v0.6.0

Highlights

• Added the Python Runner API for packaging local Python projects and exposing functions, classes, or object instances as Flame services.
• Added flmrun, a built-in Runner template application used by Runner-managed services.
• Fixed Runner package downloads for multi-object-cache deployments by preserving the cache endpoint returned from package upload.
• Reworked the Python SDK around synchronous client APIs, object-cache helpers, service sessions, and the flamepy.runner module.
• Simplified the Rust SDK with top-level connection/session helpers, typed task messages, host-service helpers, object-cache helpers, and service macros.
• Added a standalone object cache with versioned object references, pluggable storage, upload/download support, incremental object fetch, and fast paths for tabular/numpy/Arrow data.
• Added flmadm installation profiles and support for installing Flame components, SDKs, examples, and multiple Python SDK versions.
• Added flmctl deploy and object helper commands for application deployment and object-cache workflows.
• Added a Helm chart and Kind-based Kubernetes E2E workflow for static Flame installs on Kubernetes.
• Updated Helm/Kubernetes defaults for filesystem session-manager storage and the drf/gang policy set.
• Fixed Flame CLI version metadata so binaries report their package version.
• Added scheduler policy work for priority scheduling, GPU-aware DRF, dynamic policy configuration, gang/batch flows, and resource requirements.
• Added session/runtime reliability improvements including TLS, executor recovery, session bind failure recovery, task watching, node/executor persistence, and notifier-based wakeups.
• Expanded system, E2E, Runner, cache, Python SDK, Rust SDK, and storage test coverage.

Upgrade Notes

• Python SDK packaging is now versioned as 0.6.0, requires Python 3.9 or newer, and advertises Python 3.9 through 3.12.
• Runner imports should use flamepy.runner; older flamepy.rl naming was replaced.
• flmadm install now requires an explicit profile flag such as --all, --control-plane, --worker, --cache, or --client.
• Cluster policy configuration now supports multiple policies through the policies list.
• The Helm chart defaults now configure session-manager storage as fs:///var/lib/flame/session and enable only the drf and gang policies.
• Object cache references are versioned. Clients can use get_object, update_object, patch_object, upload_object, and download_object instead of passing raw object payloads through sessions.
• flmexec supports explicit runtime selection for scripts, including Python runtime selection through FLAME_PYTHON_VERSION. Its default Python runtime policy is owned by flmexec rather than the Rust SDK.

Python SDK And Runner

• Added flamepy.Runner and flamepy.runner to package local projects and run Python execution objects remotely (#298, #341, #342).
• Added Runner helpers for waiting, selecting, resolving object references, and fetching result objects (#300, #363).
• Added explicit Runner defaults for stateless functions/classes and stateful object instances (#474).
• Fixed Runner packaging for ad hoc scripts and dependency metadata generation (#471).
• Fixed Runner package URL generation to use the returned object-cache endpoint, so executor-manager downloads packages from the cache pod that stored them in multi-cache deployments (#482).
• Added support for Runner dependencies and Python-version selection in Runner/flmexec workflows (#468, #469, #473).
• Added flmrun as a built-in application template for Runner services (#285).
• Added Python service-session helpers and renamed the FlamePy agent API to service sessions (#454).
• Added custom session IDs and open_session enhancements for Python clients and Runner services (#351, #353, #354).
• Added Session.list_tasks and task watching improvements for Python clients (#359, #371).
• Aligned flamepy package metadata for the 0.6 release, including __version__, Python 3.9+ tooling targets, and Python 3.12 classifier support (#477).

Rust SDK

• Simplified the flame-rs API with direct helpers for connecting, creating/opening sessions, running typed tasks, and writing services (#456).
• Added typed service macros and typed task payload support through FlameMessage (#456).
• Added Rust object-cache helpers for putting, getting, updating, patching, uploading, downloading, and deleting objects (#427, #430, #456).
• Kept flmexec Python runtime default policy local to flmexec and removed that default from the public Rust SDK surface (#475).
• Fixed Rust SDK logger initialization when RUST_LOG is unset (#343).

Object Cache And Storage

• Added the standalone flame-object-cache service and object-cache client helpers (#244, #321).
• Added common-data/object-cache integration and object references for shared session state (#258, #259, #269, #296).
• Added LRU eviction and per-application cache behavior (#367, #358).
• Added pluggable cache storage backends, object versioning, and a standalone cache binary (#419, #427).
• Added cache upload/download support with multi-scheme downloaders (#430).
• Added incremental object get, native tabular cache path, and fast-path serialization for numpy/Arrow data (#444, #446, #464).
• Added filesystem, HTTP, and none storage engines for session manager and package/cache workflows (#339, #344, #377, #395).

