Skip to main content

AI Deploy - Troubleshooting

Stephanie R.
Updated

This page gives you a few hints on how to debug your apps if you encounter some issues.

 

Requirements

 

Instructions

How to start an AI Deploy app?

All steps for starting and working on AI Deploy are described here.

 

How long can I use my AI Deploy app?

There is no duration limitation on AI Deploy apps. An AI Deploy app runs continuously until manually interrupted by the user.

 

Building your app

Best practices and mandatory guidelines to build your app

When you are deploying your own applications and models, some guidelines must be followed. We detail them in the guide AI Deploy - Build and use a custom Docker image. Be particularly cautious about image requirements such as OVHcloud user and Docker architecture used. Otherwise, your deployment will end in FAILED status.

 

Apps examples to follow

If you need some official examples, please follow this guide where we share the source code: AI Deploy - Apps portfolio.

 

Test your app locally and in the cloud

Before paying for cloud resources, feel free to locally test your Docker image. For that, simply install Docker on your local environment.

For the building step, as explained in the mandatory guidelines linked in the previous section, your Docker image has to support at least linux/amd64 platform to be deployed correctly. Otherwise, deployment will fail.

Then perform a docker run as follows:

# Build your Docker image for at least linux/amd64 architecture
docker buildx build --platform linux/amd64,linux/arm64 ...

# Run your Docker image as OVHcloud user
docker run --rm -it --user=42420:42420 <image-identifier>

This way, you will imitate the OVHcloud user. Once validated locally, you can deploy your app first with CPUs which are cheaper compared to GPUs.

 

Deployments

My deployment has failed

An AI Deploy app has a workflow in multiple steps, the FAILED status being one of them. This state happens when OVHcloud is unable to deploy your app, meaning the infrastructure side (backend) is working fine but something is broken on the image side. You can find more details about AI Deploy workflow on the AI Deploy - Billing and lifecycle page.

Main items to troubleshoot:

  • Typography in your repository name, image, or version name. Test deploying your image locally first.
  • Your Docker image is not following mandatory guidelines, such as OVHcloud user. See AI Deploy - Build and use a custom Docker image.
  • Your Docker image is in a private registry and you did not authorize OVHcloud to access it.
  • You have reached your quotas in terms of CPUs or GPUs. You can check them via the OVHcloud Control Panel (Project Management / Quotas) or the ovhai CLI command ovhai me.

If you are using ovhai CLI, you can get more details about your command with the ovhai debug command, and ovhai app logs <app_ID> to download logs history.

 

My deployment is in error

While a deployment in the FAILED state is due to a problem with the image, repository, etc., an app in the ERROR state can occur when AI Deploy is encountering an issue.

Try redeploying your app, and modifying the targeted data center for example. As in the previous answer, when using our CLI you can get more details about your command with the ovhai debug command, and ovhai app logs <app_ID> to download logs history.

If the issue persists, please contact our support teams.

 

My Deployment seems very long

When AI Deploy initializes your app, the Docker image is pulled (downloaded) in our infrastructure and replicated over the replicas, if any. The larger the Docker image is, the longer it will take to be deployed on the AI Deploy side.

Also, since we pull the data from a registry of your choice, if this particular registry is experiencing some issues or is restricted in terms of bandwidth or throughput, it may cause some slowness.

In an ideal situation, for a Docker image of approximately 1GB, without external data linked, it should take less than 10 minutes.

 

Scaling strategies

When using AI Deploy, you can choose between two scaling strategies: static scaling and autoscaling, allowing you to scale up or down based on triggers such as CPU or RAM usage.

  • The static scaling strategy allows you to choose the number of replicas (1 to 10) on which the app will be deployed.

  • With the autoscaling strategy, it is possible to choose both the minimum number of replicas (1 by default) and the maximum number of replicas.

Find more information on the official documentation about scaling strategies.

 

My deployed app does not scale

If your app does not scale:

  • Check if you deployed your app with static scaling or autoscaling.
  • Verify triggers (CPU or RAM usage) and their value. By default, the value is at 75%.
  • Open the Monitoring dashboard of your app (Grafana dashboard is provided for each app) and check if the threshold has been reached.
  • Refer to the following load-testing tutorial which also provides a dashboard example to follow your scaling: AI Deploy - How to load test your application with Locust.

 

My deployed app is very slow

Slowness may find its roots in multiple reasons. Each deployed app is a combination of software code and resources, such as compute and network.

