NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #978 most downloaded on PyPI
Ray provides a simple, universal API for building distributed applications.
Latest stable
2.60.0
Last release 18 days ago
04 Nov 2026
Ships fairly regularly
a new release about every 3 weeks
Nearly every release is documented
notes for 59 of the last 60 stable releases
1 version withdrawn
withdrawn after publishing
9 years old
137 releases · first in 2017
Drag across the ruler to compare, or tap a release to open it.
…CA compromise, and aiohttp moves to 3.14.4 for CVE-2026-111064 and CVE-2026-111066 ( #67295 , #67298 , #67301 , #67304 , #67305 , #67350 ).
max_surge_percent, which starts replacement replicas before stopping old ones so a rolling update doesn't reduce serving capacity (#66267). A rolling update whose new replicas keep failing now stops at DEPLOY_FAILED and leaves the remaining old replicas routable, instead of replacing every healthy replica (#66252, #66290).certifi 2026.10.19 and pip 26.2.2 after the eMudhra CA compromise, and aiohttp moves to 3.14.4 for CVE-2026-111064 and CVE-2026-111066 (#67295, #67298, #67301, #67304, #67305, #67350).Dataset.write_orc() for writing ORC files (#65453)unnest() expression that expands a struct-valued expression or column into one column per field (#65241)min_bytes_per_file to write_parquet to combine small blocks toward a target output file size (#65636)write_delta through schema_mode (#65858)RAY_DATA_ACTOR_AUTOSCALER=DEFAULT restores the old one (#66035)CheckpointConfig.checkpoint_manager_cls and checkpoint_filter_cls (#65676)ray.data.list_stats_summaries() developer API for execution stats of completed datasets (#66294)sort, and rename external shuffle to disk shuffle (#66260, #64875, #66367)hash_partitions (#66075, #66341, #66269)num_cpus=1 when num_cpus is unspecified; set num_cpus=0 to opt out (#66202)ds.stats() UDF time into input prep, function body, and output build, optionally per fused stage (#65807, #65810)memory is below 1.25x its observed max USS (#65512)OutputBackpressureGuard (#65202)pyarrow.dataset on import ray.data (#66463)read_lerobot (#65284)ACCESS_DENIED writes, partition value casts, and V1 reads that find no files (#66033, #66078, #65256)RandomAccessDataset (#61287)ResourceBudget backpressure stalling a pipeline indefinitely (#64601)limit(), which changed the rows limit().filter() returned (#66143)MapOperator with non-zero concurrency into a following AllToAllOperator (#66350)target_max_block_size=None in shuffle reduce (#66451)TorchInference outputs on device-to-host transfer (#65997)delta_timestamps keys and a zero height or width in read_images (#65857, #65967)write_webdataset docs and encoder typing, and a broken PyArrow link in write_parquet (#65278, #66154)RAY_SERVE_COLUMNAR_AUTOSCALING_METRICS=0 to revert (#64281, #66104, #66105, #66070, #66240). The controller health schema also reports how much of each control loop goes to ingesting metrics (#66106).max_surge_percent deployment option starts new replicas before stopping old ones, keeping capacity at the target during a rolling update; the default of 0 keeps the existing stop-then-start behavior (#66267).RAY_SERVE_USE_PACK_SCHEDULING_STRATEGY=1, once all deployments are stable the controller migrates replicas off a partially used node so the autoscaler can release it, starting replacements before stopping the originals so capacity never dips; tune it with RAY_SERVE_NODE_COMPACTION_DELAY_S and related settings (#66108).grpcurl, grpcui, and Postman can now discover and call Serve gRPC services through the gRPC proxy without local proto files; reflection is on by default and can be disabled with gRPCOptions(enable_reflection=False) or --grpc-disable-reflection (#65854).RoundRobinRouter. Direct-ingress deployments accept multiplexed model IDs over both HTTP and gRPC (#65678), and RoundRobinRouter prefers replicas that already host the requested model (#65679).PrometheusQueryMixin lets policies scale on Prometheus metrics by keeping PromQL query results fresh in a background thread, so reads never block on network calls (#65884).RAY_SERVE_RUN_SYNC_IN_THREADPOOL=0 (#66436)._to_object_ref(). DeploymentResponse._to_object_ref() now returns as soon as the request is assigned to a replica instead of waiting for the replica's result (#65243).DEPLOY_FAILED (including across controller restarts) instead of replacing every healthy old replica; opt out with RAY_SERVE_STOP_FAILED_ROLLING_UPDATES=0 (#66290).maxconn lowered