Re: Flink native k8s integration vs. operator
Hi Robert, I agree with you, I mean that's why I was writing a K8s operator, but the restriction wasn't decided by me, it was imposed on me. I guess my thinking was rather that an operator wouldn't necessarily supersede standalone+reactive, at least not in my case, but that certainly doesn't mean an operator is a bad idea, it's just something that other users might want to keep in mind. Regards, Alexis. From: Robert Metzger Sent: Thursday, January 20, 2022 7:06 PM To: Alexis Sarda-Espinosa Cc: dev ; user Subject: Re: Flink native k8s integration vs. operator Hi Alexis, The usage of Custom Resource Definitions (CRDs). The main reason given to me was that such resources are global (for a given cluster) and that is not desired. I know that ultimately a CR based on a CRD can be scoped to a specific namespace, but customer is king… I don't think this restriction applies to many organizations. K8s operators are the de facto standard for deploying all kinds of software. There are quite many projects that used to just have a Helm chart, that are now switching over to provide operators, because they provide a much better experience. If you have more specifics on this concern that is relevant for the Flink community, I'd like to hear that. Kubernetes Service Accounts (SAs) with roles to create deployments/pods. This one is more understandable, particularly after the whole log4j debacle. Roles that manage solely deployment.scale subresources would be acceptable though. This requirement is not strictly needed to deploy Flink on K8s. Only with the native K8s integration of Flink, you need to give the Flink JVM a role that allows creating other pods. Best, Robert On Tue, Jan 18, 2022 at 5:18 PM Alexis Sarda-Espinosa mailto:alexis.sarda-espin...@microfocus.com>> wrote: Hi everyone, Since I see this is getting some traction, I’d like to add a couple things. I had been developing a Kubernetes controller for Flink as a Proof of Concept at my company; I called it Flork because it was to be a Flink Orchestrator for Kubernetes. In the end, we will most likely not use this controller due to security concerns that were communicated to me. These concerns stem from the fact that our product would be used by customers in their own Kubernetes clusters, and many customers don’t want: - The usage of Custom Resource Definitions (CRDs). The main reason given to me was that such resources are global (for a given cluster) and that is not desired. I know that ultimately a CR based on a CRD can be scoped to a specific namespace, but customer is king… - Kubernetes Service Accounts (SAs) with roles to create deployments/pods. This one is more understandable, particularly after the whole log4j debacle. Roles that manage solely deployment.scale subresources would be acceptable though. I mention these in case they prove to be relevant for others in the current context. For us, it means we have to stick with something like standalone Kubernetes + reactive/adaptive. Nevertheless, the PoC I had was already functional and, while I would have to request permission to contribute the code to the community, it might be useful for these efforts. However, I’d first ask if there is actually interest in this code, considering these are some of the “features” it currently has: * The CRD relies on the Pod Template support included in Flink itself. As such, some of the fields in the CRD are “vanilla” pod specs, and the schema reflects that because it embeds a flattened version of the schema from [1]. I’d also have a basic Helm chart ready. * The code is written in a mixture of Java and Kotlin, and is built with Gradle. I made heavy use of Kotlin Coroutines to implement some of the core logic in a non-blocking way. * The code already supports High Availability by leveraging Kubernetes leases and the corresponding helpers in Fabric8’s client. * The main deployment logic is delegated to Flink’s own flink-kubernetes module [2]. Nevertheless, my build shadows all the fabric8 classes and service definitions embedded in said module, so that the rest of the code can use other kubernetes-client versions independently. * The controller handles savepoint creation for redeployments after CR changes, e.g. upgrades. This would also work after controller fail-over with/without HA. * The code supports some extension for custom container images: classes defined in META-INF/services/ files are called as decorators for Flink’s conf file and/or the pod specs defined in the CR, and they could be copied to the image on top of an official base version. * A deployment mode without CRD could be supported --- I have some code that can run on top of the core controller and allows “embedding” a CR in a Config Map key. The translation between the CM and the core controller code is then done transparently. * I ha
Re: Flink native k8s integration vs. operator
ingress for Flink’s UI. > > > > Let me know what you think. > > > > [1] > https://github.com/kubernetes/kubernetes/blob/master/api/openapi-spec/swagger.json > > [2] https://github.com/apache/flink/tree/master/flink-kubernetes > > > > Regards, > > Alexis. > > > > *From:* Gyula Fóra > *Sent:* Montag, 17. Januar 2022 12:40 > *To:* dev > *Cc:* Xintong Song ; user > *Subject:* Re: Flink native k8s integration vs. operator > > > > Hi Yang! > > Thanks for the input! > > > > I agree with you on both points that you made. Even if we might support > both standalone and native modes in the long run, we should probably build > the first version on top of the native integration. > > This I feel will result in a much simpler, minimalistic first version that > will already support the most important features. We are familiar with your > PoC implementation and I think it's a great idea to use that as a base. > > > > As for Java / Go, I think Java is the obvious choice here. I would have to > think very hard to make any good arguments for picking Go :) > > > > Cheers, > > Gyula > > > > > > > > On Mon, Jan 17, 2022 at 10:30 AM Yang Wang wrote: > > Glad to see that the interest of this thread keeps going. And thanks > Thomas, Gyula, and Marton for driving this effort. > > I want to share my two cents about the Flink K8s operator. > > > Standalone deployment VS native K8s integration > > There is already some feature requirement issue[1] for the existing > GoogleCloudPlatform/flink-on-k8s-operator to support native K8s > integration. So I think > it will be great if the new introduced K8s operator could support native > K8s mode. I could imagine some advantages for using native mode. e.g. > dynamic allocation, > stability improvement, etc. > > Compared with standalone + reactive mode, the native K8s could not > integrate with auto-scaling(allocate/remove TaskManager pods based on > metrics) well. > Since the reconcile behavior for standalone and native K8s mode will be > different, I am not sure whether we will support them both at the very > beginning. > > > > Go VS Java > > Although most of the K8s operators are developed in Go, which could benefit > from the prosperous ecosystem and various tools. I lean to develop the K8s > operator under Flink umbrella using Java. > Then the Flink contributors will be easier to get involved. We could use > the same Kubernetes Java client with Flink. When Flink exposes some public > deployment interfaces(e.g. ApplicationDeployer) > in the future, the K8s operator will also benefit a lot from this. > > I already have a simple PoC project of this implementation[2]. Hope you > could get some inspirations from this. > > > [1]. > https://github.com/GoogleCloudPlatform/flink-on-k8s-operator/issues/168 > [2]. https://github.com/wangyang0918/flink-native-k8s-operator > > > Best, > Yang > > > > Xintong Song 于2022年1月14日周五 15:47写道： > > > Thanks for volunteering to drive this effort, Marton, Thomas and Gyula. > > > > Looking forward to the public discussion. Please feel free to reach out > if > > there's anything you need from us. > > > > Thank you~ > > > > Xintong Song > > > > > > > > On Fri, Jan 14, 2022 at 8:27 AM Chenya Zhang < > chenyazhangche...@gmail.com> > > wrote: > > > >> Thanks Thomas, Gyula, and Marton for driving this effort! It would > >> greatly ease the adoption of Apache Flink on Kubernetes and help to > address > >> the current operational pain points as mentioned. Look forward to the > >> proposal and more discussions! > >> > >> Best, > >> Chenya > >> > >> On Thu, Jan 13, 2022 at 12:15 PM Márton Balassi < > balassi.mar...@gmail.com> > >> wrote: > >> > >>> Hi All, > >>> > >>> I am pleased to see the level of enthusiasm and technical consideration > >>> already emerging in this thread. I wholeheartedly support building an > >>> operator and endorsing it via placing it under the Apache Flink > umbrella > >>> (as a separate repository) as the current lack of it is clearly > becoming > >>> an > >>> adoption bottleneck for large scale Flink users. The next logical step > is > >>> to write a FLIP to agree on the technical details, so that we can put > >>> forward the proposal to the Flink PMC for creating a new repository > with > >>> a > >>> clear purpose in mind. I volunteer to work wit
Re: Flink native k8s integration vs. operator
