Skip to main content

26 posts tagged with "2022"

View All Tags

· 4 min read
Doug Sillars

In early Feb 2022, we had our first meetup on Netflix Conductor: Using Conductor in Production. It was also co-hosted by the Netflix Conductor team.

We had two excellent talks from Maros Marasalek at FRINX and Nick Tomlin at Netflix.

After the two talks, we had 2 roadmap sessions. The first session was by the Netflix Conductor team - where they discussed recent releases, and walked through the Open Source roadmap for the coming months. The second session, by our own Viren Baraiya, introduced Orkes, and our plans for extending Netflix Conductor.

Conductor in FRINX

Maros' presentation showed how the FRINX team has integrated Conductor into their product, and how the FRINX tooling helps to build custom workflows.

Bridging human and system workflows with Conductor

Nick Tomlin works on the Finance team at Netflix, and his team has built a set of Conductor workflows that enable other Netflix teams to quickly build and share workflows.

Netflix Conductor roadmap

Our third talk was from the Netflix Conductor team - where they presented the roadmap for Netflix Conductor for the coming months.

Introducing Orkes

Finally, one of our founders (and the committer of the first line of Conductor code) Viren Baraiya presented Orkes and our roadmap:


Throughout the meetup, the attendees asked a number of great questions. They are reproduced here for visibility outside the meetup.

Hello, thanks for organizing such an event. I would like to know if there are any performance metrics for Conductor? We are planning to use it in a system with heavy traffic(multimillion requests each of them would trigger a workflow) and I would like to know if it’ll be reliable enough. Thank you :)

Conductor is horizontally scalable and we have known users scaling to handle workloads at the scale you mentioned. Here is a recent discussion on scale on Github, and a post from Netflix talking about the scale.

Conductor was built ground up for high reliability and performance. There are several companies that are running multi-million workflows on their core business flows.

I hear Netflix also is using which is cadence based. How do the two overlap at Netflix and the reason behind the use of Temporal?

Conductor is the default workflow orchestration tool of choice at Netflix. That said, engineers are free to choose the tool that’s best for their needs. The Spinnaker team at Netflix felt that Temporal is best aligned with their needs.

You mention CLI. Any thoughts about SDK as well?

Yes, we are working on that - stay tuned :) (Note: Check out the Orkes video above on things we are planning for next couple quarters)

Can AWS Lambda be hooked up as a task in workflow

AWS Lambda can be hooked up, but it's not available as a default task. There is an extension available that will allow integrating with AWS lambdas. We can send you details about this.

How often are conductor versions released to the community?

Frequent releases - upto twice per month. You can see the recent releases here:

From the previous presenter, what is the expected timeline for all those features to be released?

Most of the features we mentioned today will be released over the next 2-3 quarters in phases.

Any pointers about production level setup instructions of Dynomite (with the consideration of DR) on K8s cluster?

Please connect with us on our community slack channel or via Github discussions. Dynomite is not actively maintained anymore and is on its way towards being deprecated. There are alternatives that will offer similar features of Dynomite.

Thank you

Thank you to all the speakers for their incredible presentations, and also to our community for such great questions. This is our first of many meetups, and we hope to see all of you (and everyone else) at our next meetup.

In the meantime - don't forget to Star the Conductor repository.

For the latest updates on Conductor and Orkes, please subscribe to our YouTube channel, and follow us on Twitter.

· 5 min read
Doug Sillars

In our initial image processing workflow using Netflix Conductor, we initially built a workflow that takes one image, resizes it and uploads it to S3.

Image processing workflow diagram

In our 2nd post, we utilized a fork to create two images in parallel. When building this workflow, we reused all of the tasks from the first workflow, connecting them in a way that allows for the parallel processing of two images at once.

Two tasks are reused in both workflows: image_convert_resize and upload_toS3. This is one great advantage of using microservices - we create the service once and reuse it many times in different ways.

In this post, we'll take that abstraction a step further and replace the tasks in the two forks with a SUB_WORKFLOW. This allows us to simplify the full workflow by abstracting a frequently used set of tasks into a single task.

