Kevin W Monroe
on 16 February 2016
Though only a few weeks into 2016, the Juju Big Data team has already been busy engaging diverse communities at conferences and meetups. We’d like to share our perspective on a few of those engagements here.
SCaLE 14x / UbuCon, Jan 20-24, Pasadena
This year’s Southern California Linux Expo was especially exciting as it also hosted UbuCon and DevOps Days LA. This meant we had the chance to speak with a wide range of attendees, from casual users to hard core developers, all under one roof.
Jorge Castro and Kevin Monroe gave a talk about complexities in Big Data environments and how Juju helps simplify them. There were ~40 audience members that watched us deploy our syslog analytics bundle live on an OrangeBox in about 20 minutes. We’ve written about this bundle before, but talking through it on stage was really well received. The audience had tons of great questions about Juju and MAAS — luckily we had Marco Ceppi there to field all the hard ones 😉 The talk was recorded if you’re interested:
Outside of our formal talk, we had plenty of great hallway chats. I think one in particular sums up why we’re so excited to do what we do: We met a community member that is learning R and is interested in running Big Data jobs with SparkR, but has zero interest in setting up or configuring Hadoop. This almost led him to give up on Big Data, but he decided to come to SCaLE to see if anyone had this figured out. We pitched Juju and our Big Data offerings on Thursday night, and he was so intrigued that he wanted us to block out some time so we could put Juju on his laptop. Marco set him up with free AWS creds, and Kevin walked him through deployment of an HDFS/Spark/Zeppelin bundle. In under an hour, he was running Spark jobs in his own Big Data AWS environment. He was blown away and asked if it would be ok if he extended our Spark charm to support SparkR. Awesome.
This week wasn’t just about preaching for us. We had lots of opportunity to learn from other players in the Big Data space. During the DevOps Days open spaces event, we heard from people implementing the Lambda architecture and walked away with ideas on how we can offer speed, batch, and serving layers in our bundles. We also heard about testing methodologies that go beyond typical function/integration tests. Behavioral testing — the notion of verifying your model behaves consistently over time — was particularly interesting. This bleeds into hotspot detection, where you identify hardware, software, or data anomalies that may be affecting your model’s behavior. This is really cool stuff that we can’t wait to build into our offerings!
While we love talking about our own stuff, we are equally excited to hear from Big Data enthusiasts large and small. The community around Big Data is thriving, and we’re happy to play a role in it. Hopefully, we’ll see you at SCaLE for many years to come!
cfgmgmtcamp & Juju Charmer Summit, Feb 1-5, Ghent
Mark Shuttleworth kicked off the 2016 “Config Management Camp” with what was almost an impromptu demo deploying OpenStack on an OrangeBox, which went over well and helped get more people into the Juju track rooms later in the week. The Canonical employees taking up space had to make way for curious attendees, and we were all very pleased with the turnout. The Big Data session was no exception; a packed room, a great presentation (although we may be biased) and numerous well thought out questions. Tom Barber from meteorite.bi (who had some excellent thoughts on the event) wanted to know if we had any plans to couple with popular, fast moving projects, such as Apache Zeppelin. Luckily, Alexander Bezzubov, sitting next to him, raised his hand and replied: “I’m the release manager for Zeppelin, and yes, we’re working on it.”
Our live demo was — we thought — pretty cool, this time validated by feedback from the audience. When the presentation began we asked people to SSH into our primary Hadoop NameNode with “Funny Usernames (no profanity please)” to generate “access denied” syslog messages. These would be sent to HDFS by Apache Flume to demonstrate the ingestion and storage capabilities of our Big Data charms and subsequently processed with Apache Spark to count words.
The result of all of this calculation would be displayed graphically (beautifully) with Apache Zeppelin. What we hadn’t counted on was that we had a room full of computer-type-people who knew how to write for-loops and after a few minutes we noticed we had more than a few login attempts:
By the end of the session we had somewhere around 30,000 login attempts, and a screenshot from Zeppelin looked something like this:
While the weather in Ghent wasn’t warm or sunny, the faces of our attendees were, and we were ecstatic to share the experience with them. We’re always interested in meeting more community members. If you have questions/comments about these engagements or any of our work, reach out to us in #juju
on irc.freenode.net
.