On October 24, 2012 Cloudera announced the release of Cloudera Impala and the commercial support subscription service of Cloudera Enterprise Real Time Query (RTQ). During the Hadoop World/STRATA Conference in NYC, I was invited over to see a demonstration. Impala is a SQL based Real Time Query/Ad Hoc query engine built on top of HDFS or Hbase. As I watched the demonstration unfold, I wondered if one of the remaining technology gaps in the NOSQL arsenal had been closed. What gap you ask? Near Real Time Analytics on a NOSQL stack. Working with customers across the Cyber Security customer space, not only do they face the familiar BIGDATA horsemen of the apocalypse: Volume, Velocity and Variety but one more large challenge crept in: Time (V3T). The Near Real Time Analysis/Near Real Time Analytic capability that Cloudera Impala provides is essential in many high value use cases associated with Cyber Security: comparing current activity with observed historical norms, correlation of many disparate data sources/enrichment and automated threat detection algorithms.
When the demonstration concluded, the Cloudera representatives and I discussed the potential of performing an informal independent evaluation of Cloudera Impala against some of the common Real Time/Near Real Time use cases in Cyber Security. I agreed to step up and perform an independent evaluation as well as developing a demonstration platform for FedCyber 2012 (almost three weeks hence for inquiring minds). So let us set the field: a new BETA technology, NO prior exposure to the technology or documentation, a vendor making promises, addressing a large technology gap and three weeks to implement, seemed straight forward; no pressure.
The day after I returned from the STRATA Conference, I returned to my office and provisioned four Virtual Machines in order to build the Impala demonstration. As a committer/contributor for SherpaSurfing an open source Cyber Security solution, I have an abundance of data sets, enrichment sources, Hive data structures and services. Given the amount of time and the audience for FedCyber 2012, I decided to focus on some Intrusion Detection and Netflow related use cases for the demonstration. The data sets for the demonstration included base data sets: 20 million Netflow events, 8 million Intrusion Detection System events and enrichment: Geographic, Blacklist, Whitelist and Protocol related information. Each of the selected uses cases for this demonstration is critical to the Perform Near-Real Time Network Analysis domain in Cyber Security. The name for the demonstration system was decided to be the Impala Mission Demonstration Platform (IMDP). The IMDP was implemented based on vendor recommendations with no tuning or optimization.
The IMDP effort provided me with my first opportunity to work with Cloudera Manager. Although this post is focused on Cloudera Impala I would be remiss not to mention Cloudera Manager. I have worked with Hadoop since 1.0 and built more than a few clusters over the years. I used the installation and configuration guides provided with Cloudera Impala and followed the recommendations. One of the first recommendations was use of the Cloudera Manager. Using the Cloudera Manager (CDH 4.1), I was able to roll out a four node cluster in two hours. I was able to discover the hosts, manage services and provision them in accordance with the IMDP deployment plan. The deployment plan consisted of:
- node 1 – hbase, hdfs, impala, mapreduce
- node2 – hbase, hdfs, impala, mapreduce
- node3 – hbase(region server, master), hdfs(namenode), impala(impalad, statestore), mapreduce(job tracker, tasktracker) , hue, oozie and zookeeper
- node4 – Application Tier, Cloudera Manager
The Cloudera Manager saved at least two days of effort in deploying the cluster, the tight integration with the support portal, comprehensive help and one place to work with all properties of the entire cluster and view space consumption metrics; verdict on Cloudera Manager: Cloudera masterful, bold stroke, thumbs up.
Now that the cluster build-out completed; I shifted attention to deploying and configuring the Cloudera Impala service. Using Cloudera Manager, I deployed Impala on three nodes: three instances of Impalad and one impala state store, in a matter of minutes. I completed the deployment and configuration of the Hive MetaStore. Keeping in mind this is a BETA; the documentation was complete, but fragmented on deployment and configuration (HIVE MetaStore portion); verdict on impala deployment and configuration: solid for a BETA (needs an example hive-site.xml, configuration guide needs better flow).
At this point all configuration and deployment was completed, attention turned to building data structures and loading data. I took the Data Definition Language (DDL) scripts or data structures for ten data sources and enrichment; ported them over to Hive and tested them in less than four hours. It is worthy of mention that the data sources for this demonstration are large flat tables: netflow and intrusion detection system. Cloudera Impala uses HIVE as an Extract Transform Load (ETL) engine, using Hive I defined all of the data structures in source files which were sourced using hive shell: created a database (Sherpa). Hive was then used to load data into the tables that were just created. Creating data structures in Hive was simple as usual and loading data sets was quick (20 million netflow events in 57 seconds). Logging into impala-shell, issued a refresh of the MetaStore and I was working with data. I performed verification of the data load, all data loaded and no issues were revealed. One area of potential improvement would be more comprehensive messages on load failure. Defining the data structures and loading data using Hive was nothing new; verdict: really good; easy to use, easy to load, but need to improve failed load messages.
Finally, we moved on to the most interesting stage which is using Cloudera Impala in a series of Real Time Query (RTQ) scenarios that are common across the Cyber Security customer space. The real world scenarios selected come from the perform netflow analysis set of use case(s). In each of these scenarios, the exact same queries were executed on the same cluster using Hive and then Impala against the same data structures (database and tables). In the Hive approach, we traverse the batch processing stack and with Impala we traverse the Real Time Query (RTQ) stack performing a series of analytics. In the first use case, I ran a five tuple (sip, sport, dip, dport, protocol) summary covering bytes per packet, summing bytes and packets for a 20 million event set resulted in: identical result sets, Hive 82 seconds – Impala 6 seconds. In the second use case, I performed a summary of destination ports where the source port is 80 which resulted in: identical result sets, Hive 57 seconds, Impala 5 seconds. In the third use case, I performed correlation between netflow and intrusion detection systems, correlating netflow with intrusion detection events for several hours which resulted in: identical result sets, Hive 40 seconds, Impala sub-second. Finally, for FedCyber 2012, I developed a java based situational awareness dashboard which connected to Cloudera Impala via ODBC and executed analytics performing: correlation of blacklists, Intrusion Detection, Netflow, statistical cubes for ten hours with a refresh of every five seconds without failure or issue. The ODBC implementation easily provided the ability to export data to desktop tools (using ODBC) and common BI tools as advertised. Developing and Using Cloudera Impala verdict: This is as advertised; easy to use, easy to implement on, very fast, very flexible and more than capable of running real time analytics. The Impala shell is limited but much of the demonstration work was done using result sets so it was not an impediment.
In summation, I have worked for over a decade across the vast BIGDATA technology space covering Legacy Relational Database, Data Warehouse, and NOSQL; Cloudera Impala proved more than capable of running near real time analytics and providing mission relevance to customers with a Near Real Time (NRT) requirement. Based on my initial review Cloudera Impala appears to be a bold step in closing the gap of near real time analytics on a NOSQL stack. I did encounter some minor problems, but the few problems and limitations that were encountered in this demonstration were documented and published in the known issues document so they will not be shared; none were show stoppers.
The notes, details and all of the lessons learned, data structures and the configuration guide from the demonstration are being published out on Github under SherpaSurfing in the coming days. These documents cover everything in detail and will enable developers to replicate the demonstration platform and get a jump start on Cloudera Impala. Finally, I would like to thank two contributors: Hanh Le, Robert Webb and Six3 Systems for helping me pull this off.