from 20,000 to 4,096 to cap buffer memory (#66360).ray start --block receives SIGTERM), so Serve's existing logic migrates replicas and proxies off them instead of failing in-flight requests (#65502).max/min aggregation; the RAY_SERVE_AGGREGATE_METRICS_AT_CONTROLLER flag is removed, and because it defaulted to off this changes the default behavior (#65714). The controller also no longer re-aggregates requests just to write the scaling log line (#64654).root_path (breaking). root_path now follows the ASGI spec in every ingress mode, fixing direct ingress returning 404 for every request (including /-/healthz) and HAProxy stripping the prefix twice; uvicorn is upgraded to 0.53 (minimum 0.26). Requests that still include the root_path prefix now return 404, so strip it at your ingress or unset root_path (#66446).jinja2 import regression. from ray import serve no longer fails on slim images without jinja2 (a 2.57 regression); jinja2 is now imported lazily and declared in the serve extra (#65542).KVAwareRouter (#66325).KeyError that silently stranded requests (#66358).DeploymentUnavailableError rejections (#66323), and removing a deployment from an app with application-level autoscaling no longer raises a KeyError that stalls the application control loop (#66438).serve CLI reference moved to its own page, matching how the other libraries' API references are organized (#65943).metrics_interval_s warning. The autoscaling guide now warns that very low values can overload the controller, and the examples use the defaults (#65972).ray.io/v1. The examples, including the Serve fault-tolerance guide, moved from ray.io/v1alpha1, and a RayService example that used the unsupported serveConfig field (and so deployed zero applications) was fixed (#65336).QueueMonitorActor, its controller-side metric, and AutoscalingContext.total_pending_async_requests (which was always 0); AsyncInferenceAutoscalingPolicy already polls the broker directly (#66405).serve/_private type checking. mypy and pyrefly now scan the whole serve/_private tree instead of a hand-maintained file list, so new modules are covered by default (#65653).TorchTrainer (#66055)TorchCheckpoint, TensorflowCheckpoint, and XGBoostCheckpoint (#66179)record_stream on nested tensors in _WrappedDataLoader (#66344)_upload_to_fs_path (#65959)RunDisabled class so the integration works with upcoming wandb releases (#65736)evaluate and scikit-learn in the examples that use them (#65838, #65849)qrandint and qlograndint uniformly, and quantize when q is 1 (#65864, #65535)quantile_fraction is 0 (#65747, #65790)Repeater crash when the search finishes (#65960)MedianStoppingRule sample (#65957)resolved_vars (#65991)set_search_properties fallback for custom schedulers on Python 3.10+ (#65250)flatten_dict raising UnboundLocalError for lists with prevent_delimiter=True (#65621)mujoco-py engine and move MuJoCo environments to v4 and v5 (#65616)_update_add_metrics in the replay buffers (#66102)rllib- file prefix, apply a style pass, and fix broken links (#66062, #66060, #66214, #66239, #66338, #66234)local:// runtime_env URIs that use code already on the node without uploading it (#65245, #66302).tar.xz runtime_env archives, with archive validation unified across upload and download paths (#65742)$VAR and ${VAR} in uv package inputs and uv_pip_install_options (#65767)dispatch to run_on_slice (#65723, #66400, Note truncated.
Your coding agent can read these notes before it upgrades. Set up the MCP server →
…ray[default] , and the images. Fixes CVE-2026-111064 , CVE-2026-111066 and GHSA-2g87-pp6c-29x5 ( #67304 , pick #67487 ; reported in #67302 ).
The eMudhra intermediate CA keys were reported compromised on October 20, 2026. pip trusts its own vendored certifi bundle in addition to the system store, so updating OS ca-certificates alone doesn't protect pip. This release moves every place Ray installs or bundles a CA bundle to the fixed releases:
virtualenv 21.14.12, which bundles pip 26.2.2 for runtime_env virtualenvs. This also moves packaging to 26.3, typing-extensions to 4.16.0, distlib to 0.4.3 and python-discovery to 1.6.1 (#67301, pick #67487).certifi>=2026.10.19 floor in Ray's security requirements (#67305, pick #67487).aiohttp 3.14.3 → 3.14.4 (and yarl 1.23.0 → 1.25.1) in the bundled runtime_env agent dependencies, ray[default], and the images. Fixes CVE-2026-111064, CVE-2026-111066 and GHSA-2g87-pp6c-29x5 (#67304, pick #67487; reported in #67302).Thanks to everyone who contributed to this release!
