Scalable Compute
Compute big tasks, small tasks, and parallel tasks in the cloud without hassle.
Compute anything, change nothing
ML, AI, and data apps have varied compute needs: Some require small-scale compute, some many CPU cores, some need GPUs, and some plenty of memory.
With one line of code, you can annotate your functions with their resource requirements. Existing code and libraries work without changes. No need to learn new paradigms or frameworks.
Scalable pool of compute resources
The Outerbounds Platform comes with an auto-scaling, fully managed compute cluster, deployed on your AWS account, optimized for Metaflow workloads.
Use the cluster flexibly for all compute needs from 10,000-way hyperparameter sweeps and data processing with TBs of RAM to multi-GPU model training.
Minimize cloud costs
Workloads are monitored to improve utilization, optimize instance pools, attribute costs and help minimize your cloud invoices.
Utilize cloud savings plans, spot instances, and right-sized instances for optimal cost.
Trusted by the best
we’ve been able to entirely switch over our research experimentation process to rely on Metaflow with great success.
Bumped into Metaflow at work, plus @HamelHusain seems excited about it, so I am excited about it 🙂 But after watching the video, wow 😲
Thanks for Metaflow: it's a pleasure to use!
wow I didn't realize how much of the infrastructure needs Metaflow takes care of!
Love metaflow!
Metaflow - what a great tool!
I really appreciate the clean syntax and the seamless transitioning from dev to dev-at-scale to prod-at-scale
One of the things I liked about Metaflow from Netflix is their clear documentation on "Dealing with Failures": it's important to raise awareness about "retry" logic, transient/permanent failures, idempotent operations, etc.
By unifying our data and ML pipelines we can now automatically train models when an underlying warehouse table is updated which drastically improves efficiency and collaboration.
There's a lot of good ideas made practical in Metaflow to help ML projects and developers focus on what's important.
@MetaflowOSS rules!
Very impressive work! Easily mix multiple Python virtual environments and compute resources within a single pipeline 🤯 using #Metaflow
really love the event triggering feature in Metaflow
The MLOps piece of tech I'm most excited about is @MetaflowOSS. Just from reading the docs I'm amazed at how a large, enterprise company like Netflix seems to have really, really grasped the developer experience in the way I'd only expect 'smaller' projects to do.
i love metaflow.
I've been using Metaflow for years now. I can't endorse it nor the people behind it enough.💥Parallelize hyperparameters across containers trivially 💥Persist, retrieve, reproduce results and data artifacts from forever ago 💥Stand up an API that serves resulting data
"As an informal estimate, our #datascience team believes they were able to test twice as many models in Q1 2021 as they did in all of 2020, with simple experiments that would have taken a week now taking half a day."
Our complex, multi-stage workflows are codified and orchestrated using Metaflow.
Very impressive level of polish in the Metaflow docs/tutorials. Pretty incredible how many details have been abstracted away by Metaflow.
Metaflow was recommended at work. I spent some time with it today... it's amazing. Their tagline is on-point: "Build and manage real-life data science projects with ease."
A space I'm excited about: 'open source on top of public cloud tech.' Instead of using open source libraries (which you have to run & scale, e.g. RabbitMQ), it uses infinitely-scalable off-the-shelf cloud services (e.g. SQS). Metaflow is a good example. https://docs.metaflow.org/metaflow-on-aws/metaflow-on-aws
Maybe we'll be talking about this as the "Airflow for Data Science" in a few years. Not replacing Airflow, but in the sense that it's the default standard for orchestrating data science projects like Airflow does for data engineering pipelines.
We've had a really great experience transitioning from Argo workflows to Metaflow
I find it *extremely* satisfying that the tutorials for Metaflow (which was originally developed at Netflix) are organised in seasons and episodes 🤓
Outerbounds Platform has proven to be transformative for us. The platform has allowed us to execute and train models without the burden of worrying about infrastructure
The end-to-end ML management in Outerbounds—event triggers, experiment tracking, managed artifacts—was a game-changer for our small team.
Been using metaflow for a while now, and love it. At its best, it hugely speeds up our development cycles
Have been using Airflow in prod for several years now & looking forward to making a switch to Metaflow
Metaflow has made my daily life as an ML engineer so much easier, especially as we scale our efforts. I've never worked with a framework that works out of the box as well as metaflow.
MetaFlow has a really fantastic API. https://docs.metaflow.org/metaflow/basics. Most DAG systems could learn from it.
Metaflow has really been helping in speeding up our ds workflows
This diagram is pure perfection h/t @moritzplassnig How much data scientist cares vs how much infrastructure is needed 🤣
I haven’t yet heard a bad thing about metaflow.
Metaflow records the result of every step to S3. IMO the ability of metaflow to expose individual components of a complicated process is a big deal. Like, a really big deal.
In six weeks, a team that hadn’t used Metaflow before was able to build an ML-based model, A/B test its performance, which handily beat the old simple approach, and roll it out to production.
We are saving over 90% of the @awscloud cost in our #artificialintelligence model training phase. We've just finished the automation of our AI pipelines through the implementation of @MetaflowOSS , and the migration from http://H2O.ai
Outerbounds was like Metaflow, but with a ‘plus plus’—it made everything easier.
Metaflow has helped us avoid the anti-pattern of needing to push code to find out if something works
What happened to Kubeflow? So many companies are migrating away from it because it's a pain to configure and maintain. I find it hard to recommend Kubeflow anymore. Right now, Metaflow seems to be a much nicer option.
Thanks for a fantastic product ❤️
The team has shaved months off the time it takes to build a productionized machine learning model, and at the same time improved the overall collaboration of the data science organization with other departments
I've been extremely impressed by the Metaflow team. They have built something that's loved by everyone, so much so that they have Metaflow sad hours to encourage people say what can be improved. In my conversations with our users, I've often heard that Metaflow is one of the best software Netflix has ever built.
Built with the core idea of allowing ML engineers and data scientists to use Python as native way to work, this friendly approach has tremendously boosted Carsales productivity in ML
Metaflow is such an amazing tool!
❤️ metaflow! The people-first API design is fantastic.