* Init commit for async plasma client
* Create an eventloop model for ray/plasma
* Implement a poll-like selector base on `ray.wait`. Huge improvements.
* Allow choosing workers & selectors
* remove original design
* initial implementation of epoll-like selector for plasma
* Add a param for `worker` used in `PlasmaSelectorEventLoop`
* Allow accepting a `Future` which returns object_id
* Do not need `io.py` anymore
* Create a basic testing model
* fix: `ray.wait` returns tuple of lists
* fix a few bugs
* improving performance & bug fixing
* add test
* several improvements & fixing
* fix relative import
* [async] change code format, remove old files
* [async] Create context wrapper for the eventloop
* [async] fix: context should return a value
* [async] Implement futures grouping
* [async] Fix bugs & replace old functions
* [async] Fix bugs found in tests
* [async] Implement `PlasmaEpoll`
* [async] Make test faster, add tests for epoll
* [async] Fix code format
* [async] Add comments for main code.
* [async] Fix import path.
* [async] Fix test.
* [async] Compatibility.
* [async] less verbose to not annoy the CI.
* [async] Add test for new API
* [async] Allow showing debug info in some of the test.
* [async] Fix test.
* [async] Proper shutdown.
* [async] Lint~
* [async] Move files to experimental and create API
* [async] Use async/await syntax
* [async] Fix names & styles
* [async] comments
* [async] bug fixing & use pytest
* [async] bug fixing & change tests
* [async] use logger
* [async] add tests
* [async] lint
* [async] type checking
* [async] add more tests
* [async] fix bugs on waiting a future while timeout. Add more docs.
* [async] Formal docs.
* [async] Add typing info since these codes are compatible with py3.5+.
* [async] Documents.
* [async] Lint.
* [async] Fix deprecated call.
* [async] Fix deprecated call.
* [async] Implement a more reasonable way for dealing with pending inputs.
* [async] Fix docs
* [async] Lint
* [async] Fix bug: Type for time
* [async] Set our eventloop as the default eventloop so that we can get it through `asyncio.get_event_loop()`.
* [async] Update test & docs.
* [async] Lint.
* [async] Temporarily print more debug info.
* [async] Use `Poll` as a default option.
* [async] Limit resources.
* new async implementation for Ray
* implement linked list
* bug fix
* update
* support seamless async operations
* update
* update API
* fix tests
* lint
* bug fix
* refactor names
* improve doc
* properly shutdown async_api
* doc
* Change the table on the index page.
* Adjust table size.
* Only keeps `as_future`.
* change how we init connection
* init connection in `ray.worker.connect`
* doc
* fix
* Move initialization code into the module.
* Fix docs & code
* Update pyarrow version.
* lint
* Restore index.rst
* Add known issues.
* Apply suggestions from code review
Co-Authored-By: suquark <suquark@gmail.com>
* rename
* Update async_api.rst
* Update async_api.py
* Update async_api.rst
* Update async_api.py
* Update worker.py
* Update async_api.rst
* fix tests
* lint
* lint
* replace the magic number
<!--
Thank you for your contribution!
Please review https://github.com/ray-project/ray/blob/master/CONTRIBUTING.rst before opening a pull request.
-->
## What do these changes do?
JSON Logger now uses cloudpickle to dump the configs as welll, which pkls the functions needed for multi-agent replay.
## Related issue number
<!-- Are there any issues opened that will be resolved by merging this change? -->
auto wrap multi-agent dict and tuple spaces by keeping a policy -> preprocessor in the sampler
add some Q-learning debug stats
report min, max of custom metrics
better errors
## What do these changes do?
This passes in the right obs space to the lstm model wrapper, so that it doesn't attempt to un-flatten the already processed dict observation.
## Related issue number
Closes https://github.com/ray-project/ray/issues/3367
This PR introduces single-node fault tolerance for Tune.
## Previous behavior:
- Actors will be restarted without checking if resources are available. This can lead to problems if we lose resources.
## New behavior:
- RUNNING trials will be resumed on another node on a best effort basis (meaning they will run if resources available).
- If the cluster is saturated, RUNNING trials on that failed node will become PENDING and queued.
- During recovery, TrialSchedulers and SearchAlgorithms should receive notification of this (via `trial_runner.stop_trial`) so that they don’t wait/block for a trial that isn’t running.
Remaining questions:
- Should `last_result` be consistent during restore?
Yes; but not for earlier trials (trials that are yet to be checkpointed).
- Waiting for some PRs to merge first (#3239)
Closes#2851.
IMPALA support for multiagent was broken since IMPALA has a requirement that batch sizes be of a certain length. However multi-agent envs can create variable-length batches.
Fix this by adding zero-padding as needed (similar to the RNN case).
