Updating models and manifests in containers

Integration Partners with containerized solutions can update the models that power their containers by uploading model update artifacts to their Amazon S3 bucket. A model update requires two types of files:

  • Models: Contains the decisioning and scoring data used at run-time.

  • Manifests: Lists the model data files included in the model update to the container image.

Because model updates can be made without affecting the container code that is currently in production, you can refresh your containerized solutions without rebuilding the entire container image. This accelerates iteration speed, reduces downtime, and keeps models current without requiring a full redeployment. You can focus on improving performance and accuracy, while Index ensures secure and reliable deployment.

How does it work?

Step Description

You upload artifacts to your dedicated S3 bucket. For details on how to structure these artifacts, see Requirements and structure for S3 folders, manifests, and models.

An Index artifact sync job periodically polls these S3 buckets to check for new uploads and to perform security scanning. Index automatically scans uploaded artifacts to make sure that they:

  • Conform to supported formats (JSON, YAML, CSV, etc.).

  • Are non-executable.

  • Do not exceed the maximum supported size (≤0.5 GB per file).

Containers will be redeployed with updated artifacts:

  • Validated model data and manifest files are made available to your container.

  • Containers read files from the predefined MODEL_PATH environment variable.

  • Containers will be redeployed and begin referencing the latest manifest and associated models.

  • Model data is available in all active data centers.

Requirements and structure for S3 folders, manifests, and models

When creating your S3 folder, and your model and manifests files, be aware of the following requirements:

  • Each folder in the root bucket directory, along with the files it contains, is treated as a single model containing one manifest and one or more model data files.

  • Model folders must use CalVer naming conventions. This makes sure that Index can detect when a new model folder should be synced and apply the correct model_version tag for telemetry.

  • Model folders should not be modified in place once uploaded to S3 because they won’t be re-synced. Instead, create a new model folder with a bumped version.

  • You must grant both Read-Only and List access to Index so new artifacts can be detected and retrieved automatically.

Example S3 folder structure

It's recommended that you format your S3 folder like the following:

2025.08.20.0/ 
	manifest.json 
	model_data/ 
		CarBrand1-model.csv 
2025.08.20.1/ 
	manifest.json 
	model_data/ 
		CarBrand1-model.csv 
		CarBrand2-model.csv

Example manifest

It's recommended that you format your manifests like the following:

{ 
	"name": "Manifest for August 20th", 
	"models": ["CarBrand1-model.csv"," CarBrand2-model.csv"] 
}

Example model

It's recommended that you format your models like the following:

domain,country,weight 
carbrand1.com,CA,1 
carbrand2.com,US,0.5

Monitoring and metrics

Your container should expose the following metrics to help Index evaluate performance:

  • model_data_evaluations counter metric: Tracks which models are being evaluated by the container.

    • model_data_file tag attribute (optional): Identifies the model data file being actioned by name.

    • result tag attribute: Indicates whether the model matched a given request (true/false).

The Helm chart (the Kubernetes deployment configuration) is set up so that every metric is automatically tagged with:

  • image_version: Identifies which version of the container image is running (e.g. v1.3.2).

  • model_version: Identifies which model folder the container is referencing (e.g. 2025.01.01.0).

These tags allow Index to monitor performance trends and detect changes as new images and models are deployed.

Your responsibilities

As an Integration Partner, you are expected to:

  • Structure your S3 bucket and give Index Read-Only and List access as described above.

  • Expose required runtime metrics.

  • Make sure that your manifests correctly reference the intended model files.

  • Test new models locally before publishing.

  • Limit updates to no more than two times per day.

  • Monitor container performance.

Index responsibilities

Index is expected to:

  • Detect new artifacts in your S3 bucket automatically.

  • Perform all required security scans.

  • Deploy validated models and manifests into your container.

  • Pause deployments if issues arise.

Service level agreement

Model deployments are typically processed during Index business hours (EST) during the next business day.

If you have any questions or need assistance, contact your Index Representative.