Audience¶
Who is Seldon Deploy for?¶
Seldon Deploy provides benefits for a range of different roles involved in machine learning operations. These span from operational concerns to ML performance to management and compliance.
Data scientists can benefit from Seldon Deploy while iteratively developing their models and when these are in production.
They can take advantage of the following features in particular:
Catalog of model versions, including model metadata
Model performance monitoring, including drift and outlier detection
Explanations of model predictions, to understand and improve models
Traffic mirroring, to safely experiment with new models
Operational aspects are key to productionising ML pipelines. Seldon Deploy enables engineers to deploy safely and easily.
The key benefits that Seldon Deploy offers are:
GitOps for reliable, reproducible, auditable deployments
Canary rollouts for mitigating risk, with explicit canary promotion
Real-time monitoring of metrics including model latency, resource consumption, and error rates
Management of secrets for model artifacts and image registries
API and Python SDK for automation of common and complex workflows, such as CI/CD tasks
Managing your ML estate is no easy task, especially as new model versions are developed and your team’s scope increases.
Seldon Deploy helps managers keep on top of their estates:
Projects keep groups of related models together
Namespaces provide strong resource isolation, for example to separate development and production workloads
Access controls mitigate risk by defining who can do what and where
Catalog of model versions, including model metadata
ML systems are complex and powerful tools that need to be fair and unbiased for many real-world applications. They are often subject to regulatory and compliance checks to ensure this.
Seldon Deploy assists auditors with the following features:
Audit logs to know who did what and when
GitOps for deployments, to ensure the system state is what it’s meant to be
Explanations for model predictions, to understand why a model did what it did and to check for bias
Access controls, so that only authorized parties are allowed to see sensitive data or make changes