In the last weeks I visited several Cloud and Big Data conferences. Especially the Big Data Innovation in Boston gained me a lot of insight. Some people only consider the technology side of BigData technologies like Hadoop or Cassandra. The real driver however is a different one. Business analysts discover Big Data technologies as the means to leverage tons of existing data and ask questions about customer behavior and all sorts relationships to drive business strategy. By doing that they are pushing their IT departments to run ever bigger Hadoop environments and ever faster real time systems.
What’s interesting from a technical side is that ad-hoc analytics on existing data is allowed to take some time. However ad-hoc, implies people waiting for an answer, meaning we are talking about minutes and not hours. Another interesting insight is that Hadoop environments are never static or standalone. Most companies take in new data on a continuous basis via technologies like flume. This means Hadoop map reduce jobs need to be able to keep up with the data flow; either by adding more hardware or by optimizing them.
There are multiple drivers to BigData (actually there are a lot) but the two most important ones are these: Analytics and Technical Need for Speed. Let’s look at some of those and the resulting takeaways:
The value is in the insight not the volume
The value of BigData is in the insights that the data can provide, not the sheer volume of it. The reason that more and more companies are keeping all of their log and transaction data is that they want to gain those insights. The sheer size of the data is rather an obstacle to this goal and has been for a long time. With BigData technologies this value can be harnessed.
Don’t forget that data analysts are people too.
Ad-hoc analytics doesn’t have to be instant, but must not take hours either. It was interesting to see that time to result on ad-hoc analytics is considered very important. This is because people are doing those queries, and people don’t like to wait for hours. But even more important is that business analytics is often an iterative process. Ask a question, check the answer, refine or change the question. Hours long MapReduce jobs are prohibitive to this process.
New data is coming in all the time
Big Data Environments are constantly fed with new data. This is not really big news, but I was still surprised by the constant reiteration of this fact. The constant data growth means that ad-hoc queries get either slower over time or need to work on samples. To remedy this companies are writing scrubbing and categorizing MapReduce Jobs. These jobs basically strip out all the unimportant stuff and put cleansed, streamline easy to access data into new files. Instead of executing analytics against raw files, the analyst works on a cleansed data set. The implications are that scrubbing jobs need to be maintained all the time (as data input is changing over time) and they need to be able to keep up with the velocity of the input. MapReduce is not allowed to run for hours, but needs to be quick and iterative.
BigData is not cheap!
While it sounds obvious, it is something that is not talked about by the vendors unless specifically asked. Hadoop requires a lot of hardware and a lot of expertise. Especially the expertise is hard to come by as of yet. And while hardware might be cheap (you don’t need expensive boxes for Hadoop) the bigger the environment the higher the operational costs. That operational cost is the reason some Hadoop vendors exist on services alone and also why customers are demanding better monitoring and management solutions.
Data must be accessible at low latencies to provide value
One very interesting fact is that most adopters that use Hadoop for analytics, use it for ad-hoc analytics and scrubbing, but not as a traditional warehouse. They use MapReduce to do the heavy lifting that is usually reserved for ETL jobs and put the resulting dimensions in existing data warehouses or into a NoSQL solution like HBase, Cassandra or MongoDB. These solutions provide low latency access semantics and are then integrated in the transactional application world, e.g. to provide recommendations to the end users.
This does not absolve them from optimizing their Hadoop environment where they can, but it gives them the much needed real time access that Hadoop so far does not provide. This also makes for additional complexity that needs to be maintained and monitored.
NoSQL solutions need management and monitoring as well
NoSQL solutions are most often used to provide low latency databases with failover and horizontal scaling characteristics. As expected, practitioners quickly run into new issues like distribution and wrong access patterns. Most NoSQL solutions lack sophisticated monitoring or performance analysis tools and require experts instead. Fortunately several companies are working on providing those tools and some APM vendors work hard to support NoSQL databases similar to normal databases. This is emphasized by another interesting finding: With a fast and scalable data storage, the application itself quickly becomes the response time and scaling bottleneck.
Applications using NoSQL technologies are more complex
Most NoSQL solutions surrender more complex logic like joins in order to achieve horizontally scalable data distribution. That logic is moved to the application – arguably this is where it should be anyway. NoSQL solutions require data to be stored in a query access optimized way – de-normalization is the key. The flip side of storing data multiple times and the need to keep it in sync on updates, is that the storage logic again becomes more complex. More application logic usually means less performance.
My conclusion as a performance engineer is relatively clear: BigData requires Performance Management and Monitoring Tools to fulfill its promise in a cost effective and timely manner. Here are some suggestions on what you should think about when you start a Big Data project.
- Large Hadoop environments are hard to manage and operate. Without automation in terms of deployment, operations, monitoring and root cause analysis they quickly become unmanageable. Make sure to have a monitoring solution in place that informs you pro-actively of any infrastructure or software issues that would affect your operation. It needs to give you an easy way to pinpoint the root cause.
- The easiest way to identify new performance issues is to detect and analyze change. Adopt a life cycle and 24/7 production APM approach. It will enable you to notice changes in data and compute distribution over time. In addition a life cycle approach will allow you to immediately pin point any negative changes introduced by a new software release.
- Don’t just throw more and more hardware at the problem. While you can use cheaper hardware for Hadoop, it’s still cost. But more than that you have to consider the operational drag. Every node you add will make traditional log based analysis more complicated. Instead ensure that you have an APM solution in place that lets you understand and optimize MapReduce jobs at their core and reduce both the time and resources it takes to run them.
- Your Hadoop cluster is no island, but will always be connected in some form or the other to a real time or at least transactional system. Make sure that you have a monitoring solution in place that can support both.
- NoSQL applications tend to have more complex logic. The very performance and scalability of the store depends on correct data access and data distribution. An good monitoring solution allows you to monitor and optimize that additional complexity with ease; it also enables you to understand how your application access the data and how that access is distributed across your NoSQL cluster in your production system. The best way to ensure a scalable and fast NoSQL store is to ensure optimal distribution and access pattern.
BigData is still very much an emerging technology and its promises are huge. But in order to deliver on those promises it must be cost and time effective to those that harness its value – The Business and not just technology experts.