Cloud Native Computing Foundation KubeCon + CloudNativeCon North America 2017 (Austin), 15 Dec 2017

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Distributed Workflows for Microservices-Style Applications [I] - Yun Qin, Nirmata

Description

Distributed Workflows for Microservices-Style Applications [I] - Yun Qin, Nirmata

Microservices-style architectures solve several problems but also introduce new complexities. With Microservices, a best practice is to keep services isolated and loosely coupled. However, in the real world, it is not uncommon to encounter business logic which requires coordination across multiple business functions i,e. microservices.

The distributed workflow pattern addresses this problem. In this presentation we will describe the distributed workflow pattern and its use cases. We will then look at various implementations of this pattern, such as Netflix Conductor, AWS Simple Workflow Service and NirmataOSS Workflow.

We will end by showing a demonstration of a distributed workflow, running on a Kubernetes cluster and show how workflow managers can leverage Kubernetes features like Horizontal Pod Autoscaling.

About YUN QIN
Yun is a software enginner in Nirmata, a company deliverring integrated solutions for multi-cloud application management. Yun has extensive experience in distributed system application development and operations. Prior to joining Nirmata, Yun worked as a senior network engineer at China Unicom, one of the biggest telecommunication company in China. Yun holds a M.S. in Computer Science Engineering from Santa Clara University.
Join us for KubeCon + CloudNativeCon in Barcelona May 20 - 23, Shanghai June 24 - 26, and San Diego November 18 - 21! Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy and all of the other CNCF-hosted projects.

A

Thank you, everyone for coming to my presentation. My topic today is distributed. Workflow for micro service, tile application, so micro service architecture is adopted for breaking down the big software application into loose coupling service modules. It provides a lot of good things. Oh sorry, sorry about that. So the micro service architecture provides a lot of good things, such as failure, isolation, scalability and it's easy for the for the development, but it also introduced some new complexity.

A

Suppose you're gonna implement your business logic, a micro service design, so how you're gonna? How you gonna manage your message: traveling across the different modules and how you're gonna coordinate your task execution on different services and how you're gonna handle your errors so to make your life easier, I. Think a workflow pattern will give you the capability to implement your business service logic over loose coupling service modules.

A

So before we start I, gotta say something about myself. My name is Ming Cheng and I. Former I was a formers network engineer and now I work. As a software engineer in the matter, a company delivered the integrated solution on Motty cloud application management as well as kuben Eddy based container management. Also Yamada is a kubernetes certificate service provider.

A

Okay, so my presentation is to introduce the distributed workflow pattern and its usage in micro service style application using the matter OSS workflow library, so here's the outline so first I will briefly introduce what is microservice and what is or workflow. And what is the challenge about the workflow management of micro services? Then I will introduce you, the Nomura, OSS workflow and try to do the demo, and we were also we were talking about other solutions.

A

So what is micro service architecture? Basically in micro service architectures service model walked independent and isolated. In most case, we deploy those service module on different containers in docker word and the different service module will communicate each other through, let's say, REST API, then what is workflow so what flow basically is a cop is the coordinator of a sequence of tasks for it.

A

Let's, for taking an example, if you want to do a shopping workflow, so you first, you gonna place placed order and then you're gonna pay the order and then you're gonna get your product shipped. So there are three tasks in this workflow and each tasks are processed by different service module and they are sequential. So we need a workflow to coordinate those tasks execution, and this is a one LS tic web application. So what about you? You want to scale your application in a distributed environment.

A

That means you have multiples everything's. You have multiple instance for your web application. Accordingly, you have multiple instance for your workflow. Then you will need a scheduling module to coordinate those workflow instance.

A

Now we come to the micro services. As we mentioned before, micro services module are independent and isolated, so the challenges are how we're going to hold an executive, a synchronized over the distributed in across multiple services.

A

At the same times, we need to manage the tasks of dependency and also we need to handle the complex logic such as what if one of the tasks is felt, and you want to keep going your your workflow or if the, if the one of the tasks fail, do you want to stop all the workflow kind of things and namana OSS workflow solve those problems?

A

It is an open source library written in Java based on Apache zookeeper, which is a well know, reliable and highly available coordinator, and- and we also use the curator as the kind to keeper client it is lightweight and easy to use. So here needs to some a features of the workflow library. It is capable of manage the tasks relationship, so the tasks relationship here mainly means the tasks of dependency and it could schedule tasks in the distributed environment and the tasks type could be customized.

A

Also. It gives a good support for fault, tolerant, for example. Maybe you maybe anything will go wrong during your task execution say your service is restarting or your zookeeper cluster change state. Then the task execution will be resumed.

A

There's some key component in the workflow library first is the workflow manager. The workflow manager will reduced for a specific task type when the service initialized, so the workflow manager is responsible to submit your tasks and the give notification say. There's a note. There's a new task note there and the scheduler will manage the task note during the task execution.

A

It will ensure that most the task node will be, will not be executive trice, considering because it's a distributed system right so and for the task queue here is a wrapper of zookeeper note where we put our tasks note in and the task executor here is it's just a executives read: you can have multiple tasks threads in a in your service. Sorry, multiple tasks, executives read in your service: let's go deeper to take a look at the task, execution module.