Hi Yang! Thanks for the input! I agree with you on both points that you made. Even if we might support both standalone and native modes in the long run, we should probably build the first version on top of the native integration. This I feel will result in a much simpler, minimalistic first version that will already support the most important features. We are familiar with your PoC implementation and I think it's a great idea to use that as a base. As for Java / Go, I think Java is the obvious choice here. I would have to think very hard to make any good arguments for picking Go :) Cheers, Gyula On Mon, Jan 17, 2022 at 10:30 AM Yang Wang wrote: > Glad to see that the interest of this thread keeps going. And thanks > Thomas, Gyula, and Marton for driving this effort. > > I want to share my two cents about the Flink K8s operator. > > > Standalone deployment VS native K8s integration > > There is already some feature requirement issue[1] for the existing > GoogleCloudPlatform/flink-on-k8s-operator to support native K8s > integration. So I think > it will be great if the new introduced K8s operator could support native > K8s mode. I could imagine some advantages for using native mode. e.g. > dynamic allocation, > stability improvement, etc. > > Compared with standalone + reactive mode, the native K8s could not > integrate with auto-scaling(allocate/remove TaskManager pods based on > metrics) well. > Since the reconcile behavior for standalone and native K8s mode will be > different, I am not sure whether we will support them both at the very > beginning. > > > > Go VS Java > > Although most of the K8s operators are developed in Go, which could benefit > from the prosperous ecosystem and various tools. I lean to develop the K8s > operator under Flink umbrella using Java. > Then the Flink contributors will be easier to get involved. We could use > the same Kubernetes Java client with Flink. When Flink exposes some public > deployment interfaces(e.g. ApplicationDeployer) > in the future, the K8s operator will also benefit a lot from this. > > I already have a simple PoC project of this implementation[2]. Hope you > could get some inspirations from this. > > > [1]. > https://github.com/GoogleCloudPlatform/flink-on-k8s-operator/issues/168 > [2]. https://github.com/wangyang0918/flink-native-k8s-operator > > > Best, > Yang > > > > Xintong Song 于2022年1月14日周五 15:47写道： > > > Thanks for volunteering to drive this effort, Marton, Thomas and Gyula. > > > > Looking forward to the public discussion. Please feel free to reach out > if > > there's anything you need from us. > > > > Thank you~ > > > > Xintong Song > > > > > > > > On Fri, Jan 14, 2022 at 8:27 AM Chenya Zhang < > chenyazhangche...@gmail.com> > > wrote: > > > >> Thanks Thomas, Gyula, and Marton for driving this effort! It would > >> greatly ease the adoption of Apache Flink on Kubernetes and help to > address > >> the current operational pain points as mentioned. Look forward to the > >> proposal and more discussions! > >> > >> Best, > >> Chenya > >> > >> On Thu, Jan 13, 2022 at 12:15 PM Márton Balassi < > balassi.mar...@gmail.com> > >> wrote: > >> > >>> Hi All, > >>> > >>> I am pleased to see the level of enthusiasm and technical consideration > >>> already emerging in this thread. I wholeheartedly support building an > >>> operator and endorsing it via placing it under the Apache Flink > umbrella > >>> (as a separate repository) as the current lack of it is clearly > becoming > >>> an > >>> adoption bottleneck for large scale Flink users. The next logical step > is > >>> to write a FLIP to agree on the technical details, so that we can put > >>> forward the proposal to the Flink PMC for creating a new repository > with > >>> a > >>> clear purpose in mind. I volunteer to work with Thomas and Gyula on the > >>> initial wording on the proposal which we will put up for public > >>> discussion > >>> in the coming weeks. > >>> > >>> Best, > >>> Marton > >>> > >>> On Thu, Jan 13, 2022 at 9:22 AM Konstantin Knauf > >>> wrote: > >>> > >>> > Hi Thomas, > >>> > > >>> > Yes, I was referring to a separate repository under Apache Flink. > >>> > > >>> > Cheers, > >>> > > >>> > Konstantin > >>> > > >>> > On Thu, Jan 13, 2022 at 6:19 AM Thomas Weise wrote: > >>> > > >>> >> Hi everyone, > >>> >> > >>> >> Thanks for the feedback and discussion. A few additional thoughts: > >>> >> > >>> >> [Konstantin] > With respect to common lifecycle management > operations: > >>> >> these features are > >>> >> > not available (within Apache Flink) for any of the other resource > >>> >> providers > >>> >> > (YARN, Standalone) either. From this perspective, I wouldn't > >>> consider > >>> >> this > >>> >> > a shortcoming of the Kubernetes integration. > >>> >> > >>> >> I think time and evolution of the ecosystem are factors to consider > as > >>> >> well. The state and usage of Flink was much different when YARN > >>> >> integration was novel. Expectations are different today and the
Re: Flink native k8s integration vs. operator
Glad to see that the interest of this thread keeps going. And thanks Thomas, Gyula, and Marton for driving this effort. I want to share my two cents about the Flink K8s operator. > Standalone deployment VS native K8s integration There is already some feature requirement issue[1] for the existing GoogleCloudPlatform/flink-on-k8s-operator to support native K8s integration. So I think it will be great if the new introduced K8s operator could support native K8s mode. I could imagine some advantages for using native mode. e.g. dynamic allocation, stability improvement, etc. Compared with standalone + reactive mode, the native K8s could not integrate with auto-scaling(allocate/remove TaskManager pods based on metrics) well. Since the reconcile behavior for standalone and native K8s mode will be different, I am not sure whether we will support them both at the very beginning. > Go VS Java Although most of the K8s operators are developed in Go, which could benefit from the prosperous ecosystem and various tools. I lean to develop the K8s operator under Flink umbrella using Java. Then the Flink contributors will be easier to get involved. We could use the same Kubernetes Java client with Flink. When Flink exposes some public deployment interfaces(e.g. ApplicationDeployer) in the future, the K8s operator will also benefit a lot from this. I already have a simple PoC project of this implementation[2]. Hope you could get some inspirations from this. [1]. https://github.com/GoogleCloudPlatform/flink-on-k8s-operator/issues/168 [2]. https://github.com/wangyang0918/flink-native-k8s-operator Best, Yang Xintong Song 于2022年1月14日周五 15:47写道： > Thanks for volunteering to drive this effort, Marton, Thomas and Gyula. > > Looking forward to the public discussion. Please feel free to reach out if > there's anything you need from us. > > Thank you~ > > Xintong Song > > > > On Fri, Jan 14, 2022 at 8:27 AM Chenya Zhang > wrote: > >> Thanks Thomas, Gyula, and Marton for driving this effort! It would >> greatly ease the adoption of Apache Flink on Kubernetes and help to address >> the current operational pain points as mentioned. Look forward to the >> proposal and more discussions! >> >> Best, >> Chenya >> >> On Thu, Jan 13, 2022 at 12:15 PM Márton Balassi >> wrote: >> >>> Hi All, >>> >>> I am pleased to see the level of enthusiasm and technical consideration >>> already emerging in this thread. I wholeheartedly support building an >>> operator and endorsing it via placing it under the Apache Flink umbrella >>> (as a separate repository) as the current lack of it is clearly becoming >>> an >>> adoption bottleneck for large scale Flink users. The next logical step is >>> to write a FLIP to agree on the technical details, so that we can put >>> forward the proposal to the Flink PMC for creating a new repository with >>> a >>> clear purpose in mind. I volunteer to work with Thomas and Gyula on the >>> initial wording on the proposal which we will put up for public >>> discussion >>> in the coming weeks. >>> >>> Best, >>> Marton >>> >>> On Thu, Jan 13, 2022 at 9:22 AM Konstantin Knauf >>> wrote: >>> >>> > Hi Thomas, >>> > >>> > Yes, I was referring to a separate repository under Apache Flink. >>> > >>> > Cheers, >>> > >>> > Konstantin >>> > >>> > On Thu, Jan 13, 2022 at 6:19 AM Thomas Weise wrote: >>> > >>> >> Hi everyone, >>> >> >>> >> Thanks for the feedback and discussion. A few additional thoughts: >>> >> >>> >> [Konstantin] > With respect to common lifecycle management operations: >>> >> these features are >>> >> > not available (within Apache Flink) for any of the other resource >>> >> providers >>> >> > (YARN, Standalone) either. From this perspective, I wouldn't >>> consider >>> >> this >>> >> > a shortcoming of the Kubernetes integration. >>> >> >>> >> I think time and evolution of the ecosystem are factors to consider as >>> >> well. The state and usage of Flink was much different when YARN >>> >> integration was novel. Expectations are different today and the >>> >> lifecycle functionality provided by an operator may as well be >>> >> considered essential to support the concept of a Flink application on >>> >> k8s. After few years learning from operator experience outside of >>> >> Flink it might be a good time to fill the gap. >>> >> >>> >> [Konstantin] > I still believe that we should keep this focus on low >>> >> > level composable building blocks (like Jobs and Snapshots) in Apache >>> >> Flink >>> >> > to make it easy for everyone to build fitting higher level >>> abstractions >>> >> > like a FlinkApplication Custom Resource on top of it. >>> >> >>> >> I completely agree that it is important that the basic functions of >>> >> Flink are solid and continued focus is necessary. Thanks for sharing >>> >> the pointers, these are great improvements. At the same time, >>> >> ecosystem, contributor base and user spectrum are growing. There have >>> >> been significant additions in many areas of Flink including
