A case for predictable databases

Barzan Mozafari, Assistant Professor of Computer Science and Engineering at the University of Michigan (Ann Arbor), will be giving a talk on the predictability of performance in database systems at the OpenStack Live conference in Santa Clara, California on Tuesday, April 14.

He is a member of the Michigan Database Group and the Software Systems Lab; and passionate about building large-scale data-intensive systems, with a particular interest in database-as-a-service clouds, distributed systems, and crowdsourcing. In his research, he draws on advanced mathematical models to deliver practical database solutions. (from his OpenStack Live profile)

I asked him these 5 questions before he takes the stage on April 14.

Why is database performance so important to system administrators working with OpenStack?

Today, databases host some of the most valuable data in every enterprise. Thus, database performance is the key to unlocking the ultimate value hidden in their data. We live in a fast paced world where making the right decision a microsecond before your rival is all you need to win. Think about online advertising, recommendation systems, etc.

What are some causes of performance problems for databases in cloud installations?

The opaque nature of cloud-based systems and the decoupling of the database operators from database users. In other words, the same reasons that make clouds are attractive, they also make them perform poorly in many cases. Greater degree of uncertainty in cloud settings is often unparalleled.

What tools are you using to evaluate these problems?

We have developed a number of open source tools that we believe are quite effective in diagnosing and alleviating many common types of performance problems.

Is the University of Michigan using Database as a Service (DBaaS) in production?

Our tools are not restricted to DBaaS. Our DBSeer and DBSherlock are open source tools that can be used by anyone in academia or industry. Our focus is on bridging the gap between research and the real-world, between academia and industry. We have integrated the latest research breakthroughs into practical and easy-to-use software tools that can be used in a any performance-sensitive application.

Here at the University of Michigan Ann Arbor, we have a long legacy of database research. We value both research and real-world applications. We always start from a real world problems and end with a real-world solutions. We are not your typical academics whoe are happy with publishing their research as mere papers. We see to it that our research is practical and impactful, and solves a real-world problem. This is why we always open source software and contribute our latest breakthroughs back into the open source world.

Also, the alternative is to treat the database system as a black-box: you do not gain any insight and even if you do, because you have no mechanism of implementing your ideas into a closed-source product.

Tell us more about your OpenStack Live talk.

My talk will be about how predictability of performance has long been neglected in database systems.

For the past 40 years, we have mostly focused on reducing the average latency and have ignored that being able to guarantee a consistent and predictable level of performance is often equally important, if not more important. In my talk, I will identify a few tools that allow database administrators (DBAs) to gain insight into their workload and automatically diagnose their performance problems, and achieve and sustain predictable performance in their database system.

This article is part of the Speaker Interview Series for OpenStack Live. OpenStack Live is a conference which is designed to teach attendees about the best practices and performance considerations for operating OpenStack, taking place in Santa Clara, California on April 13 and 14, 2015.