In Spark, the shuffle primitive requires Spark executors to persist data to the local disk of the worker nodes. If executors crash, the external shuffle service can continue to serve the shuffle data that was written beyond the lifetime of the executor itself. In YARN, Mesos, and Standalone mode, the external shuffle service is deployed on every worker node. The shuffle service shares local disk with the executors that run on its node.
There are some shortcomings with the way shuffle is fundamentally implemented right now. Particularly:
- If any external shuffle service process or node becomes unavailable, all applications that had an executor that ran on that node must recompute the shuffle blocks that were lost.
- Similarly to the above, the external shuffle service must be kept running at all times, which may waste resources when no applications are using that shuffle service node.
- Mounting local storage can prevent users from taking advantage of desirable isolation benefits from using containerized environments, like Kubernetes. We had an external shuffle service implementation in an early prototype of the Kubernetes backend, but it was rejected due to its strict requirement to be able to mount hostPath volumes or other persistent volume setups.
In the following architecture discussion document (note: not an SPIP), we brainstorm various high level architectures for improving the external shuffle service in a way that addresses the above problems. The purpose of this umbrella JIRA is to promote additional discussion on how we can approach these problems, both at the architecture level and the implementation level. We anticipate filing sub-issues that break down the tasks that must be completed to achieve this goal.
Edit June 28 2019: Our SPIP is here: https://docs.google.com/document/d/1d6egnL6WHOwWZe8MWv3m8n4PToNacdx7n_0iMSWwhCQ/edit