Using Spark to speed up the diagnosis performance for big data applications

Description

Cosmos is Microsoft’s internal big data analysis platform. Everyday, it processes huge numbers of data from Microsoft services like Bing, Office, Windows, Xbox, Dynamics, etc. The DevOps team is responsible for keeping the service reliable. For each live site issue, the on-call engineer has a hard deadline to mitigate the problem. The Microsoft big data team has been working on bringing integrated development environment- (IDE) style diagnosis experience to large-scale applications. However, there were several challenges for on-call engineers to use the team’s IDE diagnosis tools: it’s slow to process complex jobs with large profiles, and the IDE may crash for jobs with a profile larger than 10G; the team provides an auto-diagnosis wizard for common issues but on-call engineers still need to digger deeper into various logging systems case by case; and it requires extra effort for on-call engineers to document their troubleshooting steps.

Ruixin Xu, Long Tian, and Yu Zhou outline an experiment run to solve these challenges, replacing the diagnosis engine with Spark and using the Jupyter notebook as the frontend. The experiment results indicate that the Spark-based solution has improved the diagnosis performance significantly, especially for complex jobs with a large profile. The Jupyter notebook also brings the benefit of fast iteration and easy knowledge share. They also outline the insights they’ve gained on the journey.