●Change8

ray-2.42.0

📦 rayView on GitHub →
✨ 13 features🐛 8 fixes⚡ 1 deprecations🔧 16 symbols

Summary

This release adds audio/video reading to Ray Data, introduces new Train and Tune callbacks, stabilizes the DeploymentHandle API, and brings numerous bug fixes and performance enhancements across Ray libraries including Serve, RLlib, and Core.

Migration Steps

  1. Replace usage of num_rows_per_file with min_rows_per_file in Ray Data APIs.

✨ New Features

  • Added read_audio and read_video functions to Ray Data.
  • Introduced Train v2 user-facing callback interface.
  • Added TuneReportCallback for propagating intermediate Train results to Tune.
  • Cache metrics in replica and report on an interval in Ray Serve.
  • Cache expensive calls to inspect.signature in Ray Serve.
  • Removed extra pickle serialization for gRPCRequest in Ray Serve.
  • Shared LongPollClient for Routers in Ray Serve.
  • DeploymentHandle API is now stable in Ray Serve.
  • Added metrics to replay buffers in RLlib.
  • Enhanced node-failure tolerance with a new API stack in RLlib.
  • Split AddStates... connectors into AddTimeDimToBatchAndZeroPad and AddStatesFromEpisodesToBatch in RLlib.
  • Only get single node info rather than all when needed in Ray Core.
  • Introduced with_tensor_transport API in Ray Core.

🐛 Bug Fixes

  • Fixed non-deterministic read tasks bug in Ray Data.
  • Fixed BayesOptSearch documentation.
  • Fixed batched requests hanging after request cancellation bug in Ray Serve.
  • Fixed tqdm manager thread safety issue.
  • Fixed token expiration for Ray autoscaler.
  • Re-introduced mixin-replay-buffer pass for IMPALA/APPO when replay-ratio=0, fixing a memory leak.
  • Fixed MetricsLogger race conditions.
  • Fixed APPO/IMPALA bug for > 1 Learner actor.

🔧 Affected Symbols

ray.data.read_audioray.data.read_videoray.data.num_rows_per_fileray.data.min_rows_per_fileray.train.TrainCallbackV2ray.tune.TuneReportCallbackray.serve.DeploymentHandleray.serve.LongPollClientray.serve.gRPCRequestray.rllib.replay_buffersray.rllib.metrics_loggerray.rllib.connectors.AddTimeDimToBatchAndZeroPadray.rllib.connectors.AddStatesFromEpisodesToBatchray.core.with_tensor_transportray.core.get_node_inforay.autoscaler.token

⚡ Deprecations

  • Deprecated num_rows_per_file in Ray Data in favor of min_rows_per_file.