Metadata for Custom ML Runtimes

This topic addresses the metadata for custom Runtimes.

All new custom Runtimes must override the Edition metadata of existing Runtimes. The rest of the metadata can be overriden to communicate the expectations for the consumers of the image. Both the Docker label and the environment variable match in a custom Runtime image. In order to add or register a custom Runtime to a deployment, the user facing metadata combination should be unique in that deployment. For example, of the following, Editor, Edition, Kernel, Version, and Maintenance Version, at least the later should be incremented for adding a next iteration of the same image.

See the following reference table for more details on ML Runtime metadata.

Environment variable Docker Label Description Override in custom runtime
ML_RUNTIME_METADATA_VERSION Metadata version Not allowed
ML_RUNTIME_EDITOR CDSW/CML Editor installed in the image. Allowed
ML_RUNTIME_EDITION Edition of the image, a notion of the Runtime capabilities. Required
ML_RUNTIME_DESCRIPTION Longer description of the Runtime image capabilities. Recommended
ML_RUNTIME_KERNEL Main kernel included in the image, e.g., Python 3.8 Not allowed
ML_RUNTIME_SHORT_VERSION Main version of the image, e.g., 1.0. This shows up as Version in the selection screen. Recommended
ML_RUNTIME_FULL_VERSION Full version consists of the short version + maintenance version, e.g., 1.0.1 Recommended
ML_RUNTIME_MAINTENANCE_VERSION Maintenance version must be an integer, e.g., 1. Only the largest maintenance version of the set of the same short version images are visible for users to select. Increment this number if you create a drop in replacement of an existing Runtime. Start with 1 with a new version or edition. Recommended
ML_RUNTIME_CUDA_VERSION CUDA version installed in the image. Not allowed
ML_RUNTIME_GIT_HASH Git hash of runtime source Not allowed
ML_RUNTIME_GBN Cloudera internal build number Not allowed