Correction to Recent End Point BenchmarkMay 29, 2015
About a month ago, End Point published a NoSQL database benchmark between Cassandra, MongoDB, HBase, and Couchbase. As we reviewed the configuration details, we discovered two errors had been made in configuring Cassandra and HBase.
First, Cassandra’s durability option was not set to ensure no possibility of data loss. Second, HBase had been configured to write data to two replicas instead of just one.
End Point agreed with us that their settings for both Cassandra and HBase resulted in test results that were in error, and that the benchmarks for both databases needed to be repeated. End Point performed that work and the benchmark report has now been revised and reposted.
What, in general, changed? For Cassandra, it was no surprise that the Load and Insert-Mostly tests saw the largest declines in throughput and increases in latency, while more nominal decreases in throughput were seen in the Balanced and Read-Modify-Write tests. Some tests such as Read-Mostly and Analytic actually saw modest throughput gains for Cassandra.
Where HBase was concerned, its performance naturally improved in tests such as Insert-Mostly and Load, while its throughput performance somewhat declined in the other tests.
In the end, as the revised report shows, Cassandra still exceeds its competitors in the benchmarks where both throughput and latency are concerned. Of course, every use case is different, and while general benchmarks such as these can serve as a guide to how databases perform under certain conditions, you need to conduct your own tests to see how each handles your expected workload.
Along with End Point, we regret that incorrect results for Cassandra and HBase were issued in the original report and are satisfied with the results in the revised document. The updated report is now available for download.
SHARE THIS PAGE