Skip to main content



Back-end administration (for DevOps).

What happens to models that are running on a node that goes down?

The models are immediately replicated among the remaining nodes in the Cluster.

Can Promote installations be automated?

Versions previous to 2018.4 can be automated. Version 2018.4 and beyond cannot be automated since the installation is done through a Web UI.

Are the Dev, Staging, and Production environments physically separated?

The Dev, Staging, and Production environments are not physically separated. When a new model is deployed, a container is created with the model dependencies, and the user that created the model has their username appended to the container and replica names that are created. If the model is transferred to the Production environment a copy of the model container is created with two replicas, and the production username replaces the old username for that model.

Why does Promote require at least three nodes?

We use the Raft consensus. Please visit the article on Docker's site.

Can I persist data in existing databases?

No, the Promote platform itself does not support this feature, but it's possible for you to create your own feature.

Can Promote be installed behind a proxy?

Yes, but it depends on how the proxy is set up and how it's being used with Promote based on the system requirements.

Why do I need my own external load balancer?

If the Master node goes down, the load balancer is required to redirect incoming traffic to the remaining nodes in the cluster.

Can Promote be installed on machines without outbound internet traffic?

No, because if you have to restart Promote the platform will attempt to ping the Quay repository for image updates or other repositories for package updates.

Model deployment & management (for data scientists)

Is there a limit to the number of models that can be deployed to a Promote cluster?

Yes, 100 is the limit. This limit is dependent upon the number of replications defined for each model, and whether you are increasing or decreasing model replications. There is a default of two replications for each model when it is first deployed.

Can Promote models be set up to be automatically retrained?

No, Promote does not currently provide this feature however, you can potentially try retraining models with a scripting language.

Can I connect to external databases from my deployed models?

Yes, but it is required that you install the appropriate connectors to the container layers that are hosting the predictive models.

Can I track model performance metrics with Promote?

No, model metrics cannot be tracked, but you can track how often a model is accepting HTTP requests via the model history and model analytics tabs.

Model inference (for engineers)

Is there a way to prioritize prediction requests?

No, the prediction requests are handled randomly, via round-robin.

How does Promote communicate prediction errors?

An error is returned through the HTTP request. This error can be reviewed in the prediction analytics section.

What is the latency overhead of Promote?

Using a simple Hello World model, the latency between sending a POST and receiving a response is 100 milliseconds.

Can Promote handle batch data?

No, Promote currently handles real-time data only.