Re: Flink native k8s integration vs. operator
Thanks for volunteering to drive this effort, Marton, Thomas and Gyula. Looking forward to the public discussion. Please feel free to reach out if there's anything you need from us. Thank you~ Xintong Song On Fri, Jan 14, 2022 at 8:27 AM Chenya Zhang wrote: > Thanks Thomas, Gyula, and Marton for driving this effort! It would greatly > ease the adoption of Apache Flink on Kubernetes and help to address the > current operational pain points as mentioned. Look forward to the proposal > and more discussions! > > Best, > Chenya > > On Thu, Jan 13, 2022 at 12:15 PM Márton Balassi > wrote: > >> Hi All, >> >> I am pleased to see the level of enthusiasm and technical consideration >> already emerging in this thread. I wholeheartedly support building an >> operator and endorsing it via placing it under the Apache Flink umbrella >> (as a separate repository) as the current lack of it is clearly becoming >> an >> adoption bottleneck for large scale Flink users. The next logical step is >> to write a FLIP to agree on the technical details, so that we can put >> forward the proposal to the Flink PMC for creating a new repository with a >> clear purpose in mind. I volunteer to work with Thomas and Gyula on the >> initial wording on the proposal which we will put up for public discussion >> in the coming weeks. >> >> Best, >> Marton >> >> On Thu, Jan 13, 2022 at 9:22 AM Konstantin Knauf >> wrote: >> >> > Hi Thomas, >> > >> > Yes, I was referring to a separate repository under Apache Flink. >> > >> > Cheers, >> > >> > Konstantin >> > >> > On Thu, Jan 13, 2022 at 6:19 AM Thomas Weise wrote: >> > >> >> Hi everyone, >> >> >> >> Thanks for the feedback and discussion. A few additional thoughts: >> >> >> >> [Konstantin] > With respect to common lifecycle management operations: >> >> these features are >> >> > not available (within Apache Flink) for any of the other resource >> >> providers >> >> > (YARN, Standalone) either. From this perspective, I wouldn't consider >> >> this >> >> > a shortcoming of the Kubernetes integration. >> >> >> >> I think time and evolution of the ecosystem are factors to consider as >> >> well. The state and usage of Flink was much different when YARN >> >> integration was novel. Expectations are different today and the >> >> lifecycle functionality provided by an operator may as well be >> >> considered essential to support the concept of a Flink application on >> >> k8s. After few years learning from operator experience outside of >> >> Flink it might be a good time to fill the gap. >> >> >> >> [Konstantin] > I still believe that we should keep this focus on low >> >> > level composable building blocks (like Jobs and Snapshots) in Apache >> >> Flink >> >> > to make it easy for everyone to build fitting higher level >> abstractions >> >> > like a FlinkApplication Custom Resource on top of it. >> >> >> >> I completely agree that it is important that the basic functions of >> >> Flink are solid and continued focus is necessary. Thanks for sharing >> >> the pointers, these are great improvements. At the same time, >> >> ecosystem, contributor base and user spectrum are growing. There have >> >> been significant additions in many areas of Flink including connectors >> >> and higher level abstractions like statefun, SQL and Python. It's also >> >> evident from additional repositories/subprojects that we have in Flink >> >> today. >> >> >> >> [Konstantin] > Having said this, if others in the community have the >> >> capacity to push and >> >> > *maintain* a somewhat minimal "reference" Kubernetes Operator for >> Apache >> >> > Flink, I don't see any blockers. If or when this happens, I'd see >> some >> >> > clear benefits of using a separate repository (easier independent >> >> > versioning and releases, different build system & tooling (go, I >> >> assume)). >> >> >> >> Naturally different contributors to the project have different focus. >> >> Let's find out if there is strong enough interest to take this on and >> >> strong enough commitment to maintain. As I see it, there is a >> >> tremendous amount of internal investment going into operationalizing >> >> Flink within many companies. Improvements to the operational side of >> >> Flink like the operator would complement Flink nicely. I assume that >> >> you are referring to a separate repository within Apache Flink, which >> >> would give it the chance to achieve better sustainability than the >> >> existing external operator efforts. There is also the fact that some >> >> organizations which are heavily invested in operationalizing Flink are >> >> allowing contributing to Apache Flink itself but less so to arbitrary >> >> github projects. Regarding the tooling, it could well turn out that >> >> Java is a good alternative given the ecosystem focus and that there is >> >> an opportunity for reuse in certain aspects (metrics, logging etc.). >> >> >> >> [Yang] > I think Xintong has given a strong point why we introduced >> >> the native K8s in
Re: Flink native k8s integration vs. operator
Thanks Thomas, Gyula, and Marton for driving this effort! It would greatly ease the adoption of Apache Flink on Kubernetes and help to address the current operational pain points as mentioned. Look forward to the proposal and more discussions! Best, Chenya On Thu, Jan 13, 2022 at 12:15 PM Márton Balassi wrote: > Hi All, > > I am pleased to see the level of enthusiasm and technical consideration > already emerging in this thread. I wholeheartedly support building an > operator and endorsing it via placing it under the Apache Flink umbrella > (as a separate repository) as the current lack of it is clearly becoming an > adoption bottleneck for large scale Flink users. The next logical step is > to write a FLIP to agree on the technical details, so that we can put > forward the proposal to the Flink PMC for creating a new repository with a > clear purpose in mind. I volunteer to work with Thomas and Gyula on the > initial wording on the proposal which we will put up for public discussion > in the coming weeks. > > Best, > Marton > > On Thu, Jan 13, 2022 at 9:22 AM Konstantin Knauf > wrote: > > > Hi Thomas, > > > > Yes, I was referring to a separate repository under Apache Flink. > > > > Cheers, > > > > Konstantin > > > > On Thu, Jan 13, 2022 at 6:19 AM Thomas Weise wrote: > > > >> Hi everyone, > >> > >> Thanks for the feedback and discussion. A few additional thoughts: > >> > >> [Konstantin] > With respect to common lifecycle management operations: > >> these features are > >> > not available (within Apache Flink) for any of the other resource > >> providers > >> > (YARN, Standalone) either. From this perspective, I wouldn't consider > >> this > >> > a shortcoming of the Kubernetes integration. > >> > >> I think time and evolution of the ecosystem are factors to consider as > >> well. The state and usage of Flink was much different when YARN > >> integration was novel. Expectations are different today and the > >> lifecycle functionality provided by an operator may as well be > >> considered essential to support the concept of a Flink application on > >> k8s. After few years learning from operator experience outside of > >> Flink it might be a good time to fill the gap. > >> > >> [Konstantin] > I still believe that we should keep this focus on low > >> > level composable building blocks (like Jobs and Snapshots) in Apache > >> Flink > >> > to make it easy for everyone to build fitting higher level > abstractions > >> > like a FlinkApplication Custom Resource on top of it. > >> > >> I completely agree that it is important that the basic functions of > >> Flink are solid and continued focus is necessary. Thanks for sharing > >> the pointers, these are great improvements. At the same time, > >> ecosystem, contributor base and user spectrum are growing. There have > >> been significant additions in many areas of Flink including connectors > >> and higher level abstractions like statefun, SQL and Python. It's also > >> evident from additional repositories/subprojects that we have in Flink > >> today. > >> > >> [Konstantin] > Having said this, if others in the community have the > >> capacity to push and > >> > *maintain* a somewhat minimal "reference" Kubernetes Operator for > Apache > >> > Flink, I don't see any blockers. If or when this happens, I'd see some > >> > clear benefits of using a separate repository (easier independent > >> > versioning and releases, different build system & tooling (go, I > >> assume)). > >> > >> Naturally different contributors to the project have different focus. > >> Let's find out if there is strong enough interest to take this on and > >> strong enough commitment to maintain. As I see it, there is a > >> tremendous amount of internal investment going into operationalizing > >> Flink within many companies. Improvements to the operational side of > >> Flink like the operator would complement Flink nicely. I assume that > >> you are referring to a separate repository within Apache Flink, which > >> would give it the chance to achieve better sustainability than the > >> existing external operator efforts. There is also the fact that some > >> organizations which are heavily invested in operationalizing Flink are > >> allowing contributing to Apache Flink itself but less so to arbitrary > >> github projects. Regarding the tooling, it could well turn out that > >> Java is a good alternative given the ecosystem focus and that there is > >> an opportunity for reuse in certain aspects (metrics, logging etc.). > >> > >> [Yang] > I think Xintong has given a strong point why we introduced > >> the native K8s integration, which is active resource management. > >> > I have a concrete example for this in the production. When a K8s node > >> is down, the standalone K8s deployment will take longer > >> > recovery time based on the K8s eviction time(IIRC, default is 5 > >> minutes). For the native K8s integration, Flink RM could be aware of the > >> > TM heartbeat lost and allocate a new one
