Ray

Breaking Changes
- The `timeout` argument of `ray.wait` now uses seconds instead of milliseconds. 3706

Core
- Limit default redis max memory to 10GB. 3630
- Define a `Node` class to manage Ray processes. 3733
- Garbage collection of actor dummy objects. 3593
- Split profile table among many keys in the GCS. 3676
- Automatically try to figure out the memory limit in a docker container. 3605
- Improve multi-threading support. 3672
- Push a warning to all users when large number of workers have been started. 3645
- Refactor code `ray.ObjectID` code. 3674

Tune
- Change log handling for Tune. 3661
- Tune now supports resuming from cluster failure. 3309, 3725, 3657, 3681
- Support Configuration Merging for Suggestion Algorithms. 3584
- Support nested PBT mutations. 3455

RLlib
- Add starcraft multiagent env as example. 3542
- Allow development without needing to compile Ray. 3623
- Documentation for I/O API and multi-agent improvements. 3650
- Export policy model checkpoint. 3637
- Refactor PyTorch custom model support. 3634

Autoscaler
- Add an initial_workers option. 3530
- Add kill and get IP commands to CLI for testing. 3731
- GCP allow manual network configuration. 3748

Known Issues:
- Object broadcasts on large clusters are inefficient. 2945

Core
- Added experimental option to limit Redis memory usage. 3499
- Added option for restarting failed actors. 3332
- Fixed Plasma TensorFlow operator memory leak. 3448
- Fixed compatibility issue with TensorFlow and PyTorch. 3574
- Miscellaneous code refactoring and cleanup. 3563 3564 3461 3511
- Documentation. 3427 3535 3138
- Several stability improvements. 3592 3597

RLlib
- Multi-GPU support for Multi-agent PPO. 3479
- Unclipped actions are sent to learner. 3496
- `rllib rollout` now also preprocesses observations. 3512
- Basic Offline Data API added. 3473
- Improvements to metrics reporting in DQN. 3491
- AsyncSampler no longer auto-concats. 3556
- QMIX Implementation (Experimental). 3548
- IMPALA performance improvements. 3402
- Better error messages. 3444
- PPO performance improvements. 3552

Autoscaler
- Bug fixes, botocore handling fix. 3454 3503 3447

Ray Tune
- Lambdas now require `tune.function` wrapper. 3457
- Custom loggers, sync functions, and trial names are now supported. 3465
- Improvements to fault tolerance. 3414
- Variant Generator docs clarification. 3583
- `trial_resources` now renamed to `resources_per_trial`. 3580

Modin
- Modin 0.2.5 is now bundled with Ray
- `import modin` after `import ray`

Breaking Changes
- Renamed `_submit` to `_remote`. 3321
- Object store memory capped at 20GB by default. 3243
- Now `ray.global_state.client_table()` returns a list instead of a dictionary.
- Renamed `ray.global_state.dump_catapult_trace` to `ray.global_state.chrome_tracing_dump`.

Known Issues
- The Plasma TensorFlow operator leaks memory. 3404
- Object broadcasts on large clusters are inefficient. 2945
- Ape-X leaks memory. 3452
- Action clipping can impede learning (please set clip_actions: False as a workaround) 3496

Core
- New raylet backend on by default and legacy backend removed. 3020 3121
- Support for Python 3.7. 2546
- Support for fractional resources (e.g., GPUs).
- Added `ray stack` for improved debugging (to get stack traces of Python processes on current node). 3213
- Better error messages for low-memory conditions. 3323
- Log file names reorganized under `/tmp/ray/`. 2862
- Improved timeline visualizations. 2306 3255

Modin
- Modin is shipped with Ray. After running `import ray` you can run `import modin`. 3109

RLlib
- Multi agent support for Ape-X and IMPALA. 3147
- Multi GPU support for IMPALA. 2766
- TD3 optimizations for DDPG. 3353
- Support for Dict and Tuple observation spaces. 3051
- Support for parametric and variable-length action spaces. 3384
- Support batchnorm layers. 3369
- Support custom metrics. 3144

Autoscaler
- Added `ray submit` for submitting scripts to clusters. 3312
- Added `--new` flag for ray attach. 2973
- Added option to allow private IPs only. 3270

Tune
- Support for fractional GPU allocations for trials. 3169
- Better checkpointing and setup. 2889
- Memory tracking and notification. 3298
- Bug fixes for `SearchAlgorithm`s. 3081
- Add a `raise_on_failed_trial` flag in run_experiments. 2915
- Better handling of node failures. 3238

Training
- Experimental support for distributed SGD. 2858 3033

API
- Add `ray.is_initialized()` to check if `ray.init()` has been called. 2818

Fixes and Improvements
- Fix issue in which `ray stop` fails to kill plasma object store. 2850
- Remove dependence on `psutil`. 2892

RLlib
- Set better default for VF clip PPO parameter to avoid silent performance degradation. 2921
- Reward clipping should default to off for non-Atari environments. 2904
- Fix LSTM failing to train on truncated sequences. 2898

Tune
- Fixed a small bug in trial pausing and cleaned up error messages. 2815

Breaking Changes
- Local mode has changed from `ray.init(driver_mode=ray.PYTHON_MODE)` to `ray.init(local_mode=True)` to improve clarity.

Autoscaler and Cluster Setup
- Added many convenience commands such as `ray up`, `ray attach`, `ray exec`, and `ray rsync` to simplify launching jobs with Ray.
- Added experimental support for local/on-prem clusters.

RLlib
- Added the IMPALA algorithm.
- Added the ARS algorithm.
- Added the A2C variant of A3C.
- Added support for distributional DQN.
- Made improvements to multiagent support.
- Added support for model-based rollouts and custom policies.
- Added initial set of reference Atari results.

Tune
- `SearchAlgorithm`s can now be used separately from `TrialScheduler`s and are found in `ray.tune.suggest`.
- All `TrialScheduler`s have been consolidated under `ray.tune.schedulers`.
- Minor API changes:
- For `Experiment` configuration, `repeat` has been renamed to `num_samples`.
- Now, `register_trainable` is handled implicitly.