Hi all, Sorry for being late to the party. I went through the SPIP doc and I think this is a great proposal! I left a comment in the SPIP doc a couple days ago, but I don't see much activity there and no one replied, so I wanted to cross-post it here to get some feedback.
I'm Shiqi Sun, and I work for Big Data Platform in Salesforce. My team has been running the Spark on k8s operator <https://github.com/GoogleCloudPlatform/spark-on-k8s-operator> (OSS from Google) in my company to serve Spark users on production for 4+ years, and we've been actively contributing to the Spark on k8s operator OSS and also, occasionally, the Spark OSS. According to our experience, Google's Spark Operator has its own problems, like its close coupling with the spark version, as well as the JVM overhead during job submission. However on the other side, it's been a great component in our team's service in the company, especially being written in golang, it's really easy to have it interact with k8s, and also its CRD covers a lot of different use cases, as it has been built up through time thanks to many users' contribution during these years. There were also a handful of sessions of Google's Spark Operator Spark Summit that made it widely adopted. For this SPIP, I really love the idea of this proposal for the official k8s operator of Spark project, as well as the separate layer of the submission worker and being spark version agnostic. I think we can get the best of the two: 1. I would advocate the new project to still use golang for the implementation, as golang is the go-to cloud native language that works the best with k8s. 2. We make sure the functionality of the current Google's spark operator CRD is preserved in the new official Spark Operator; if we can make it compatible or even merge the two projects to make it the new official operator in spark project, it would be the best. 3. The new Spark Operator should continue being spark agnostic and continue having this lightweight/separate layer of submission worker. We've seen scalability issues caused by the heavy JVM during spark-submit in Google's Spark Operator and we implemented an internal version of fix for it within our company. We can continue the discussion in more detail, but generally I love this move of the official spark operator, and I really appreciate the effort! In the SPIP doc. I see my comment has gained several upvotes from someone I don't know, so I believe there are other spark/spark operator users who agree with some of my points. Let me know what you all think and let's continue the discussion, so that we can make this operator a great new component of the Open Source Spark Project! Thanks! Shiqi On Mon, Nov 13, 2023 at 11:50 PM L. C. Hsieh <vii...@gmail.com> wrote: > Thanks for all the support from the community for the SPIP proposal. > > Since all questions/discussion are settled down (if I didn't miss any > major ones), if no more questions or concerns, I'll be the shepherd > for this SPIP proposal and call for a vote tomorrow. > > Thank you all! > > On Mon, Nov 13, 2023 at 6:43 PM Zhou Jiang <zhou.c.ji...@gmail.com> wrote: > > > > Hi Holden, > > > > Thanks a lot for your feedback! > > Yes, this proposal attempts to integrate existing solutions, especially > from CRD perspective. The proposed schema retains similarity with current > designs, while reducing duplicates and maintaining a single source of truth > from conf properties. It also tends to be close to native integration with > k8s to minimize schema changes for new features. > > For dependencies, packing everything is the easiest way to get started. > It would be straightforward to add --packages and --repositories support > for Maven dependencies. It's technically possible to pull dependencies in > cloud storage from init containers (if defined by user). It could be tricky > to design a general solution that supports different cloud providers from > the operator layer. An enhancement that I can think of is to add support > for profile scripts that can enable additional user-defined actions in > application containers. > > Operator does not have to build everything for k8s version > compatibility. Similar to Spark, operator can be built on Fabric8 client( > https://github.com/fabric8io/kubernetes-client) for support across > versions, given that it makes similar API calls for resource management as > Spark. For tests, in addition to fabric8 mock server, we may also borrow > the idea from Flink operator to start minikube cluster for integration > tests. > > This operator is not starting from scratch as it is derived from an > internal project which has been working in prod scale for a few years. It > aims to include a few new features / enhancements, and a few > re-architecture mostly to incorporate lessons learnt for designing CRD / > API perspective. > > Benchmarking operator performance alone can be nuanced, often tied to > the underlying cluster. There's a testing strategy that Aaruna & I > discussed in a previous Data AI summit, involves scheduling wide (massive > light-weight applications) and deep (single application request a lot of > executors with heavy IO) cases, revealing typical bottlenecks at the k8s > API server and scheduler performance.Similar tests can be performed for > this as well. > > > > On Sun, Nov 12, 2023 at 4:32 PM Holden Karau <hol...