+Yi Hi Tom,
Thank you for your feedback on Samza's architecture. Pluggability has been a differentiator that has enabled us to support a wide range of use-cases - from stand-alone deployments to managed services, from streaming to batch inputs and integrations with various systems from Kafka, Kinesis, Azure to ElasticSearch. Thanks for your ideas on integrating Samza and Kubernetes. Let me formalize your intuition a bit more. The following four aspects are key to running Samza with any environment. 1. Liveness detection/monitoring: This provides a means for discovering the currently available processors in the group and discovering when a processor is no longer running. The different JC implementations we have rely on Zk, Yarn or AzureBlobStore for liveness detection. 2. Partition-assignment/coordination: Once there is agreement on the available processors, this is just a matter of computing assignments. Usually, (1) and (2) will require you to identify each processor and to agree on the available processors in the group. For example, when the ClusterBasedJC starts a container, it is assigned a durable ID. 3. Resource management: This focusses on whether you want your containers to be managed / started by Samza itself or have something external to Samza that starts it. While the former allows you to run a managed service, the latter allows for more flexibility in your deployment environments. We use both models heavily at LinkedIn. As an example, the ClusterBasedJC requests resources from YARN and starts the containers itself. The ZkBasedJC assumes a more general deployment model and allows containers to be started externally and relies on Samza only for (1) and (2). 4. Auto-scaling: Here again, you can build auto-scaling right into Samza if there's support for resource management or do it externally. Having said this, you can implement this integration with Kubernetes at multiple-levels depending on how we choose to tackle the above aspects. ">> My intuition is that I need a JobCoordinator/Factory in the form of a server that sets up Watches on the appropriate Kubernetes resources so that when I perform the action in [4.1] *something* happens. " This alternative does seem more complex. Hence, I would not go down this path as the first-step. For a start, I would lean on the side of simplicity and recommend the following solution: - Configure your Samza job to leverage the existing ZkBasedJC. - Start multiple instances of your job by running the *run-app.sh* script. I believe Kubernetes has good support for this as well. - Configure Kubernetes to auto-scale your instances on-demand depending on load. - As new instances join and leave, Samza will automatically re-distribute partitions among them. Additionally, we would be thrilled if you could contribute your learnings back to the community - in the form of a blog-post / documentation to Samza itself on running with Kubernetes. Please let us know should you need any further help. Here's an example to get you started: https://github.com/apache/samza-hello-samza/tree/master/src/main/java/samza/examples/wikipedia/application Best, Jagdish On Sat, Jan 27, 2018 at 8:54 AM, Tom Davis <t...@recursivedream.com> wrote: > Hi there! First off, thanks for the continued work on Samza -- I looked > into many DC/stream processors and Samza was a real standout with its > smart architecture and pluggable design. > > I'm working on a custom StreamJob/Factory for running Samza jobs on > Kubernetes. Those two classes are pretty straight forward: I create a > Deployment in Kubernetes with the appropriate number of Pods (a number > <= the number of Kafka partitions I created the input topic with). Now > I'm moving onto what executes in the actual Docker containers and I'm a > bit confused. > > My plan was to mirror as much as possible what the YarnJob does > which is setup an environment that will work with `run-jc.sh`. However, > I don't need ClusterBasedJobCoordinator because Kubernetes is not an > offer-based resource negotiator; if the JobCoordinator is running it > means, by definition, it received the appropriate resources. So a > PassThroughJobCoordinator with appropriate main() method seemed like the > ticket. Unfortunately, the PTJC doesn't actually seem to *do* anything > -- unlike the CBJC which has a run-loop and presumably schedules > containers and the like. > > I saw the preview documentation on flexible deployment, but it didn't > totally click for me. Perhaps because it was also my first introduction > to the high-level API? (I've just been writing StreamTask impls) > > Here's a brief description of the workflow I'm envisioning, perhaps > someone could tell me the classes I should implement and what sorts of > containers I might need running in the environment to coordinate > everything? > > 1. I create a topic in Kafka with N partitions > 2. I start a job configured to run N-X containers > 2.1. If my topic has 4 partitions and I have low load, I might want > X to start at 3 so I only have 1 task instance > 3. Samza is configured to send all partitions to task instance 1 > 4. Later, load increases. > 4.1. I use Kubernetes to scale the job to 4 pods/containers > 4.2. Samza re-configures such that the new containers receive work > > My intuition is that I need a JobCoordinator/Factory in the form of a > server that sets up Watches on the appropriate Kubernetes resources so > that when I perform the action in [4.1] *something* happens. Or perhaps > I should use ZkJobCoordinator? Presumably as pods/containers come and go > they will cause changes in ZK that will trigger Task restarts or > whatever logic the coordinator employs? > > Okay, I'll stop rambling now. Thanks in advance for any tips! > > - Tom > -- Jagadish V, Graduate Student, Department of Computer Science, Stanford University