Scheduling And Runtime Management

• Added batch-session support and gang-related E2E coverage (#401, #407).
• Added priority scheduling and dynamic scheduler policy configuration (#428, #432).
• Added GPU-aware DRF scheduling with resource requirements (#434).
• Added resource requirement support in the priority plugin (#442).
• Added configurable scheduling interval and moved executor limits into limits (#373, #396).
• Added application URL support, installer metadata, and application cache-key improvements (#287, #421, #425).

CLI, Installation, And Local Development

• Added flmadm for installation, profile-based component selection, systemd integration, user-local installs, examples, and uninstall flows (#334, #338, #421, #468).
• Added flmctl deploy and object helper commands (#459).
• Added JSON output for session commands and improved CLI display behavior (#215, #439).
• Changed Flame CLI metadata to derive --version output from package versions and use XFLOPS <[email protected]> as the author (#488).
• Added Podman support and local development helpers (#212).
• Added Docker, compose, and local system-test workflows for multi-component clusters (#345, #347, #466, #470, #472).

Kubernetes And Helm

• Added charts/flame, a static Helm chart for installing Flame's session manager, executor manager, object cache, client configuration, ServiceAccount, services, persistent volumes, and Helm test resources (#479).
• Added chart values and schema coverage for images, service ports, storage, runtime volumes, TLS Secret mounting, client config, component enablement, and static object-cache replica counts (#479).
• Updated chart defaults and Kubernetes E2E overrides to use filesystem session-manager storage and only the drf and gang scheduler policies (#488).
• Added a Kind-based Kubernetes E2E workflow that builds Flame images, installs the Helm chart, runs Helm tests, and verifies flmctl, flmping, and the Python Pi Runner example from an in-cluster console pod (#479).

Reliability, Recovery, And Observability

• Added TLS support and certificate-generation fixes (#388, #411).
• Added Flame recovery flows and fixed executor state recovery behavior (#386, #403, #465).
• Added node/executor persistence for flmctl and session-manager recovery paths (#383).
• Added notifier-based wakeups to reduce busy waiting (#417).
• Added task/event management and task event recording in Rust and Python services (#238, #239, #241, #242).
• Improved executor working-directory validation, stdout/stderr log placement, and cleanup on release (#360, #361, #362).

Examples And Documentation

• Added API reference pages, SDK guides, Runner setup guide, Rust API tutorial, and local-development guide (#451, #461).
• Added and updated examples for Python/Rust Pi, OpenAI agents, LangChain agents, SRA, crawler, Candle, RL, replay buffer, and TorchRL DQN (#264, #299, #424, #441, #450, #453, #462).
• Added the Python Pi example to flmadm --with-examples installs and made its workload configurable for fast E2E validation (#479).
• Added design documents for the major 0.6 feature areas, including Runner, cache, storage, scheduler, TLS, recovery, app installer, Helm installation, and SDK simplification.
• Updated README positioning to "A Distributed Engine for AI" and refreshed project badges (#443).

Testing

• Added Python SDK unit coverage for cache, core, Runner, and service behavior.
• Added Runner storage regression coverage for returned object-cache endpoints and TLS endpoint normalization (#482).
• Added Rust SDK tests for typed services, macros, object cache, and integration flows.
• Added E2E coverage for sessions, cache, Runner, flmexec, applications, system flows, and scheduler behavior.
• Added Kubernetes install coverage for the Helm chart, multi-replica object cache, in-cluster client configuration, and Runner-based Python Pi execution (#479).
• Added release-branch E2E coverage for the Python Runner command path (#485).
• Added daily system tests and aligned the system workflow with the CI job format (#466, #470, #472).

Fixes

• Fixed fair-share allocation and priority-plugin session desired values (#276, #436).
• Fixed cache/session state updates, local instance behavior, task failure reporting, and shutdown queue handling (#214, #235, #243, #315).
• Fixed package/dependency caching through UV_CACHE_DIR and Runner dependency handling (#356, #357).
• Fixed Runner package downloads in multi-replica object-cache deployments by using the uploaded object's returned cache endpoint (#482).
• Fixed Docker image rebuild behavior and added missing tooling to the executor-manager image (#348, #385).
• Fixed Unix socket path length handling for host shim communication (#405).
• Fixed flmping output and README/example drift (#217, #256, #328).

Contributors

Thanks to everyone who contributed changes in this release cycle:

• Klaus Ma (@k82cn)
• Zhe Jin (@jinzhejz)
• Shida Qiu (@SataQiu)
• Xavier Chang (@Monokaix)