is a subgroup of NoSQL technology that I had heard of in passing but was really unaware of what problem they were trying to solve. Riak had the best answers for scaling and operational-ease. With homogeneous nodes and consistent hashing, Riak promises that adding and removing nodes are seemless. CouchDB and MongoDB sounded like a 'me too' answer so I'm interested to find out what that really means for each, or better yet what it doesn't mean. But the concepts of document-oriented databases really meshes well with ESPN's current fantasy user database. Our fantasy user profiles are stored in a traditional RDBMS as serialized maps of maps, one row for each user. Since its serialized to a BLOB column its completely opaque to reporting and analytics. To keep that model but have vendor support for divining information and having transparency into it sounds exciting. I really need to look into these. Riak definitely won this round of first impressions.
This was a technology I was referred to by a colleague and read through their site last week. I was far from impressed then since it seems much too low-level for my taste, similar to my impressions of Carbonado which we use at ESPN. The lightning talk by Flinn Mueller got me a little more interested. He seems to be doing interesting things but from an analytics and reporting perspective. He was vague on how loads the data from his primary store and what the scale of the data is, so my first impression: its a toy. I'm sure that's an unfair characterization but I'm not trying to be fair tonight. But honestly, Tokyo Cabinet makes no bones that it punts on horizontal scaling which is the deal breaker for me.
I looked at Hypertable (as in read their website) about 18 months ago on the suggestion of a colleague when discussing HBase around the same time. This conference didn't change my opinion, which is "It's HBase but written in C". It doesn't seem to bring anything else to the table which to me is a blocker. JVM implementations are available for all the operating systems I use and so I don't like the idea of needing to find the right binary to download for a given box. When it comes to Java vs. C, I choose Java but I'm also extremely biased as I've been a Java developer nearly my entire career.
Well I saved HBase for last. It's the one I've had the most experience with, though that experience can still be measured in hours. As I hinted at earlier, this conference gave me the first impression that HDFS is a weight around the neck of HBase. I was surprised to get that feeling from the room, since my impression has been purely positive so far. It is also getting a lot of flack from the 'single point of failure' problem associated with the current HDFS architecture's Name Node. Apparently performance is a dog since it was "only" designed to be highly distributed with no promise of when you'll get your data. This burden seems to carry over to HBase. But after talking to Ryan Rawson one-on-one at the end of the HBase lab, it's clear he is of the strong opinion that its getting a bad wrap. He also makes very convincing arguments about the scale of what HBase is currently doing in real production environments vs. competitors like Cassandra. It's very pursuasive and you can read more of the details in a very active thread I kicked off on the HBase user group earlier this week.
Disclaimer: Though I mention my employer ESPN in this post, these are my own personal opinions and don't represent the opinions of the company. The final decision on this stuff is "above my pay-grade" as they say.
Anyway, I was scribbling on my whiteboard for 4 days straight this past summer trying to wrap my head around eventual consistency, multiple entity versions, and vector clocks.
These articles and talks by the CTO at Amazon were helpful:
And then, somewhat unrelated, I really liked this article about the Big Table type of architectures.
In the end, we decided to prototype and build on both the eventual consistency architecture for availability and the Big Table architecture for ACID requirements, trading away consistency and availability respectively. We're planning to expose both of these systems to our middleware through a unified interface. Data will be handled by our own db management system which will choose the appropriate data model architecture based on the specifications given to it by the client code.
Incidently, we are not directly running any of the systems featured at NoSQL Live. We are indirectly using them through platform services at Google, Amazon, and RackSpace.
We'll see how it goes, but prototyping and testing has shown really great results. At the very least, it's been a lot of fun.
We use Narwhal (http://github.com/280north/narwhal) and try to follow the new CommonJS specs as much as possible. (http://commonjs.org/)