I had an amazing time at DEBS 2016! It was great to meet other researchers in my field, discuss my ideas (and have them dissected) and make some new friends.
I was extremely excited to meet Karthik Ramasamy who is the head of real time analytics at Twitter. He gave a very interesting talk about the evolution of Twitter’s new distributed stream processing engine (Heron) and how they use it in production. This included the issues they were having with sizing the resource requirements of their streaming topologies and setting the parallelism of their topology components. This process currently has to be done manually and can take up to a week to complete. As Heron is an evolution of the Apache Storm system I use in my research, I hope that my research could be applied to solving this issue with Heron.
I am going to my first academic conference! Distributed Event Based Systems (DEBS) in Irvine, California which takes place at the end of June this year. I will be presenting as part of their Doctoral Symposium, where I will explain the basis, current progress and future plan for my PhD research and get feedback from experts in the field!
DEBS is one of the main sources of literature for my PhD research and I am incredibly excited to be able to attend. As part of the Doctoral Symposium I submitted a 4 page report that will be published (in summerised form) in the DEBS proceedings. I hope to be able to publish a copy of the report here soon once I have got my head around the copyright issues.
The presentation slides for the talk can be found here.
After the talk there will be chance to have a play with Spark yourself. Firstly you need to download the Spark binaries and unpack them to a folder on your local machine.
I also have a few USB keys with the binaries on, so wave your hand if the network is slow.
Run the REPL
If you want to run an interactive PySpark interpreter then run the following command inside the spark directory:
The master flag lets you specify the master node of the cluster, however the local option allows you to run this on your machine with the specified number of worker threads. Be careful not to set this number two high (certainly no higher than the number of CPU cores you have available) or you will freeze your machine.
If you want to be able to import packages not in the standard library (or numpy) the you can include the .egg, .zip or .py files as an argument. You can then use them as a standard import:
Run a script
You can write your scripts and submit them to spark using the spark-submit script in the bin folder. This works similarly to the pySpark shell:
Remember to add the following import to your script:
The spark-submit script will handle the dependencies for this automatically. For your scripts you also need to form a spark context object yourself (unlike for the pySpark shell where this is provided for you):
Supplying the master address in the script is optional, if you do you don’t need to use the --master flag in the spark-submit call.
The easiest way to get up and running with Spark is to try the word count example. You can run this via the shell or copy and paste the code into a script and use spark-submit: