The next edition of the HP Discover Performance Podcast Series highlights how Infinity Insurance Companies in Birmingham, Alabama has been deploying a new data architecture—native column store databases—to improve productivity for their analysis and business intelligence (BI) queries.
To learn more about how Infinity has improved their performance and their results for their business analytics, BriefingsDirect interviewed Barry Ralston, Assistant Vice President for Data Management at Infinity Insurance Companies. The discussion, which took place at the recent HP Discover 2013 Conference in Las Vegas, is moderated by Dana Gardner, Principal Analyst at Interarbor Solutions.
Among other findings, Ralston and his team has seen a 100 times improvement in their top 12 worst-performing queries or longest-running queries when moving from a row-store-based Oracle Exadata implementation to a column store-based HP Vertica deployment. [Disclosure: HP is a sponsor of BriefingsDirect podcasts.]
Here are some excerpts:
Gardner: What was it that you've been doing with your BI and data warehousing that prompted you to seek an alternative?
Ralston: Like many companies, we have constructed an enterprise data warehouse deployed to a row-store technology. In our case, it was initially Oracle RAC and then, eventually, the Oracle Exadata engineered hardware/software appliance.
We were noticing that analysis that typically occurs in our space wasn’t really optimized for execution via that row store. Based on my experience with Vertica, we did a proof of concept with a couple of other alternative and analytic store-type databases. We specifically chose Vertica to achieve higher productivity and to allow us to focus on optimizing queries and extracting value out of the data.
Gardner: What does Infinity Insurance Companies do? How big are you, and how important is data and analysis to you?
Ralston: We are billion-dollar property and casualty company, headquartered in Birmingham, Alabama. Like any insurance carrier, data is key to what we do. But one of the things that drew me to Infinity, after years of being in a consulting role, was the idea of their determination to use data as a strategic weapon, not just IT as a whole, but data specifically within that larger IT as a strategic or competitive advantage.
Gardner: You have quite a bit of internal and structured data. Tell me a bit what happened when you moved into a Vertica environment, first in the proof of concept phase and then into production?
Ralston: For the proof of concept, we took the most difficult or worst-performing queries from our Exadata implementation and moved that entire enterprise data warehouse set into a Vertica deployment on three Dual Hex Core, DL380 type machines. We're running at the same scale, with the same data, with the same queries.
We took the top 12 worst-performing queries or longest-running queries from the Exadata implementation, and not one of the proof of concept queries ran less than 100 times faster. It was an easy decision to make in terms of the analytic workload, versus trying to use the Oracle row-store technology.
Gardner: Let’s dig into that a bit. I'm not a computer scientist and I don’t claim to fully understand the difference between row store, relational, and the column-based approach for Vertica. Give us the quick "Data Architecture 101" explanation of why this improvement is so impressive? [Learn more about the upcoming Vertica conference in Boston Aug. 5.]
Ralston: The original family of relational databases—the current big three are Oracle, SQL Server and DB2—are based on what we call row-storage technologies. They store information in blocks on disks, writing an entire row at a time.
If you had a record for an insured, you might have the insured's name, the date the policy went into effect, the date the policy next shows a payment, etc. All those attributes were written all at the same time in series to a row, which is combined into a block.
So storage has to be allocated in a particular fashion, to facilitate things like updates. It’s an optimal way of storing data for transaction processing. For now, it’s probably the state-of-the-art for that. If I am running an accounting system or a quote system, that’s the way to go.
Analytic queries are fundamentally different than transaction-processing queries. Think of the transaction processing as a cash register. You ring up a sale with a series of line items. Those get written to that row store database and that works well.
But when I want to know the top 10 products sold to my most profitable 20 percent of customers in a certain set of regions in the country, those set-based queries don’t perform well without major indexing. Often, that relates back to additional physical storage in a row-storage architecture.
Column store databases—Vertica is a native column store database—store data fundamentally differently than those row stores. We might break down a record into an entire set of columns or store distinctly. This allows me to do a couple of different things from an architectural level.
First and foremost, I can sort, compress, and organize the data on disk much more efficiently. Compression has been recently added to row-storage architectures, but in a row-storage database, you largely have to compress at the entirety of a row.
I can’t choose an optimal compression algorithm for just a date, because in that row, I will have text, numbers, and dates. In a column store, I can apply specific compression algorithm to the data that's in that column. So date gets one algorithm, a monotone increasing key like a surrogate key you might have in a dimensional data warehouse, has a different encoding algorithm, etc.
This is sorting. How data gets retrieved is fundamentally different, another big point for row-storage databases at query time. I could say, "Tell me all the customers that bought a product in California, but I only want to know their last name."
If I have 20 different attributes, a row-storage database actually has to read all the attributes off of disk. The query engine eliminates the ones I didn’t ask for in the eventual results, but I've already incurred the penalty of the input-output (I/O). This has a huge impact when you think of things like call detail records in telecom which have a 144-some odd columns.
If I'm only asking against a column store database, "Give me all the people who have last names, who bought a product in California," I'm essentially asking the database to read two columns off disk, and that’s all that’s happening. My I/O factors are improved by an order of 10 or in the case of the CDR, 1 in 144.
Gardner: You can’t just go back and increase your I/O improvements in those relational environments by making it in-memory or cutting down on the distance between the data and the processing? That only gets you so far, and you can only throw hardware at it so much. So fundamentally, it’s all about the architecture.
Ralston: Absolutely correct. You've seen a lot of these—I think one of the fun terms around this is "unnatural acts with data," as to how data gets either scattered or put into a cache or other things. Every time you introduce one of these mechanisms, you're putting another bottleneck between near real-time analytics and getting the data from a source system into a user’s hands for analytics. Think of a cache. If you’re going to cache, you’ve got to warm that cache up to get an effect.
If I'm streaming data in from a sensor, real-time location servers, or something like that, I don’t get a whole lot of value out of the cache to start until it gets warmed up. I totally agree with your point there, Dana, that it’s all about the architecture.
In short, in leveraging Vertica, the underlying architecture allows me to create a playfield, if you will, for business analysts. They don’t necessarily have to be data scientists to enjoy it and be able to relate things that have a business relationship between each other, but not necessarily one that’s reflected in the data model, for whatever reason.
Obviously in a row storage architecture, and specifically within dimensional data warehouses, if there is no index between a pair of columns, your performance begins to suffer. Vertica creates no indexes and it’s self-indexing the data via sorting and encoding.
So if I have an end user who wants to analyze something that’s never been analyzed before, but has a semantic relationship between those items, I don’t have to re-architect the data storage for them to get information back at the speed of their decision.
Gardner: What about opening this up to some new types of data and/or giving your users the folks in the insurance company the opportunity to look to external types of queries and learn more about markets, where they can apply new insurance products and grow the top line?
Ralston: That's definitely part of our strategic plan. Right now, 100 percent of the data being leveraged at Infinity is structured. We're leveraging Vertica to manage all that structured data, but we have a plan to leverage Hadoop and the Vertica Hadoop connectors, based on what I'm seeing around HAVEn, the idea of being able to seamlessly structured, non-structured data from one point.
Insurance is an interesting business in that, as my product and pricing people look for the next great indicator of risk, we essentially get to ride a wave of that competitive advantage for as long a period of time as it takes us to report that new rate to a state. The state shares that with our competitors, and then our competitors have to see if they want to bake into their systems what we’ve just found.
So we can use Vertica as a competitive hammer, Vertica plus Hadoop to do things that our competitors aren’t able to do. Then, I’ve delivered what my CIO is asking me in terms of data as a competitive advantage.