Re: Flink native k8s integration vs. operator
Hi All, I am pleased to see the level of enthusiasm and technical consideration already emerging in this thread. I wholeheartedly support building an operator and endorsing it via placing it under the Apache Flink umbrella (as a separate repository) as the current lack of it is clearly becoming an adoption bottleneck for large scale Flink users. The next logical step is to write a FLIP to agree on the technical details, so that we can put forward the proposal to the Flink PMC for creating a new repository with a clear purpose in mind. I volunteer to work with Thomas and Gyula on the initial wording on the proposal which we will put up for public discussion in the coming weeks. Best, Marton On Thu, Jan 13, 2022 at 9:22 AM Konstantin Knauf wrote: > Hi Thomas, > > Yes, I was referring to a separate repository under Apache Flink. > > Cheers, > > Konstantin > > On Thu, Jan 13, 2022 at 6:19 AM Thomas Weise wrote: > >> Hi everyone, >> >> Thanks for the feedback and discussion. A few additional thoughts: >> >> [Konstantin] > With respect to common lifecycle management operations: >> these features are >> > not available (within Apache Flink) for any of the other resource >> providers >> > (YARN, Standalone) either. From this perspective, I wouldn't consider >> this >> > a shortcoming of the Kubernetes integration. >> >> I think time and evolution of the ecosystem are factors to consider as >> well. The state and usage of Flink was much different when YARN >> integration was novel. Expectations are different today and the >> lifecycle functionality provided by an operator may as well be >> considered essential to support the concept of a Flink application on >> k8s. After few years learning from operator experience outside of >> Flink it might be a good time to fill the gap. >> >> [Konstantin] > I still believe that we should keep this focus on low >> > level composable building blocks (like Jobs and Snapshots) in Apache >> Flink >> > to make it easy for everyone to build fitting higher level abstractions >> > like a FlinkApplication Custom Resource on top of it. >> >> I completely agree that it is important that the basic functions of >> Flink are solid and continued focus is necessary. Thanks for sharing >> the pointers, these are great improvements. At the same time, >> ecosystem, contributor base and user spectrum are growing. There have >> been significant additions in many areas of Flink including connectors >> and higher level abstractions like statefun, SQL and Python. It's also >> evident from additional repositories/subprojects that we have in Flink >> today. >> >> [Konstantin] > Having said this, if others in the community have the >> capacity to push and >> > *maintain* a somewhat minimal "reference" Kubernetes Operator for Apache >> > Flink, I don't see any blockers. If or when this happens, I'd see some >> > clear benefits of using a separate repository (easier independent >> > versioning and releases, different build system & tooling (go, I >> assume)). >> >> Naturally different contributors to the project have different focus. >> Let's find out if there is strong enough interest to take this on and >> strong enough commitment to maintain. As I see it, there is a >> tremendous amount of internal investment going into operationalizing >> Flink within many companies. Improvements to the operational side of >> Flink like the operator would complement Flink nicely. I assume that >> you are referring to a separate repository within Apache Flink, which >> would give it the chance to achieve better sustainability than the >> existing external operator efforts. There is also the fact that some >> organizations which are heavily invested in operationalizing Flink are >> allowing contributing to Apache Flink itself but less so to arbitrary >> github projects. Regarding the tooling, it could well turn out that >> Java is a good alternative given the ecosystem focus and that there is >> an opportunity for reuse in certain aspects (metrics, logging etc.). >> >> [Yang] > I think Xintong has given a strong point why we introduced >> the native K8s integration, which is active resource management. >> > I have a concrete example for this in the production. When a K8s node >> is down, the standalone K8s deployment will take longer >> > recovery time based on the K8s eviction time(IIRC, default is 5 >> minutes). For the native K8s integration, Flink RM could be aware of the >> > TM heartbeat lost and allocate a new one timely. >> >> Thanks for sharing this, we should evaluate it as part of a proposal. >> If we can optimize recovery or scaling with active resource management >> then perhaps it is worth to support it through the operator. >> Previously mentioned operators all rely on the standalone model. >> >> Cheers, >> Thomas >> >> On Wed, Jan 12, 2022 at 3:21 AM Konstantin Knauf >> wrote: >> > >> > cc dev@ >> > >> > Hi Thomas, Hi everyone, >> > >> > Thank you for starting this discussion and sorry for chiming
Re: Flink native k8s integration vs. operator
Hi Thomas, Yes, I was referring to a separate repository under Apache Flink. Cheers, Konstantin On Thu, Jan 13, 2022 at 6:19 AM Thomas Weise wrote: > Hi everyone, > > Thanks for the feedback and discussion. A few additional thoughts: > > [Konstantin] > With respect to common lifecycle management operations: > these features are > > not available (within Apache Flink) for any of the other resource > providers > > (YARN, Standalone) either. From this perspective, I wouldn't consider > this > > a shortcoming of the Kubernetes integration. > > I think time and evolution of the ecosystem are factors to consider as > well. The state and usage of Flink was much different when YARN > integration was novel. Expectations are different today and the > lifecycle functionality provided by an operator may as well be > considered essential to support the concept of a Flink application on > k8s. After few years learning from operator experience outside of > Flink it might be a good time to fill the gap. > > [Konstantin] > I still believe that we should keep this focus on low > > level composable building blocks (like Jobs and Snapshots) in Apache > Flink > > to make it easy for everyone to build fitting higher level abstractions > > like a FlinkApplication Custom Resource on top of it. > > I completely agree that it is important that the basic functions of > Flink are solid and continued focus is necessary. Thanks for sharing > the pointers, these are great improvements. At the same time, > ecosystem, contributor base and user spectrum are growing. There have > been significant additions in many areas of Flink including connectors > and higher level abstractions like statefun, SQL and Python. It's also > evident from additional repositories/subprojects that we have in Flink > today. > > [Konstantin] > Having said this, if others in the community have the > capacity to push and > > *maintain* a somewhat minimal "reference" Kubernetes Operator for Apache > > Flink, I don't see any blockers. If or when this happens, I'd see some > > clear benefits of using a separate repository (easier independent > > versioning and releases, different build system & tooling (go, I > assume)). > > Naturally different contributors to the project have different focus. > Let's find out if there is strong enough interest to take this on and > strong enough commitment to maintain. As I see it, there is a > tremendous amount of internal investment going into operationalizing > Flink within many companies. Improvements to the operational side of > Flink like the operator would complement Flink nicely. I assume that > you are referring to a separate repository within Apache Flink, which > would give it the chance to achieve better sustainability than the > existing external operator efforts. There is also the fact that some > organizations which are heavily invested in operationalizing