Article summary
Simple Testing Can Prevent Most Critical Failures: An Analysis of Production Failures in Distributed Data-Intensive Systems
This post reviews a study by Ding Yuan and colleagues of 198 user-reported production failures in Cassandra, HBase, HDFS, Hadoop MapReduce and Redis. The study found that almost all failures need three or fewer nodes to reproduce, that most need more than one input event in a particular order, and that the error logs usually record the triggering events. It also found that 92% of the catastrophic failures came from incorrect handling of non-fatal errors, and that simple testing of error handling code could have prevented most of them. The authors' static checker, Aspirator, located 143 bugs and bad practices in nine distributed systems that were fixed or confirmed by their developers.
Read the original on It Will Never Work in Theory Archived copy
Omnifeed summarized this article with AI from its text (an archived copy), checked . Omnifeed isn’t affiliated with It Will Never Work in Theory. How we use AI