A

So when the workflow managers submits a task, it will create a new task note in the scale in the task queue. This is task type specific and the executor as a consumer will listen to those task queue. If there's a new node come in, they will be notified and then one of the task executor say we have multiple tasks. Executor instance right, so one of the executives read will pick up the tasks and executor it.

A

So the scheduler here will ensure that no other tasks executor pick up the same task here and also the scheduler here will make sure that if the task has a sub task, it will extracted sub tasks and then executive that so the whole executive model is decentralized. The workflow manager instance and the task executor could be on the same service, but it's not not necessary and it works well in the distributed system and the task could be a singer, could be executive, a synchronized.

A

So here the diagram many show how we manage the task, a dependency, the task, the task of dependency, is represented by a directed acyclic graph. So the those tasks has no dependency like task 1 task. Two and task 3 can be executives concurrently, but for task 4 and task 5. Take example here should be wait to the to destroy to the 1 2 3 finished executive that it will get started so this. So we can. We can handle the concurrent in a sequential tactics. You executor here.

A

Here I give some example code for how to beaut workflow you might need a workflow manager builder tribute the workflow and pass by some related parameters like in which namespace this namespace is its many related to the zookeeper namespace. So the namespace to where you want to put it your task in, and so the auto cleaner here, is to say that when the tasks be complete, when task exaggeration be completed, you wanna, you want to automatically clean up those tasks, notes from your task queue.

A

So how many times you want to wait after the completion and how often you want to run your auto cleaner. You could define here. Also here we can add in task executor. The task executor means means how many concurrent tasks executives. Rather, you want for this task. Type yeah so, and also here, is an example code for how to write the executor.

A

Take it simple here is I, just I just get a logger here to print to that the task could be executive. So this is a one executor. When you adding the way, adding multiple executor, it will exact you to task. At the same time, okay now I'm going I'm gonna some demo to see how how our workflow library works in our production.

A

So this is a new matter per web applications. I'm gonna create a new environment here.

A

Use a stop me or print.

A

Okay, there's a lot of if there's a lot of service here in the in this application. So each service is a you can you can think it's the micro service so why I want to demonstrate here is that I want to show you how we manage the tasks, a dependency in our workflow?

A

If you're looking to the briefing blueprint to see the shop me and go to see the service here, you can see the dependency so for cats, lock, it has no dependents and the customer also, but for deals it dependents on catalog, service and customer service, and the loyalty here is also is the similar things. So what I'm gonna do here is try to update the image of all those tasks, all those service- ok, let's say, there's a there's a service file here, let's just remove that.

A

Yeah, this is how my mother managed the micro service, so I'm gonna do a the environment update to all the service here, I choose an image.

A

Sorry I choose a new image for all those service. Let's say what happened.

A

So, underneath for each service, the tag at the image update is it is a independent task, but the whole workflow, including all the tasks here and will take care of the dependency of the task for what what we have seen before the catalog and the customers has no dependency, so those tasks will be executive first and for the task, like deals or I, think she's loyalty orders are dependent on those on the tasks on the previous tasks, so it will be executive later after those tasks. After the previous tasks be completed,.

A

Yeah you can see that, for those tasks has no dependency, it is executive concurrently and for those tasks has dependency, we'll wait for the previous tasks finished executive. Then it starts executive.

A

Okay, so yeah this is. This is what we call a workflow. So basically, it consists of multiple tasks and tasks over different microservices and we just managed to do it here, yeah, okay, so this is finished.

A

Let's go back to the slides.

A

What I just show is how we is how we use our workflow. You use the dimana workflow library to manage the workflow in our productions. Of course, there's other great solutions in the industry, such as Netflix conductor, which is the orchestrator, but also provides a feature that you can define your task execution using a JSON DSL.

A

Others like integrating call workflows, it's a cloud application, so you can use that in your in your cloud. Applications across multiple machines. These both are great solutions and great and integrations solutions. It's just. We start with our zookeeper and the curator framework, and we just want to write a library which is light, weighted and easy to use easy to plug in so we develop our own yeah.

A

So that's pretty much about it. So thank you for coming so I just wanted to see if you guys have any questions about that.

A

A

Well, from our experience, a single task workflow always and for what? Why show before it's a like wrap the tasks, wrap tactics to task a chain like you can do dozen of tasks in your workflow, but yes, I, haven't tested I mean in our current production. We haven't faced the issues that we have hundreds of tasks there, so yeah, maybe I will we were trial at technically I think it supports any other questions.

A

A

Our production, you mean production or workflow library, there's two different thing, because our workflow library is the open source library. There's a github link on the slides, but our production is yeah. It's a protection bill in this production, yeah.

A

I'm not so sure about that, but I think if you're using, if you already use, because we use a zookeeper and curator sometimes it's to be honest, it's a it's kind of difficult to monitor and troubleshooting them. But so, if you already have your zookeeper cluster runs, maybe you can have a try.

A

Okay, so, by the way our company have boosts downstairs, so if you guys have any interests- or you want to know more about a workflow library or our production, you're very welcome to come to our booth. Okay, thank you.