Flink are > allowing contributing to Apache Flink itself but less so to arbitrary > github projects. Regarding the tooling, it could well turn out that > Java is a good alternative given the ecosystem focus and that there is > an opportunity for reuse in certain aspects (metrics, logging etc.). > > [Yang] > I think Xintong has given a strong point why we introduced > the native K8s integration, which is active resource management. > > I have a concrete example for this in the production. When a K8s node is > down, the standalone K8s deployment will take longer > > recovery time based on the K8s eviction time(IIRC, default is 5 > minutes). For the native K8s integration, Flink RM could be aware of the > > TM heartbeat lost and allocate a new one timely. > > Thanks for sharing this, we should evaluate it as part of a proposal. > If we can optimize recovery or scaling with active resource management > then perhaps it is worth to support it through the operator. > Previously mentioned operators all rely on the standalone model. > > Cheers, > Thomas > > On Wed, Jan 12, 2022 at 3:21 AM Konstantin Knauf > wrote: > > > > cc dev@ > > > > Hi Thomas, Hi everyone, > > > > Thank you for starting this discussion and sorry for chiming in late. > > > > I agree with Thomas' and David's assessment of Flink's "Native Kubernetes > > Integration", in particular, it does actually not integrate well with the > > Kubernetes ecosystem despite being called "native" (tooling, security > > concerns). > > > > With respect to common lifecycle management operations: these features > are > > not available (within Apache Flink) for any of the other resource > providers > > (YARN, Standalone) either. From this perspective, I wouldn't consider > this > > a shortcoming of the Kubernetes integration. Instead, we have been > focusing > > our efforts in Apache Flink on the operations of a single Job, and left > > orchestration and lifecycle management that spans multiple Jobs to > > ecosystem projects. I still believe that we should keep this focus on low > > level composable building blocks (like J
Re: Flink native k8s integration vs. operator
Hi everyone, Thanks for the feedback and discussion. A few additional thoughts: [Konstantin] > With respect to common lifecycle management operations: these features are > not available (within Apache Flink) for any of the other resource providers > (YARN, Standalone) either. From this perspective, I wouldn't consider this > a shortcoming of the Kubernetes integration. I think time and evolution of the ecosystem are factors to consider as well. The state and usage of Flink was much different when YARN integration was novel. Expectations are different today and the lifecycle functionality provided by an operator may as well be considered essential to support the concept of a Flink application on k8s. After few years learning from operator experience outside of Flink it might be a good time to fill the gap. [Konstantin] > I still believe that we should keep this focus on low > level composable building blocks (like Jobs and Snapshots) in Apache Flink > to make it easy for everyone to build fitting higher level abstractions > like a FlinkApplication Custom Resource on top of it. I completely agree that it is important that the basic functions of Flink are solid and continued focus is necessary. Thanks for sharing the pointers, these are great improvements. At the same time, ecosystem, contributor base and user spectrum are growing. There have been significant additions in many areas of Flink including connectors and higher level abstractions like statefun, SQL and Python. It's also evident from additional repositories/subprojects that we have in Flink today. [Konstantin] > Having said this, if others in the community have the capacity to push and > *maintain* a somewhat minimal "reference" Kubernetes Operator for Apache > Flink, I don't see any blockers. If or when this happens, I'd see some > clear benefits of using a separate repository (easier independent > versioning and releases, different build system & tooling (go, I assume)). Naturally different contributors to the project have different focus. Let's find out if there is strong enough interest to take this on and strong enough commitment to maintain. As I see it, there is a tremendous amount of internal investment going into operationalizing Flink within many companies. Improvements to the operational side of Flink like the operator would complement Flink nicely. I assume that you are referring to a separate repository within Apache Flink, which would give it the chance to achieve better sustainability than the existing external operator efforts. There is also the fact that some organizations which are heavily invested in operationalizing Flink are allowing contributing to Apache Flink itself but less so to arbitrary github projects. Regarding the tooling, it could well turn out that Java is a good alternative given the ecosystem focus and that there is an opportunity for reuse in certain aspects (metrics, logging etc.). [Yang] > I think Xintong has given a strong point why we introduced the native K8s integration, which is active resource management. > I have a concrete example for this in the production. When a K8s node is > down, the standalone K8s deployment will take longer > recovery time based on the K8s eviction time(IIRC, default is 5 minutes). For > the native K8s integration, Flink RM could be aware of the > TM heartbeat lost and allocate a new one timely. Thanks for sharing this, we should evaluate it as part of a proposal. If we can optimize recovery or scaling with active resource management then perhaps it is worth to support it through the operator. Previously mentioned operators all rely on the standalone model. Cheers, Thomas On Wed, Jan 12, 2022 at 3:21 AM Konstantin Knauf wrote: > > cc dev@ > > Hi Thomas, Hi everyone, > > Thank you for starting this discussion and sorry for chiming in late. > > I agree with Thomas' and David's assessment of Flink's "Native Kubernetes > Integration", in particular, it does actually not integrate well with the > Kubernetes ecosystem despite being called "native" (tooling, security > concerns). > > With respect to common lifecycle management operations: these features are > not available (within Apache Flink) for any of the other resource providers > (YARN, Standalone) either. From this perspective, I wouldn't consider this > a shortcoming of the Kubernetes integration. Instead, we have been focusing > our efforts in Apache Flink on the operations of a single Job, and left > orchestration and lifecycle management that spans multiple Jobs to > ecosystem projects. I still believe that we should keep this focus on low > level composable building blocks (like Jobs and Snapshots) in Apache Flink > to make it easy for everyone to build fitting higher level abstractions > like a FlinkApplication Custom Resource on top of it. For example, we are > currently contributing multiple improvements [1,2,3,4] to the REST API and > Application Mode that in our experience will make it easier to manage > A
Re: Flink native k8s integration vs. operator
cc dev@ Hi Thomas, Hi everyone, Thank you for starting this discussion and sorry for chiming in late. I agree with Thomas' and David's assessment of Flink's "Native Kubernetes Integration", in particular, it does actually not integrate well with the Kubernetes ecosystem despite being called "native" (tooling, security concerns). With respect to common lifecycle management operations: these features are not available (within Apache Flink) for any of the other resource providers (YARN, Standalone) either. From this perspective, I wouldn't consider this a shortcoming of the Kubernetes integration. Instead, we have been focusing our efforts in Apache Flink on the operations of a single Job, and left orchestration and lifecycle management that spans multiple Jobs to ecosystem projects. I still believe that we should keep this focus on low level composable building blocks (like Jobs and Snapshots) in Apache Flink to make it easy for everyone to build fitting higher level abstractions like a FlinkApplication Custom Resource on top of it. For example, we are currently contributing multiple improvements [1,2,3,4] to the REST API and Application Mode that in our experience will make it easier to manage Apache Flink with a Kubernetes operator. Given this background, I suspect a Kubernetes Operator in Apache Flink would not be a priority for us at Ververica - at least right now. Having said this, if others in the community have the capacity to push and *maintain* a somewhat minimal "reference" Kubernetes Operator for Apache Flink, I don't see any blockers. If or when this happens, I'd see some clear benefits of using a separate repository (easier independent versioning and releases, different build system & tooling (go, I assume)). Looking forward to your thoughts, Konstantin [1] https://issues.apache.org/jira/browse/FLINK-24275 [2] https://issues.apache.org/jira/browse/FLINK-24208 [3] https://cwiki.apache.org/confluence/display/FLINK/FLIP-194%3A+Introduce+the+JobResultStore [4] https://issues.apache.org/jira/browse/FLINK-24113 On Mon, Jan 10, 2022 at 2:11 PM Gyula Fóra wrote: > Hi All! > > This is a very interesting