Deprecations: ray_remote_args for read APIs ( #65501 ) and for Dataset transformations ( #65228 ); min_rows_per_file with partitioned Parquet writes
Highlights
🎉 New Features:
BackpressureConfig lets you return 429 instead of 503 when a request is rejected for exceeding max_queued_requests, and optionally attach a Retry-After header. This separates deliberate load shedding from "the deployment is broken" for load balancers, retry policies, and availability SLOs, and brings the HTTP path to parity with gRPC's RESOURCE_EXHAUSTED. Defaults are unchanged (#65193).TracingConfig model (enabled, exporter_import_path, sampling_ratio) can be passed to serve.start(tracing_config=...) or set in a Serve config file, replacing environment-variable-only setup and wiring tracing through the proxy and replica paths. Setup errors now fail fast instead of being silently swallowed (#63273).min_replicas=0; the autoscaler preserves the final 1 → 0 transition instead of rounding it back up to a full gang (#65575).💫 Enhancements:
jinja2 is now imported lazily and declared in the serve extra, so it is no longer pulled in on every Serve import (#65542, #65988).deploy_mode / ServeDeployMode, use_new_handle_api on DeploymentHandle.options, and the RAY_AGENT_ADDRESS deprecation path are gone, and a # in a deployment name is now a hard ValueError since it is the replica ID delimiter (#65215).ray.serve.request_router now declares __all__, making the custom-router extension API visible to the API consistency checks (#65485).🔨 Fixes:
ActorUnschedulableError and killed the controller loop; the handle is now cleaned up gracefully. This showed up with GCS fault tolerance backed by Redis, where the new cluster restores proxy actors pinned to pre-upgrade node IDs (#65076).📖 Documentation:
# restriction on deployment names and why it exists, so the new ValueError has something to point at (#65489).llms.txt, high-traffic landing pages and card grids converted from RST to MyST, explicit image versions in place of floating latest tags, and the "Ray Pod" term retired from the Kubernetes docs (#65115, #65467, #65472, #65339, #65423).🏗 Architecture refactoring:
controller, deployment_state, replica, application_state), the proxy and data-plane modules (proxy, proxy_state, haproxy), and the scheduling and routing modules (deployment_scheduler, router, request_router). Beyond annotations, this tightened several honest-return-type and ReplicaID-vs-str key confusions in state-machine code where type confusion becomes persistent state corruption (#64753, #65332, #65410).Many thanks to all those who contributed to this release!
@mikemikimike, @ShockYoungCHN, @elliot-barn, @aaronlinear, @spencer-p, @nadongjun, @marwan116, @Kunchd, @avigyabb, @suppagoddo, @harshit-anyscale, @n3sfan, @Smallfu666, @bveeramani, @pjdurden, @Jenson97, @gangli113, @ronny-anyscale, @wufeiye, @preneond, @ankushbbbr, @yuhuan130, @ixsanpe, @dossett, @rayhhome, @ArturNiederfahrenhorst, @robertnishihara, @CrossManger, @hogeheer499-commits, @iamjustinhsu,
Note truncated.
Fix out-of-scope block callbacks being registered on blocks passed through from input to output ( #66369 , cherry-pick #66404 ). Fixes spurious callba
Thanks to everyone who contributed to this release!
@marwan116, @jeffreywang88, @dstrodtman, @owenowenisme, @elliot-barn
Upgrade bundled dependencies: log4j 2.25.4, jackson-databind 2.18.8 ( CVE-2026-54512 , CVE-2026-54513 ), gson 2.11.0, aiohttp, idna, and azure ( #6426…
LLMRouter ingress replica, the routing decision is made there, tokens are transmitted out-of-band so the engine does not re-tokenize, KV lifecycle events are broadcast to every ingress replica (#64642, #64920, #64949, #65010, #65095). KV cache and token aware routing is also aware of CPU KV caches, so offloaded KV cache blocks count toward a replica's cache hit (#65063).RAY_enable_task_events_to_dashboard_head on, the task event buffer is replaced by the ray event recorder, events are exported from the aggregator agent to a task events head that keeps an in-memory store, and the state APIs and ray.timeline read from it (#64835, #65028, #65123, #65160, #65218). Enabling the feature removes task event ingestion and serving from the GCS hot path.SubslicePlacementGroup for gang scheduling on TPU subslices, single-host TPU support in SlicePlacementGroup, and resource accounting for tpu7x and multi-core chips (#64578, #64079, #64058). This lets TPU slices and subslices be reserved and trained on without external gang-scheduling glue.Dataset.with_columns for multi-column expression projection (#63858)write_delta for Delta