Hi Ofir, I can provide some information about use cases for Apache Celeborn.
Apache Celeborn can be deployed on K8s and standalone, both are widely used in production environment by users. The largest cluster I know contains more than 1,000 Celeborn workers. Celeborn is specially beneficial for large scale shuffle with high parallelism, which usually causes long fetch wait time or even fetch failure. We have seen serveral times speedup for jobs with large scale shuffle. Besides, with Celeborn, Spark on K8s can achive better Dynamic Resource Allocation because executors don't need to store shuffle data locally, also the pods don't need a large disk space. Celeborn is relatively easy to operate, especially for its graceful rolling upgrade and backward compatibility (across two successive versions). You can find more information including user feedbacks here[1]. I recommend you to try it out, and the community is happy to help :) Regards, Keyong Zhou [1] https://news.apache.org/foundation/entry/apache-software-foundation-announces-new-top-level-project-apache-celeborn On 2024/06/06 09:08:31 Ofir Manor wrote: > Hi, > Regarding the external shuffle service on K8S and especially the push-based > variant that was merged in 3.2: > > 1. > Are there plans to make it supported and work out-of-the-box in 4.0? > 2. > Did anyone make it work for themselves in 3.5 or earlier? If so, can you > share your experience and what was needed to make it work? > > As a fallback, someone using one of the new shuffle projects with K8S such as > Apache Uniffle or Apache Celeborn and can share some feedback? Performance, > stability, added complexity etc? > Thanks, > Ofir > --------------------------------------------------------------------- To unsubscribe e-mail: dev-unsubscr...@spark.apache.org