discussion. > > I think many users find it confusing what deployment mode to choose when > considering a new production application on Kubernetes. With all the > options of native, standalone and different operators this can get tricky :) > > I really like the idea that Thomas brought up to have at least a minimal > operator implementation in Flink itself to cover the most common production > job lifecycle management scenarios. I think the Flink community has a very > strong experience in this area to create a successful implementation that > would benefit most production users on Kubernetes. > > Cheers, > Gyula > > On Mon, Jan 10, 2022 at 4:29 AM Yang Wang wrote: > >> Thanks all for this fruitful discussion. >> >> I think Xintong has given a strong point why we introduced the native K8s >> integration, which is active resource management. >> I have a concrete example for this in the production. When a K8s node is >> down, the standalone K8s deployment will take longer >> recovery time based on the K8s eviction time(IIRC, default is 5 minutes). >> For the native K8s integration, Flink RM could be aware of the >> TM heartbeat lost and allocate a new one timely. >> >> Also when introducing the native K8s integration, another hit is that we >> should make the users are easy enough to migrate from YARN deployment. >> They already have a production-ready job life-cycle management system, >> which is using Flink CLI to submit the Flink jobs. >> So we provide a consistent command "bin/flink run-application -t >> kubernetes-application/yarn-application" to start a Flink application and >> "bin/flink cancel/stop ..." >> to terminate a Flink application. >> >> >> Compared with K8s operator, I know that this is not a K8s >> native mechanism. Hence, I also agree that we still need a powerful K8s >> operator which >> could work with both standalone and native K8s modes. The major >> difference between them is how to start the JM and TM pods. For standalone, >> they are managed by K8s job/deployment. For native, maybe we could simply >> create a submission carrying the "flink run-application" arguments >> which is derived from the Flink application CR. >> >> Make the Flink's active resource manager can talk to the K8s operator is >> an interesting option, which could support both standalone and native. >> Then Flink RM just needs to declare the resource requirement(e.g. 2 * >> <2G, 1CPU>, 2 * <4G, 1CPU>) and defer the resource allocation/de-allocation >> to the K8s operator. It feels like an intermediate form between native >> and standalone mode :) >> >> >> >> Best, >> Yang >> >> >> >> Xintong Song 于2022年1月7日周五 12:02写道： >> >>> Hi folks, >>> >>> Thanks for the discussion. I'd like to share my two cents on this topic. >>> >>> Firstly, I'd like to clarify my understanding of the concepts "native >>> k8s inte
Re: Flink native k8s integration vs. operator
Hi All! This is a very interesting discussion. I think many users find it confusing what deployment mode to choose when considering a new production application on Kubernetes. With all the options of native, standalone and different operators this can get tricky :) I really like the idea that Thomas brought up to have at least a minimal operator implementation in Flink itself to cover the most common production job lifecycle management scenarios. I think the Flink community has a very strong experience in this area to create a successful implementation that would benefit most production users on Kubernetes. Cheers, Gyula On Mon, Jan 10, 2022 at 4:29 AM Yang Wang wrote: > Thanks all for this fruitful discussion. > > I think Xintong has given a strong point why we introduced the native K8s > integration, which is active resource management. > I have a concrete example for this in the production. When a K8s node is > down, the standalone K8s deployment will take longer > recovery time based on the K8s eviction time(IIRC, default is 5 minutes). > For the native K8s integration, Flink RM could be aware of the > TM heartbeat lost and allocate a new one timely. > > Also when introducing the native K8s integration, another hit is that we > should make the users are easy enough to migrate from YARN deployment. > They already have a production-ready job life-cycle management system, > which is using Flink CLI to submit the Flink jobs. > So we provide a consistent command "bin/flink run-application -t > kubernetes-application/yarn-application" to start a Flink application and > "bin/flink cancel/stop ..." > to terminate a Flink application. > > > Compared with K8s operator, I know that this is not a K8s > native mechanism. Hence, I also agree that we still need a powerful K8s > operator which > could work with both standalone and native K8s modes. The major difference > between them is how to start the JM and TM pods. For standalone, > they are managed by K8s job/deployment. For native, maybe we could simply > create a submission carrying the "flink run-application" arguments > which is derived from the Flink application CR. > > Make the Flink's active resource manager can talk to the K8s operator is > an interesting option, which could support both standalone and native. > Then Flink RM just needs to declare the resource requirement(e.g. 2 * <2G, > 1CPU>, 2 * <4G, 1CPU>) and defer the resource allocation/de-allocation > to the K8s operator. It feels like an intermediate form between native and > standalone mode :) > > > > Best, > Yang > > > > Xintong Song 于2022年1月7日周五 12:02写道： > >> Hi folks, >> >> Thanks for the discussion. I'd like to share my two cents on this topic. >> >> Firstly, I'd like to clarify my understanding of the concepts "native k8s >> integration" and "active resource management". >> - Native k8s integration means Flink's master interacts with k8s' api >> server directly. It acts like embedding an operator inside Flink's master, >> which manages the resources (pod, deployment, configmap, etc.) and watches >> / reacts to related events. >> - Active resource management means Flink can actively start / terminate >> workers as needed. Its key characteristic is that the resource a Flink >> deployment uses is decided by the job's execution plan, unlike the opposite >> reactive mode (resource available to the deployment decides the execution >> plan) or the standalone mode (both execution plan and deployment resources >> are predefined). >> >> Currently, we have the yarn and native k8s deployments (and the recently >> removed mesos deployment) in active mode, due to their ability to request / >> release worker resources from the underlying cluster. And all the existing >> operators, AFAIK, work with a Flink standalone deployment, where Flink >> cannot request / release resources by itself. >> >> From this perspective, I think a large part of the native k8s integration >> advantages come from the active mode: being able to better understand the >> job's resource requirements and adjust the deployment resource accordingly. >> Both fine-grained resource management (customizing TM resources for >> different tasks / operators) and adaptive batch scheduler (rescale the >> deployment w.r.t. different stages) fall into this category. >> >> I'm wondering if we can have an operator that also works with the active >> mode. Instead of talking to the api server directly for adding / deleting >> resources, Flink's active resource manager can talk to the operator (via >> CR) about the resources the deployment needs, and let the operator to >> actually add / remove the resources. The operator should be able to work >> with (active) or without (standalone) the information of deployment's >> resource requirements. In this way, users are free to choose between active >> and reactive (e.g., HPA) rescaling, while always benefiting from the >> beyond-deployment lifecycle (upgrades, savepoint management, etc.) and >> alig
Re: Flink native k8s integration vs. operator