@pigscanfly.ca> > wrote: > >> > >> To be clear: I am generally supportive of the idea (+1) but have some > follow-up questions: > >> > >> Have we taken the time to learn from the other operators? Do we have a > compatible CRD/API or not (and if so why?) > >> The API seems to assume that everything is packaged in the container in > advance, but I imagine that might not be the case for many folks who have > Java or Python packages published to cloud storage and they want to use? > >> What's our plan for the testing on the potential version explosion (not > tying ourselves to operator version -> spark version makes a lot of sense, > but how do we reasonably assure ourselves that the cross product of > Operator Version, Kube Version, and Spark Version all function)? Do we have > CI resources for this? > >> Is there a current (non-open source operator) that folks from Apple are > using and planning to open source, or is this a fresh "from the ground up" > operator proposal? > >> One of the key reasons for this is listed as "An out-of-the-box > automation solution that scales effectively" but I don't see any discussion > of the target scale or plans to achieve it? > >> > >> > >> > >> On Thu, Nov 9, 2023 at 9:02 PM Zhou Jiang <zhou.c.ji...@gmail.com> > wrote: > >>> > >>> Hi Spark community, > >>> > >>> I'm reaching out to initiate a conversation about the possibility of > developing a Java-based Kubernetes operator for Apache Spark. Following the > operator pattern ( > https://kubernetes.io/docs/concepts/extend-kubernetes/operator/), Spark > users may manage applications and related components seamlessly using > native tools like kubectl. The primary goal is to simplify the Spark user > experience on Kubernetes, minimizing the learning curve and operational > complexities and therefore enable users to focus on the Spark application > development. > >>> > >>> Although there are several open-source Spark on Kubernetes operators > available, none of them are officially integrated into the Apache Spark > project. As a result, these operators may lack active support and > development for new features. Within this proposal, our aim is to introduce > a Java-based Spark operator as an integral component of the Apache Spark > project. This solution has been employed internally at Apple for multiple > years, operating millions of executors in real production environments. The > use of Java in this solution is intended to accommodate a wider user and > contributor audience, especially those who are familiar with Scala. > >>> > >>> Ideally, this operator should have its dedicated repository, similar > to Spark Connect Golang or Spark Docker, allowing it to maintain a loose > connection with the Spark release cycle. This model is also followed by the > Apache Flink Kubernetes operator. > >>> > >>> We believe that this project holds the potential to evolve into a > thriving community project over the long run. A comparison can be drawn > with the Flink Kubernetes Operator: Apple has open-sourced internal Flink > Kubernetes operator, making it a part of the Apache Flink project ( > https://github.com/apache/flink-kubernetes-operator). This move has > gained wide industry adoption and contributions from the community. In a > mere year, the Flink operator has garnered more than 600 stars and has > attracted contributions from over 80 contributors. This showcases the level > of community interest and collaborative momentum that can be achieved in > similar scenarios. > >>> > >>> More details can be found at SPIP doc : Spark Kubernetes Operator > https://docs.google.com/document/d/1f5mm9VpSKeWC72Y9IiKN2jbBn32rHxjWKUfLRaGEcLE > >>> > >>> Thanks, > >>> > >>> -- > >>> Zhou JIANG > >>> > >> > >> > >> -- > >> Twitter: https://twitter.com/holdenkarau > >> Books (Learning Spark, High Performance Spark, etc.): > https://amzn.to/2MaRAG9 > >> YouTube Live Streams: https://www.youtube.com/user/holdenkarau > > > > > > > > -- > > Zhou JIANG > > > > --------------------------------------------------------------------- > To unsubscribe e-mail: dev-unsubscr...@spark.apache.org > >
