Polyaxon

Polyaxon

Standalone platform

Open-source, managed control plane, and commercial enterprise features.

All editions (free and paid) allow on-premises deployment.
The community edition is full on-premise and can be deployed on any k8s cluster.
The cloud edition is a hybrid deployment where the control plane is managed and the workload, code, logs, metrics, models, and artifacts are on the customer’s private cloud account or on-prem cluster(s).
The enterprise edition allows to deploy both the control plane and agents (data plane) on the private cloud account or the on-prem cluster(s).

Seat-based license pricing, more info on https://polyaxon.com/pricing/.

SLAs can be purchased.

Community Edition comes with community support.
All paying customers have access to email and slack support.
Additional support priority and SLAs can be purchased.

No user management in the community edition.
User management is available in all paid versions with organization-level roles, ACL, and RBAC, all users can additionally issue scoped tokens to interact with the APIs on their behalf.
Business and Enterprise plans have access to advanced restrictions, additional team-level roles, and management.
SSO with SAML is available to enterprise customers.

Advanced logical restrictions and native integration with K8S RBAC.
Workload, data, code, artifacts, logs, and models are always on the customer's clusters in all editions.

All editions can be deployed in minutes.
The hybrid deployment with cloud requires the least amount of maintenance since the agent is self-healing and only executes instructions from the control plane.
Artifacts are stored in/fetched from a blob storage, a volume, or a NAS managed and configured by the customer on their clusters using our deployment config file.
Datasets and connections are accessed and configured by the customer on their clusters using our deployment config file.

Support multiple types of artifacts stores, connections, and backends.
Workload can easily scale to any large number of parallel executions supported by a k8s cluster, on a single namespace or multi-namespace/multi-cluster configuration, we have an abstraction that is k8s-native for controlling the executions, archiving logs, collecting outputs and meta-data without disrupting the user’s logic.
The platform features an API and a set of language SDKs to integrate and interoperate with external systems.

"The platform makes the experiments and operations reproducible, portable, and repeatable while being language and framework agnostic

Every operation in Polyaxon is authored using a powerful specification and packaging format called Polyaxonfile, the specification tracks dependencies, inputs, outputs, artifacts, environments, and runtime of an operation to schedule on Kubernetes.
The platform provides a built-in extensive tracking system to log information for source code, parameters, data, metrics, tags, and logs.
All operations scheduled with the platform will go through a lineage auto-detection powered by a compiler and the environment executor to track additional information about its contexts, IO params, statuses, metrics, hyperparams, source code, data, visualizations, artifacts, and resources used in each experiment.

Polyaxon python library is a tool that supports logging several types of metadata, events, and artifact types.
Users can log single events/artifacts or versioned events/artifacts within the same execution.
The API/CLI/UI provides logs streaming and archives rendering interface.
The UI provides both a rich dashboard and artifacts rendering engine that handles several formats: line charts, bar charts, box plots, scatter plots, histograms, parallel coordinate plots, … audio, video, and image events, resources consumption, custom charts based on Matplotlib/Plotly/Bokeh/Altair/Vega, notebooks rendering, Markdown, Netron.
Users can also build and schedule Tensorboards and custom dashboards using several integrations.

Users can use a table comparison and a multi-run dashboard comparison together or separately.
All metadata, lineage information, and tracked events can be compared using one or both tools.
Both the table comparison and the events dashboard are managed with a search bar, several quick filters, and a sorting logic
Users can also schedule multi-run Tensorboards, Hiplot, or build custom downstream operations that listen to several experiments.

Provides a query language to search, filter, and sort runs.
The query languages handle scalar filters for metrics, value filters for inputs/outputs, boolean filters for flags, DateTime filters for execution timeline and state transitions, text and regex search capabilities, filters based on resources, status, lineages, agents, queues, and dependencies.
Users can store their search specifications and share them with the rest of the team
Besides the UI, the search abstraction and query language are also available via CLI/Client/APIs for runs, projects, models, and components.

Members can create new models during the experimentation phase, experiments can be bookmarked and, optionally, successful experiment's artifacts can be promoted to the model registry to be reviewed and tagged by an admin.
Each model has a version (tag) similar to a docker image and can go through several maturity stages controlled by a model admin.
Experiments that are promoted to the model registry are automatically locked and cannot be edited, resumed, or removed.

The platform comes with a DAG abstraction, schedules based on DateTime or cron specification, and external and internal events consumption to trigger operations.
Cluster admins can easily integrate a git provider using an auth token or an ssh connection.
The documentation page features several integrations with Github actions and examples to work with Argo workflows, Airflow, or Kubeflow pipelines.

The documentation page features over hundreds of integrations and examples to work with external systems, libraries, and frameworks.
Additionally, users can extend the system using the API, CLI, or their Polyaxonfile manifest by exposing their containers with inputs and outputs and environment requirements.
Each integration can be mixed with additional presets to preconfigure resources, scheduling, connections, and access checks.