Thanks all for this fruitful discussion. I think Xintong has given a strong point why we introduced the native K8s integration, which is active resource management. I have a concrete example for this in the production. When a K8s node is down, the standalone K8s deployment will take longer recovery time based on the K8s eviction time(IIRC, default is 5 minutes). For the native K8s integration, Flink RM could be aware of the TM heartbeat lost and allocate a new one timely. Also when introducing the native K8s integration, another hit is that we should make the users are easy enough to migrate from YARN deployment. They already have a production-ready job life-cycle management system, which is using Flink CLI to submit the Flink jobs. So we provide a consistent command "bin/flink run-application -t kubernetes-application/yarn-application" to start a Flink application and "bin/flink cancel/stop ..." to terminate a Flink application. Compared with K8s operator, I know that this is not a K8s native mechanism. Hence, I also agree that we still need a powerful K8s operator which could work with both standalone and native K8s modes. The major difference between them is how to start the JM and TM pods. For standalone, they are managed by K8s job/deployment. For native, maybe we could simply create a submission carrying the "flink run-application" arguments which is derived from the Flink application CR. Make the Flink's active resource manager can talk to the K8s operator is an interesting option, which could support both standalone and native. Then Flink RM just needs to declare the resource requirement(e.g. 2 * <2G, 1CPU>, 2 * <4G, 1CPU>) and defer the resource allocation/de-allocation to the K8s operator. It feels like an intermediate form between native and standalone mode :) Best, Yang Xintong Song 于2022年1月7日周五 12:02写道： > Hi folks, > > Thanks for the discussion. I'd like to share my two cents on this topic. > > Firstly, I'd like to clarify my understanding of the concepts "native k8s > integration" and "active resource management". > - Native k8s integration means Flink's master interacts with k8s' api > server directly. It acts like embedding an operator inside Flink's master, > which manages the resources (pod, deployment, configmap, etc.) and watches > / reacts to related events. > - Active resource management means Flink can actively start / terminate > workers as needed. Its key characteristic is that the resource a Flink > deployment uses is decided by the job's execution plan, unlike the opposite > reactive mode (resource available to the deployment decides the execution > plan) or the standalone mode (both execution plan and deployment resources > are predefined). > > Currently, we have the yarn and native k8s deployments (and the recently > removed mesos deployment) in active mode, due to their ability to request / > release worker resources from the underlying cluster. And all the existing > operators, AFAIK, work with a Flink standalone deployment, where Flink > cannot request / release resources by itself. > > From this perspective, I think a large part of the native k8s integration > advantages come from the active mode: being able to better understand the > job's resource requirements and adjust the deployment resource accordingly. > Both fine-grained resource management (customizing TM resources for > different tasks / operators) and adaptive batch scheduler (rescale the > deployment w.r.t. different stages) fall into this category. > > I'm wondering if we can have an operator that also works with the active > mode. Instead of talking to the api server directly for adding / deleting > resources, Flink's active resource manager can talk to the operator (via > CR) about the resources the deployment needs, and let the operator to > actually add / remove the resources. The operator should be able to work > with (active) or without (standalone) the information of deployment's > resource requirements. In this way, users are free to choose between active > and reactive (e.g., HPA) rescaling, while always benefiting from the > beyond-deployment lifecycle (upgrades, savepoint management, etc.) and > alignment with the K8s ecosystem (Flink client free, operating via kubectl, > etc.). > > Thank you~ > > Xintong Song > > > > On Thu, Jan 6, 2022 at 1:06 PM Thomas Weise wrote: > >> Hi David, >> >> Thank you for the reply and context! >> >> As for workload types and where native integration might fit: I think >> that any k8s native solution that satisfies category 3) can also take >> care of 1) and 2) while the native integration by itself can't achieve >> that. Existence of [1] might serve as further indication. >> >> The k8s operator pattern would be an essential building block for a >> k8s native solution that is interoperable with k8s ecosystem tooling >> like kubectl, which is why [2] and subsequent derived art were >> created. Specifically the CRD allows us to directly express the >> concep
Re: Flink native k8s integration vs. operator
Hi folks, Thanks for the discussion. I'd like to share my two cents on this topic. Firstly, I'd like to clarify my understanding of the concepts "native k8s integration" and "active resource management". - Native k8s integration means Flink's master interacts with k8s' api server directly. It acts like embedding an operator inside Flink's master, which manages the resources (pod, deployment, configmap, etc.) and watches / reacts to related events. - Active resource management means Flink can actively start / terminate workers as needed. Its key characteristic is that the resource a Flink deployment uses is decided by the job's execution plan, unlike the opposite reactive mode (resource available to the deployment decides the execution plan) or the standalone mode (both execution plan and deployment resources are predefined). Currently, we have the yarn and native k8s deployments (and the recently removed mesos deployment) in active mode, due to their ability to request / release worker resources from the underlying cluster. And all the existing operators, AFAIK, work with a Flink standalone deployment, where Flink cannot request / release resources by itself. >From this perspective, I think a large part of the native k8s integration advantages come from the active mode: being able to better understand the job's resource requirements and adjust the deployment resource accordingly. Both fine-grained resource management (customizing TM resources for different tasks / operators) and adaptive batch scheduler (rescale the deployment w.r.t. different stages) fall into this category. I'm wondering if we can have an operator that also works with the active mode. Instead of talking to the api server directly for adding / deleting resources, Flink's active resource manager can talk to the operator (via CR) about the resources the deployment needs, and let the operator to actually add / remove the resources. The operator should be able to work with (active) or without (standalone) the information of deployment's resource requirements. In this way, users are free to choose between active and reactive (e.g., HPA) rescaling, while always benefiting from the beyond-deployment lifecycle (upgrades, savepoint management, etc.) and alignment with the K8s ecosystem (Flink client free, operating via kubectl, etc.). Thank you~ Xintong Song On Thu, Jan 6, 2022 at 1:06 PM Thomas Weise wrote: > Hi David, > > Thank you for the reply and context! > > As for workload types and where native integration might fit: I think > that any k8s native solution that satisfies category 3) can also take > care of 1) and 2) while the native integration by itself can't achieve > that. Existence of [1] might serve as further indication. > > The k8s operator pattern would be an essential building block for a > k8s native solution that is interoperable with k8s ecosystem tooling > like kubectl, which is why [2] and subsequent derived art were > created. Specifically the CRD allows us to directly express the > concept of a Flink application consisting of job manager and task > manager pods along with associated create/update/delete operations. > > Would it make sense to gauge interest to have such an operator as part > of Flink? It appears so from discussions like [3]. I think such > addition would significantly lower the barrier to adoption, since like > you mentioned one cannot really run mission critical streaming > workloads with just the Apache Flink release binaries alone. While it > is great to have multiple k8s operators to choose from that are > managed outside Flink, it is unfortunately also evident that today's > hot operator turns into tomorrow's tech debt. I think such fate would > be less likely within the project, when multiple parties can join > forces and benefit from each other's contributions. There were similar > considerations and discussions around Docker images in the past. > > Out of the features that you listed it is particularly the application > upgrade that needs to be solved through an external process like > operator. The good thing is that many folks have already thought hard > about this and in existing implementations we see different strategies > that have their merit and production mileage (certainly applies to > [2]). We could combine the best of these ideas into a unified > implementation as part of Flink itself as starting point. > > Cheers, > Thomas > > > [1] https://github.com/wangyang0918/flink-native-k8s-operator > [2] https://github.com/lyft/flinkk8soperator > [3] https://lists.apache.org/thread/fhcr5gj1txcr0fo4s821jkp6d4tk6080 > > > On Tue, Jan 4, 2022 at 4:04 AM David Morávek wrote: > > > > Hi Thomas, > > > > AFAIK there are no specific plans in this direction with the native > integration, but I'd like to share some thoughts on the topic > > > > In my understanding there are three major groups of workloads in Flink: > > > > 1) Batch workloads > > 2) Interactive workloads (Both Batch and Streamin
Re: Flink native k8s integration vs. operator
