* Implement Actor checkpointing
* docs
* fix
* fix
* fix
* move restore-from-checkpoint to HandleActorStateTransition
* Revert "move restore-from-checkpoint to HandleActorStateTransition"
This reverts commit 9aa4447c1e3e321f42a1d895d72f17098b72de12.
* resubmit waiting tasks when actor frontier restored
* add doc about num_actor_checkpoints_to_keep=1
* add num_actor_checkpoints_to_keep to Cython
* add checkpoint_expired api
* check if actor class is abstract
* change checkpoint_ids to long string
* implement java
* Refactor to delay actor creation publish until checkpoint is resumed
* debug, lint
* Erase from checkpoints to restore if task fails
* fix lint
* update comments
* avoid duplicated actor notification log
* fix unintended change
* add actor_id to checkpoint_expired
* small java updates
* make checkpoint info per actor
* lint
* Remove logging
* Remove old actor checkpointing Python code, move new checkpointing code to FunctionActionManager
* Replace old actor checkpointing tests
* Fix test and lint
* address comments
* consolidate kill_actor
* Remove __ray_checkpoint__
* fix non-ascii char
* Loosen test checks
* fix java
* fix sphinx-build
* Factor out starting Ray processes.
* Detect flags through environment variables.
* Return ProcessInfo from start_ray_process.
* Print valgrind errors at exit.
* Test valgrind in travis.
* Some valgrind fixes.
* Undo raylet monitor change.
* Only test plasma store in valgrind.
* add marvil policy graph
* fix typo
* add offline optimizer and enable running marwil
* fix loss function
* add maintaining the moving average of advantage norm
* use sync replay optimizer for unifying
* remove offline optimizer and use sync replay optimizer
* format by yapf
* add imitation learning objective
* fix according to eric's review
* format by yapf
* revise
* add test data
* marwil
Rename `xray_test.py` to `mini_test.py` and use that in the documentation. Right now we suggest that people run `runtest.py`, but that often doesn't succeed and takes too long.
* Limit Redis max memory to 10GB/shard by default.
* Update stress tests.
* Reorganize
* Update
* Add minimum cap size for object store and redis.
* Small test update.
This PR introduces cluster-level fault tolerance for Tune by checkpointing global state. This occurs with relatively high frequency and allows users to easily resume experiments when the cluster crashes.
Note that this PR may affect automated workflows due to auto-prompting, but this is resolvable.