· 5 min read
Doug Sillars

In recent posts, we have built several image processing workflows with Conductor. In our first post, we created an image processing workflow for one image - where we provide an image along with the desired output dimensions and format. The workflow output is a link on Amazon S3 to the desired file.

In the 2nd example, we used the FORK System task to create multiple images in parallel. The number of images was hardcoded in the workflow - as FORK generates exactly as many paths as are coded into the workflow.

Several images are hardcoded in the workflow, but only 2 images are created. When it comes to image generation, there is often a need for more images (as new formats become popular) or sizes - as more screens are supported.

Luckily, Conductor supports this flexibility and has a feature to specify the number of tasks to be created at runtime. In this post, we'll demonstrate the use of dynamic forks, where the workflow splitting is done at runtime.

Learn how to create a dynamic fork workflow in this post!

· 8 min read
Doug Sillars

In our previous post on image processing workflows, we built a Netflix Conductor workflow that took an image input and then ran 2 tasks: The first task resizes and reformats the image, and the second task uploads the image to an AWS S3 bucket.

With today's varied screen sizes and varied browser support, it is a common requirement that the image processing pipeline must create multiple images with different sizes and formats of each image.

To do this with a Conductor workflow, we'll utilize the FORK operation to create parallel processes to generate multiple versions of the same image. The FORK task creates multiple parallel processes, so that each image will be created asynchronously - ensuring a fast and efficient process.

In this post, our workflow will create 2 versions of the same image - a jpg and a webp.

· 2 min read

What is Conductor

Conductor is a Microservices orchestration platform from Netflix, released under Apache 2.0 Open Source License.

Design for failures

Failures and service degradation are the facts of any system; this is especially true with large interconnected systems running in the cloud. Conductor is designed with principles that systems can and will go down, degrade in performance and any dependencies should be able to handle such failures.

Tasks in Conductor

Conductor workflows are the orchestration of many activities known as tasks. Each task represents ideally a stateless worker who does the work and produces output given a specific input. The tasks are typically running outside the Conductor server and there are many factors that could affect their availability.

Designing for failures

Conductor allows you to define your stateful applications that can handle failures and temporary degradation of services and without having to write code.

Configuring tasks to handle failures

Each task in Conductor can be configured in a way how it responds to availability events such as:

  1. Failures
  2. Timeouts
  3. Rate limits.

Here is a sample task definition:

"createdBy": "user",
"name": "sample_task_name_1",
"description": "This is a sample task for demo",
"responseTimeoutSeconds": 10,
"timeoutSeconds": 30,
"timeoutPolicy": "TIME_OUT_WF",
"retryCount": 3,
"retryLogic": "FIXED",
"retryDelaySeconds": 5,
"rateLimitPerFrequency": 0,
"rateLimitFrequencyInSeconds": 1

retry* parameters specify how to handle cases where the task execution fails and retries can be configured to be with fixed delay or exponential backoff. Similarly timeout* parameters specify how much time to give for a task to complete execution and if the task should be marked as 'Timed Out' if it runs longer than that.

More Details

Follow us at for the source code and updates.

· 9 min read
Doug Sillars

There are many tools available to work with images - resizing, changing the format, cropping, changing colors, etc. Tools like Photoshop require a lot of manual work to create an image. Online tools for image processing are also extremely popular. But, rather than doing the work manually, or paying for a service to modify your images, wouldn't it be cool to have a workflow that does image resizing for you automatically? In this post, we'll build just this using Conductor to orchestrate the microservices involved, and to create an API-like surface for image processing.

In this post, we'll run Conductor locally on your computer. The Conductor workflow consists of two tasks. The first task reads an image and resizes it according to the parameters provided (labeled "image_convert_resize_ref" in the image below). The second task ("image_toS3_ref" below) takes the resized image and saves it to an Amazon S3 bucket.

Image processing workflow diagram in Conductor

Using a microservice architecture for this process allows for easy swapping of components, and allows for easy extension of the workflow - easily adding additional image processing steps (or even swapping in and out different processes for different workflows). We could also easily change the location of the saved file based on different parameters.