Neptune ML limits - Amazon Neptune
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

Neptune ML limits

  • The types of inference currently supported are node classification, node regression, edge classification, edge regression and link prediction (see Neptune ML capabilities).

  • The maximum graph size that Neptune ML can support depends on the amount of memory and storage required during data preparation, model training, and inference.

    • The maximum size of memory of a SageMaker data-processing instance is 768 GB. As a result, the data-processing stage fails if it needs more than 768 GB of memory.

    • The maximum size of memory of a SageMaker training instance is 732 GB. As a result, the training stage fails if it needs more than 732 GB of memory.

  • The maximum size of an inference payload for a SageMaker endpoint is 6 MiB. As a result, inductive inference fails if the subgraph payload exceeds this size.

  • Neptune ML is currently available only in Regions where Neptune and the other services it depends on (such as Amazon Lambda, Amazon API Gateway and Amazon SageMaker) are all supported.

    There are differences in China (Beijing) and China (Ningxia) having to do with the default use of IAM authentication, as is explained here along with other differences.

  • The link prediction inference endpoints launched by Neptune ML currently can only predict possible links with nodes that were present in the graph during training.

    For example, consider a graph with User and Movie vertices and Rated edges. Using a corresponding Neptune ML link-prediction recommendation model, you can add a new user to the graph and have the model predict movies for them, but the model can only recommend movies that were present during model training. Although the User node embedding is calculated in real-time using its local subgraph and the GNN model, and can therefore change with time as users rate movies, it's compared to the static, pre-computed movie embeddings for the final recommendation.

  • The KGE models supported by Neptune ML only work for link prediction tasks, and the representations are specific to vertices and edge types present in the graph during training. This means that all vertices and edge types referred to in an inference query must have been present in the graph during training. Predictions for new edge types or vertices cannot be made without retraining the model.

SageMaker resource limitations

Depending on your activities and resource usage over time, you may encounter error messages saying that you've exceeded your quota (ResourceLimitExceeded). and you need to scale up your SageMaker resources, follow the steps in the Request a service quota increase for SageMaker resources procedure on this page to request a quota increase from Amazon Support.

SageMaker resource names correspond to Neptune ML stages as follows:

  • The SageMaker ProcessingJob is used by Neptune data processing, model training, and model transform jobs.

  • The SageMaker HyperParameterTuningJob is used by Neptune model training jobs.

  • The SageMaker TrainingJob is used by Neptune model training jobs.

  • The SageMaker Endpoint is used by Neptune inference endpoints.