If you are experiencing slowness, here are some actions to investigate:

  • Open the Monitoring dashboard for your app (Grafana dashboard is provided for each app) and check if some resources are reaching 90/100%, such as RAM, CPU, GPU, or network. You can also check the overall latency.
  • If nothing is visible, it can be an issue between the client (where the query comes from) and the deployed app. As an example, if you are contacting your apps from a geographically distant point, it will add latency. Try reducing the distances in your architecture.
  • Your Docker image itself may be the root cause. Try running your Docker image locally, and query your app locally. Some apps might be heavy to run or not well-optimized.

 

My deployment has crashed

Like any cloud product, AI Deploy might experience hardware or software failures over time. To mitigate the risk on your side, please deploy your app on at least two replicas, allowing us to provide high availability. At this time, all replicas are in the same region, but it will prevent them from a physical server failure.

Another root cause may be your own Docker image, for example, by writing an uncontrolled amount of data into your working directory.

We also recommend orchestrating your workflow with third-party tools such as Airflow, Prefect, Dagster, or Kestra, allowing you to relaunch an app once it has crashed.

If your app crashes and you are using ovhai CLI, you can get more information with ovhai app logs <app_ID> to download the logs history.

 

My data is not synchronized back

AI Deploy does not synchronize back your remote data. Please follow official guidelines to build and use a custom Docker image.

 

Is it possible to update the image used by my app?

To make your app use the updated version of a Docker image, follow this guide.

 

Connectivity

I don't understand how I can connect to my app

AI Deploy provides an HTTP endpoint for each deployed app. You can find your endpoint via the OVHcloud control panel (Public Cloud / AI Deploy / My app / Access URL), API, or CLI.

An HTTP endpoint will look like this: https://<unique_id>.app.us-east-va.ai.cloud.ovh.us

Your app will be directly exposed to this HTTP endpoint and linked to a port (by default, port 8080).

Depending on what you deployed, you then just have a REST endpoint or a Web interface. You can refer to our Getting Started guide for full explanations.

 

I'm unable to connect (unauthorized)

When you deploy an app, you can opt for unrestricted access (open to the internet) or secured access.

While unrestricted access means that everyone is authorized, secured access will require credentials. Two options are available:

  • An AI user. It can be seen as a user and password restriction. Quite simple but not a lot of granularity.
  • An AI token (preferred solution). A token is very effective since you can link them with labels. For example, a token for a specific app ID, for a team, ...

If you selected restricted access, don't forget to generate an applicative token.

 

I need more than one port to be exposed

By design, AI Deploy links your app to one HTTP endpoint and one port (default is 8080). If you need more than one port, the best practice is to split your deployment into multiple apps. If you cannot afford it, you can tweak your HTTP endpoint as follows: https://<unique_id>-<specific_port>.app.<region>.ai.cloud.ovh.us.

For example, your default app URL, starting with the app's ID and which accesses the default port, is https://00000000-0000-0000-0000-000000000000.app.us-east-va.ai.cloud.ovh.us. If you want to access port 9000, you will have to append the port number to your app's URL after the app's unique ID: https://00000000-0000-0000-0000-000000000000-9000.app.us-east-va.ai.cloud.ovh.us.

This way, you will be routed to this specific port even after the app has been launched.

You can also use gRPC on your AI Deploy apps. To do this, specify a port when you launch your app with the ovhai CLI, by using the --grpc-port <GRPC_PORT> attribute. You will then get your gRPC Address in your task info.

 

Billing

I don't understand how much it will cost to deploy an app

The AI Deploy pricing model is quite simple compared to competitors. You pay for the compute resources (CPUs/GPUs) during the lap of time you will use them.

  • Basic example: If you deploy one app with 2 x GPU at $1.00 for 6 hours, you will pay $12.00 at the end (2 x $1.00 x 6h), no matter the number of calls or users received.

Prices are shown statically on our official website, inside our Public Cloud section. For a dynamic estimation, use the OVHcloud Control Panel. An estimation will be available before launching a deployment.

Also, for more detailed information, please refer to our AI Deploy - Billing and lifecycle page.

 

I'm unable to get a "pay per call" deployment

So far, only a "pay per minute" model is available. We also share the ambition for a "pay per call" model, but it is not available for now.

 

Go further

For more information and tutorials, please see our other AI & Machine Learning support guides or explore the guides for other OVHcloud products and services.

If you need training or technical assistance to implement our solutions, contact your sales representative or click on this link to get a quote and ask our Professional Services experts for a custom analysis of your project.

Related articles