Lake, with catalog support (#64923, #65079)ignore_missing_paths and skip_paths to read_parquet on DatasourceV2 (#65118)delta_timestamps (temporal windows) to read_lerobot (#64877)DataContext.max_consecutive_actor_init_deaths (#64846)RAY_DATA_HASH_SHUFFLE_MAP_TASK_TARGET_INPUT_BYTES in DataContext (#65103)Limit into ReadFiles when it sits directly on it, and extract FileIndexer.list_file_infos (#65167, #65168)BlockRefCounter for object store memory estimation and remove BlockRefCounter.clear() (#64456, #64521)BatchIterator and iter_torch_batches, and allow a custom collate_fn with a custom device (#64994, #64967)sort_reduce memory multiplier to 3x and stop using estimated_input_blocks as the shuffle partition count (#65176, #65296, #65335)PlacementGroupSchedulingStrategy is in use (#64417)reports_custom_op_stats to should_report_custom_op_stats across all MapTransformFn variants (#64461, #64515)allocated to reserved and add type aliases in the autoscaling coordinator (#64997, #65096)ray_remote_args_fn and Dataset.zip (#64963, #65111)read_lance or nested pickle objects could execute arbitrary code (#64881)null[pyarrow] in to_pandas (#65187)iter_torch_batches device resolution and typing (#65059, #64947)isolate_read_workers to DatasourceV2 (#65191)tf-keras to the text_embedding pip packages (#64889)use_datasource_v2 docstring default and an incorrect default_map_logical_memory_enabled reference (#65155, #65091)Retry-After header for backpressure rejections (#65193, #65319)RUNNING replicas round-robin instead of every tick (#64690)CreatePlacementGroupRequest.runtime_env as a dict (#64892)ASGIService bypassing token authentication (#65189)PreemptingState (#64360)contains_tensor check and add a serialization check for the results return value (#64930)PlacementGroupCleaner to the head node (#64705)HyperOptSearch dropping tune.choice categories that are constant dicts (#64537)test_multi_trial_reuse_with_failing and decide test_experiment_restore completion from measured progress (#64526, #65212)LLMRouter ingress replica, decide KV/token routing there, broadcast KV lifecycle events to all ingress replicas, and make selection and reservation atomic (#64642, #64920, #64949, #65010)apply_checkpoint_info (#62962)az:// Blob streaming with RunAI Streamer (#64819, #64825)target_qf_twin head in IQL target prediction (#64932)use_kl_loss in the PPO Torch and TF policies (#61562)KeyError in the multi-agent module-to-env connector (#64803)test_env_runner callback-count tests (#64807, #64989)ray.timeline are rerouted to it, with reconciliation on worker death and job completion (#64835, #65028, #65057, #65123, #65141, #65160, #65218, #65247, #65288)SubslicePlacementGroup for gang scheduling on TPU subslices, support single-host TPUs in SlicePlacementGroup, and add a per_slice_pgs parameter (#64578, #64079, #64072)mps) support and an Intel GPU ZE_AFFINITY_MASK mapping (#38464, #64440)tpu7x and multi-core chips, add gb200/gb300 accelerator constants, and add TTNPU custom accelerator resources (#64058, #65009, #61554)RedisContext::Connect non-fatal on connection failure (#64299)io_context (#65024)ActorPool instead of recycling them (#64646)label_domain to topology strategy in the scheduling policy (#64384)StateSchema column order in filter_fields (#65052)StreamResponse on an empty log stream (#62296)ObjectRefStreamEndOfStreamError from _get_next_ref_n (#64602)FunctionDescriptor rebuild in CallSiteString (#64874)canceled_tasks_ its own mutex to break a lock-order cycle (#65393, #65620)collect() in OpenTelemetryMetricRecorder (#64946)check_signals instead of exiting the process (#65184, #65400)OwnerDiedError during graceful raylet shutdown (#64899)ray.get on refs from a non-restartable streaming generator when those objects are lost (#64756)Push hits a stale local_objects_ mirror (#64916)spill_manager_objects_bytes reporting the restored object count instead of restored bytes (#65013)RDTManager that could SIGSEGV (#64558)Status::operator<< against an OK status, and fix UB in StatusOr swap and assignment on error-state operands (#64983, #64799)ResourceRequest fields in operator== (#64838)kubernetes import (#64962)ray-torch release test image (py3.14, cu12.8) and a hello_world_py314 smoke release test (#65114, #64857)pybase64 into the ML release-test image (#64980, Note truncated.
Fix out-of-scope block callbacks being registered on blocks passed through from input to output ( #66369 , cherry-pick #66403 ). Fixes spurious callba
Thanks to everyone who contributed to this release!
@marwan116, @jeffreywang88, @elliot-barn
One column per quarter.