Strata + Hadoop World 2015 - Conference or Trade Show?
Fresh off the plane from a few days at Strata + Hadoop World, I thought I’d
draft a quick wrap up of the
conference trade show. This
being my first attendance, expectations were quite high and I was
unfortunately left frustrated and a bit disappointed.
The event was held at the San Jose Convention Center and despite the buildings timid facade, the interior boasts a sprawling labyrinth of meeting rooms, public spaces and hidden hallways. The organizers did well to provide plenty of workspace, conspicuous power and some of the best wifi I have experienced at a gathering of this size. From a technical execution standpoint, the only blemish was sparse catering between scheduled breaks. Finding a cup of coffee to prop yourself up during the extended afternoon sessions was an effort better fit for a scavenger hunt. That said, the catering, when it was around, was on point with a fair diversity of offerings and a pleasant staff.
Not unlike other gatherings of this size, the event pivoted around a sizable expo hall with plenty of variety on display. During Wednesday’s expo crawl, I found everything from two to three person startups to traditional enterprise integration firms to the multinationals; most all with polished pitches aimed at the myriad of personalities and business roles present.
I zig-zagged my way through a sea of drones, “Bacon Scientist” t-shirts and
overly caffeinated sales professionals and was struck by a rather horrifying
realization - nearly every third vendor was brandishing demos about how their
SQL/ACID transaction solution for Hadoop was “the fastest” or “the only”.
While many vendors are claiming that SQL will ultimately bring the masses to Hadoop, I worry that this very useful but still corner case capability for the platform has reached snake oil status. More on this topic in my next post.
The crawl ended at a modest 6:30pm leaving out-of-towners to fend for themselves that evening. As is often the case on the road, Twitter saved a boring evening of hotel room email when Alistair Croll tipped me off to an event hosted by The Hive. The meet and greet sandwiched a panel talk skillfully hosted by Alistair with Oscar Celma of Pandora, Jennifer Kennedy of SoundHound and Dr. Douglas Eck of Google’s Play Music.
As is often the case when listening to companies that have direct access to the emotional centers of our brains, I found myself completely enchanted with the possibilities and absolutely terrified for the future. Perhaps the deepest and most unsettling settling insights came late in the evening as Dr. Eck explained the nuances of down voting a piece of music. The overly eager group brandishing libations were lobed a doozie - “How can you tell if someone down voting every song on Coda hates Led Zeppelin or just the album?” Blend the difficulty of that question with the threat of overplaying a beloved artist to a diehard fan and the complexities at hand become apparent. As I stood there, Dr. Eck layered on the deep emotional connection we have to music, the difficulty of sustaining a listener’s emotional state - positive or negative, and the goal of providing a rich and long lasting experience to expose the user to more advertising. I closed out a rather late evening wondering how I would fair against two more days of thought provoking content.
The following morning I found myself sitting in an overly dark room, amongst a sea of dimly lit laptops listening to the keynote addresses. In turn - Cloudera, MemSql, SalesForce, MapR and IBM speakers leveraged identical formats to fill 10-15 minute slots with predictably bold statements about the future of “Big Data”, each culminating with a terse and unashamed pitch for their various products. On the tail of these presentations we encountered a meandering spiel by Dj Patil, the newly appointed US Chief Data Scientist, which included a short but warmly welcomed spot by President Obama. Like papers strewn across a desk, the disjointed talk had clearly been disrupted by Mr. Patil’s recent appointment and, unfortunately, it fell short of the momentum inducing spot it had been positioned for. For those audience members seeking substance and had not yet walked out, two excellent talks punctuated the tail of the keynotes. The information dense, though poorly organized presentation from Solomon Hsiang on “Data-Driven Policy” brought to bear the complexities of interleaving data at global scale and the nervously delivered, but delightful and thought provoking talk from Poppy Crum of Dolby clearly illuminated our susceptibility to sensory augmentation through conflicting information sources. In all, the balance in the keynotes would foreshadow the remainder of the content at Strata + Hadoop World.
I kicked off the sessions with a talk on Apache Kafka which was about as informative as the project’s homepage. Kafka is exactly what you would expect to get from merging the clustering approaches of Hadoop, the WAL of a product like HBase or Cassandra and a message queue like RabbitMQ - all together very interesting stuff. I would also say that behind mentions of Spark, it was the 2nd most talked about project in the sessions. Unfortunately in every OmniGraffle systems chart that was presented, replacing each occurrence of “Kafka” with “Tibco” and “Data Bus” with “Enterprise Message Bus” and it could easily have been 2005 all over again. My point here is that the message bus in enterprise architecture has found its place and the considerations, such as asynchronicity, decoupling and message schema validation via middleware should be taken into account when building systems with Kafka. If you are interested in getting a 10K foot overview, check out this SE-Radio podcast on the project. Last note of caution is that Kafka is not case hardened from a security perspective yet; those features and functionality are however on the roadmap.
On the point of security, the Kafka session was followed up with an interesting presentation on the future of Hadoop security. Core issues such as user proxying and restricting data access resonated with well with the audience. Security after the fact on the platform is the current modus operandi and there were encouraging statements indicating that the trend may be reversing as common practices solidify. I am not well informed on the topic, but my instinct says that we will see all the common RDBMS security controls ported or projected into the ecosystem very soon.
The remainder of the early afternoon was filled with two sessions on data visualizations, both of which confirmed the immaturity of the field and the trail blazing that is being done daily by teams within the big data visualization space. A quick post presentation chat with Etan Lightstone of New Relic re-introduced the work that Ernest Dichter did for Betty Crocker wherein adding an egg made housewives feel more involved in the process of baking; while relevant this is actually not true (see here). Either way, the chat did reignite concerns I have about giving users in data driven jobs just enough data to help them draw their own, albeit guided, conclusions.
Around 5pm that evening I finally encountered the session I had been waiting for. Scott Donaldson of FINRA gave a stunning architectural and implementation review of their fraud detection systems. While the juicy details like cluster sizing were held back, he detailed an architecture which houses several petabytes of data and ingests a massive number of discretely schemed transactions per day. Their solution is an interesting blend of bulk raw transaction data, data warehousing, network graphs to trace transaction flow between entities, data extraction into relational stores for deep analysis and an intense UI that enables investigators to exceed their prior investigational contexts.
The following day picked up with an architectural patterns for Hadoop talk that while deeply interesting was for the most part a transposition of enterprise architectural patterns mixed with what has become standard data retention procedures of event sourcing systems. On the back of that session, I had set high hopes for a deep dive into Elastic Search only to find a cursory review of Kibana and its relationship to ElasticSearch. Adding to the mediocre nature of the day was an introduction to using Parquet for ETL that landed in similar fashion to the prior day’s Kafka talk. We were presented with primarily a review of the internal structures of Parquet with little to any commentary as to its form and fit within an architecture. The product itself is quite intriguing as a columnar data store that provides data packing for significant disk savings and increased search performance - for those in need of that feature set.
In all, the conference was marginal at best. Though it’s obviously possible that I simply landed on a slew of sessions that missed reaching the potential of their abstracts, my instincts say that content like FINRA will remain uncommon until implementors on the platform are willing to open up about what they are really doing behind closed doors. So if you are just starting out with the platform, the conference as a trade show will introduce you to the big players in the field and provide enough orientation to sound informed during your first sales call. Personally, should tickets fall into my lap next year and the ecosystem has see some significant changes then I might bear the cost of a flight and hotel room, otherwise I will be finding a different conference for my big data interests.
Check back next week for a follow-up post on the state of big data post Strata & Hadoop World.