June 2014 Summaries

The benchmark by Thumbtack Technology for Aerospike has been analyzed with regard to its methodology and results. While the test was performed on bare metal machines with SSD disks, there were issues with benchmark hygiene such as not properly isolating read runs or accounting for ongoing compactions in Cassandra. The benchmark also measures different things in each tested system, particularly in the insert and read-heavy workloads. Further investigation revealed that Aerospike and Couchbase's faster write times compared to Cassandra could be attributed to their asynchronous buffering of writes, while Cassandra synchronously writes to its commitlog by default. The author suggests using a standardized methodology for benchmarking databases to ensure accurate results.

The text discusses four simple rules that should clarify most queries when using DataStax drivers for Cassandra, such as C#, Python, or Java. These rules involve configuring a Cluster instance, using prepared statements, and utilizing batch operations. A Cluster instance allows for the configuration of connection handling aspects like contact points, request routing policy, retry and reconnection policies, etc. Prepared statements offer performance benefits by being parsed and prepared on Cassandra nodes, ready for future execution. Batch operations combine multiple data modification statements into a single logical operation, ensuring atomicity in execution. The text also mentions some specific scenarios where these rules might need slight tweaking, such as using a predefined number of Sessions when interacting with a large number of keyspaces and handling null values in prepared statements.

This article discusses strategies for bulk loading data into Titan at varying scales, focusing on hundreds of millions and billions of edges using Faunus as the loading tool. It provides a step-by-step guide to loading the DocGraph dataset with approximately 1 million vertices and 154 million edges using a single Hadoop node running in pseudo-distributed mode. The article also demonstrates how to load the Friendster social network dataset, which represents a graph with 117 million vertices and 2.5 billion edges, using a four-node Hadoop cluster. It emphasizes that while there are common strategies for loading data at different scales, the actual approach must be adapted to the specific data and domain.