Hi David, Thank you for the reply and context! As for workload types and where native integration might fit: I think that any k8s native solution that satisfies category 3) can also take care of 1) and 2) while the native integration by itself can't achieve that. Existence of [1] might serve as further indication. The k8s operator pattern would be an essential building block for a k8s native solution that is interoperable with k8s ecosystem tooling like kubectl, which is why [2] and subsequent derived art were created. Specifically the CRD allows us to directly express the concept of a Flink application consisting of job manager and task manager pods along with associated create/update/delete operations. Would it make sense to gauge interest to have such an operator as part of Flink? It appears so from discussions like [3]. I think such addition would significantly lower the barrier to adoption, since like you mentioned one cannot really run mission critical streaming workloads with just the Apache Flink release binaries alone. While it is great to have multiple k8s operators to choose from that are managed outside Flink, it is unfortunately also evident that today's hot operator turns into tomorrow's tech debt. I think such fate would be less likely within the project, when multiple parties can join forces and benefit from each other's contributions. There were similar considerations and discussions around Docker images in the past. Out of the features that you listed it is particularly the application upgrade that needs to be solved through an external process like operator. The good thing is that many folks have already thought hard about this and in existing implementations we see different strategies that have their merit and production mileage (certainly applies to [2]). We could combine the best of these ideas into a unified implementation as part of Flink itself as starting point. Cheers, Thomas [1] https://github.com/wangyang0918/flink-native-k8s-operator [2] https://github.com/lyft/flinkk8soperator [3] https://lists.apache.org/thread/fhcr5gj1txcr0fo4s821jkp6d4tk6080 On Tue, Jan 4, 2022 at 4:04 AM David Morávek wrote: > > Hi Thomas, > > AFAIK there are no specific plans in this direction with the native > integration, but I'd like to share some thoughts on the topic > > In my understanding there are three major groups of workloads in Flink: > > 1) Batch workloads > 2) Interactive workloads (Both Batch and Streaming; eg. SQL Gateway / > Zeppelin Notebooks / VVP ...) > 3) "Mission Critical" Streaming workloads > > I think the native integration fits really well in the first two categories. > Let's talk about these first: > > 1) Batch workloads > > You don't really need to address the upgrade story here. The interesting > topic is how to "dynamically" adjust parallelism as the workload can change > between stages. This is where the Adaptive Batch Scheduler [1] comes into > play. To leverage the scheduler to the full extend, it needs to be deployed > with the remote shuffle service in place [2], so the Flink's Resource Manager > can free TaskManagers that are no longer needed. > > This can IMO work really well with the native integration as there is really > clear approach on how the Resource Manager should decide on what resources > should be allocated. > > 2) Interactive workloads > > Again, the upgrade story is not really interesting in this scenario. For > batch workloads, it's basically the same as the above. For streaming one this > gets tricky. The main initiative that we current have in terms of auto > scaling / re-scaling of the streaming workloads is the reactive mode > (adaptive scheduler) [3]. > > I can totally see how the reactive mode could be integrated in the native > integration, but with the application mode, which is not really suitable for > the interactive workloads. For integration with session cluster, we'd first > need to address the "scheduling" problem of how to distribute newly available > resources between multiple jobs. > > What's pretty neat here is that the interpreter (zeppelin, sql gw, ...) have > a really convenient way of spinning up a new cluster inside k8s. > > 3) "Mission Critical" Streaming workloads > > This one is IMO the primary reason why one would consider building a new > operator these days as this needs a careful lifecycle management of the > pipeline. I assume this is also the use case that you're investigating, am I > correct? > > I'd second the requirements that you've already stated: > a) Resource efficiency - being able to re-scale based on the workload, in > order to keep up with the input / not waste resources > b) Fast recovery > c) Application upgrades > > I personally don't think that the native integration is really suitable here. > The direction that we're headed is with the standalone deployment on > Kubernetes + the reactive mode (adaptive scheduler). > > In theory, if we want to build a really cloud (Kube
Re: Flink native k8s integration vs. operator
Hi Thomas, AFAIK there are no specific plans in this direction with the native integration, but I'd like to share some thoughts on the topic In my understanding there are three major groups of workloads in Flink: 1) Batch workloads 2) Interactive workloads (Both Batch and Streaming; eg. SQL Gateway / Zeppelin Notebooks / VVP ...) 3) "Mission Critical" Streaming workloads I think the native integration fits really well in the first two categories. Let's talk about these first: 1) Batch workloads You don't really need to address the upgrade story here. The interesting topic is how to "dynamically" adjust parallelism as the workload can change between stages. This is where the Adaptive Batch Scheduler [1] comes into play. To leverage the scheduler to the full extend, it needs to be deployed with the remote shuffle service in place [2], so the Flink's Resource Manager can free TaskManagers that are no longer needed. This can IMO work really well with the native integration as there is really clear approach on how the Resource Manager should decide on what resources should be allocated. 2) Interactive workloads Again, the upgrade story is not really interesting in this scenario. For batch workloads, it's basically the same as the above. For streaming one this gets tricky. The main initiative that we current have in terms of auto scaling / re-scaling of the streaming workloads is the reactive mode (adaptive scheduler) [3]. I can totally see how the reactive mode could be integrated in the native integration, but with the application mode, which is not really suitable for the interactive workloads. For integration with session cluster, we'd first need to address the "scheduling" problem of how to distribute newly available resources between multiple jobs. What's pretty neat here is that the interpreter (zeppelin, sql gw, ...) have a really convenient way of spinning up a new cluster inside k8s. 3) "Mission Critical" Streaming workloads This one is IMO the primary reason why one would consider building a new operator these days as this needs a careful lifecycle management of the pipeline. I assume this is also the use case that you're investigating, am I correct? I'd second the requirements that you've already stated: a) Resource efficiency - being able to re-scale based on the workload, in order to keep up with the input / not waste resources b) Fast recovery c) Application upgrades I personally don't think that the native integration is really suitable here. The direction that we're headed is with the standalone deployment on Kubernetes + the reactive mode (adaptive scheduler). In theory, if we want to build a really cloud (Kubernetes) native stream processor, deploying the pipeline should be as simple as deploying any other application. It should be also simple to integrate with CI & CD environment and the fast / frequent deploy philosophy. Let's see where we stand and where we can expand from there: a) Resource efficiency We already have the reactive mode in place. This allows you to add / remove task managers by adjusting the TM deployment (`kubectl scale ...`) and Flink will automatically react to the available resources. This is currently only supported with the Application Mode, that is limited to a single job (which should be enough for this kind of workload). The re-scaling logic is left completely up to the user and can be as simple as setting up a HPA (Horizontal Pod Autoscaler). I tend to think in the direction, that we might want to provide a custom k8s metrics server, that allows HPA to query the metrics from JM, to make this more flexible and easy to use. As this looks really great in theory, there are still some shortcomings that we're actively working on addressing. For this feature to be really widely adopted, we need to make the re-scaling experience as fast as possible, so we can re-scale often to react to the input rate. This could be currently a problem with large RocksDB states as this involves full re-balance of the state (splitting / merging RocksDB instances). The k8s operator approach has the same / even worse limitation as it involves taking a savepoint a re-building the state from it. b) Fast recovery This is IMO not as different from the native mode (although I'd have to check whether RM failover can reuse task managers). This involves frequent and fast checkpointing, local recovery (which is still not supported in reactive mode, but this will be hopefully addressed soon) and working directory efforts [4]. c) Application upgrades This is the topic I'm still struggling with a little. Historically this involves external lifecycle management (savepoint + submitting a new job). I think at the end of the day, with application mode on standalone k8s, it could be as simple as updating the docker image of the JM deployment. If I think about the simplest upgrade scenario, simple in-place restore from the latest checkpoint, it may be fairly simple to implement. What I
