The Ship List: MissingLink.ai February Edition

Published in

MissingLink Deep Learning Platform

4 min readFeb 28, 2019

Since our public launch in September, the MissingLink team has been working hard to improve the product and add new features users have requested.

You told us deep learning teams require tools that are reliable, secure, scalable, collaborative, modular, and most of all easy-to-use. Based on that feedback, we have added a number of features designed to streamline the entire deep learning pipeline.

Here’s a rundown of everything the MissingLink team has improved or shipped since our public launch last fall.

Microsoft Azure + MissingLink: Deep Learning for Computer Vision At Scale

If you’re serious about deep learning, you’ve probably run experiments in the cloud or at least thought about it. Microsoft Azure is an excellent option for scaling your deep learning experiments, but it can be complex to manage if you’re working as a team.

MissingLink makes running computer vision projects with massive datasets on Azure easy to set up and easy to operate. With MissingLink, you can provision multiple GPUs to run heavy workloads with ease, which allows you to forget about infrastructure and frees you up to build winning experiments.

If you haven’t done so already, sign up for MissingLink and integrate the SDK with Azure to scale deep learning like a pro.

Experiment Management Redesign

We kicked off 2019 by giving experiment management a facelift. The new console brings all the most important information into view, including hyperparameters and artifacts, and is designed to make everything just a bit more accessible. Log in to your console and try it out now.

Version Control for Data at Petabyte Scale

While it was already in use near the end of 2018, data management officially launched in early 2018. Thanks to a competitive cloud landscape, MissingLink users can already store petabytes of data. However, being able to store and reproduce human-readable queries efficiently is a game changer for data scientists.

If you’ve ever mixed data streams that you shouldn’t have or couldn’t figure out which data set caused a failure in production, MissingLink Data Management was made for you. We’ve launched petabyte-scale data management support for AWS, GCP, and Azure storage solutions already, and will have more configurations to come in 2019.

A One-click, Scalable Solution for Hybrid Environments

We built MissingLink with resource management in mind, but interviews with AI teams revealed just how complicated, buggy and manual managing hybrid environments could be. Deep learning experts reluctantly learned how to spin machines up and down, while also ensuring that they were able to extract all the code, data, logs, and other artifacts in and out of remote machines.

We determined that we should abstract resource management into a one-click solution that would be just as powerful as managing your own machines on-premises or in the cloud. By the end of 2018, we launched support for resource management on-premise, AWS, and Microsoft Azure.

Get Started in Minutes with Streamlined Onboarding Process

We recently released a more streamlined wizard to get you up and running quickly and painlessly. The wizard walks you through a few simple steps:

Create a Resource Group.
Select if you want to use a public cloud (AWS, Azure or GCP) or a local environment.
Configure some additional parameters such as your Docker image and Git repository

After that, you’re done! The wizard will automatically generate the command line to run your experiment locally or in the cloud.

Webhooks: Powerful Automation for Your Deep Learning Experiments

Pushing your model to production is one of the most critical yet challenging tasks in deep learning. In December 2018, we released webhooks to make it easy for deep learning teams to automate their pipeline. Once you configure a webhook URL for your organization, each time a MissingLink event (experiment start, end or fail) occurs, a POST request is sent to this URL with information about the event.

This unlocks a world of automation possibilities, including triggering a deployment to production, recording a bug when a job crashes and interaction with other systems when jobs are queued or experiments end.

Read more and see a detailed tutorial about Webhooks here.

Increase Team Collaboration and Visibility with a Visual Job Dashboard

Until recently, we didn’t provide a great deal of detail on our job dashboard. To improve this experience, we created a much more detailed page for each job, including information such as the name, status, duration, and owner of each. You can also access detailed logs for each job, the Github link to the code and query details. In the page, you can also change the job priority while it’s running.