On the ServeHead level, it is talking to serve api and controller to do deployment and clean up now. With this pr, it hides the deployment clean up logic into server.run() for code cleanness and easy to refactor in the future.
This PR modifies the KubeRay e2e autoscaling test so that one of its scaling commands is sent via the Ray Job Submission API.
This validates that the Ray Job Submission API works with KubeRay and, in particular, that the Ray Dashboard is correctly exposed.
Updating W&B Ray Tune Integration with new standards. Adding support to wandb service, the soon to be default way for multiprocessing + wandb run logging.
Co-authored-by: Kai Fricke <krfricke@users.noreply.github.com>
This PR allows for Ray to disable metrics collection. It was possible with RAY_enable_metrics_collection, but it didn't fully disable collection because there was a metrics collection happening from agent that wasn't properly disabled. This PR also adds tests.
Users get error messages from client/server on actor failures, even if they already try-except'd the error. For example:
```
import ray
ray.init("ray://localhost:10001")
try:
ray.get_actor("doesnotexist")
except ValueError:
pass
```
Will still generate the log `Caught schedule exception` and `Exception from actor creation is ignored in destructor. To receive this exception in application code, call a method on the actor reference before its destructor is run.`. Reduce the level of these logs to debug by default.
Fixes checkpoints not being recorded in Tune's checkpoint manager if the first checkpoint has None value. This also deflakes `test_checkpoint_manager.py::CheckpointManagerTest`.
**TL;DR:** Don't clear for eager, clear all but non-lazy input blocks if lazy, clear everything if pipelining.
This PR provides more efficient and intuitive block clearing semantics for eager mode, lazy mode, and pipelining, while still supporting multiple operations applied to the same base dataset, i.e. fan-out. For example, two different map operations are applied to the same base `ds` in this example:
```python
ds = ray.data.range(10).map(lambda x: x+1)
ds1 = ds.map(lambda x: 2*x)
ds2 = ds.map(lambda x: 3*x)
```
If naively clear the blocks when executing the map to produce `ds1`, the map producing `ds2` will fail.
### Desired Semantics
- **Eager mode** - don’t clear input blocks, thereby supporting fan-out from cached data at any point in the stage chain without triggering unexpected recomputation.
- **Lazy mode** - if lazy datasource, clear the input blocks for every stage, relying on recomputing via stage lineage if fan-out occurs; if non-lazy datasource, do not clear source blocks for execution plan when executing first stage, but do clear input blocks for every subsequent stage.
- **Pipelines** - Same as lazy mode, although the only fan-out that can occur is from the pipeline source blocks when repeating a dataset/pipeline, so unintended intermediate recomputation will never happen.
#17581 introduced a warning about excess queuing for actors. Unfortunately since Ray 1.10.0, the metric used became wrong for async actors, resulting in bogus warnings when they are called more than 5000 times, even though there are not 5000 pending tasks.
The difference between 1.9.2 and 1.10.0 is that async actors tasks skip the queue in CoreWorkerClient::PushActorTask. However CoreWorkerClient::ClientProcessedUpToSeqno uses max_finished_seq_no_ which is never updated when the queue is skipped.
I think that a better metric for the amount of tasks that are pending submissions is the size of the internal queue CoreWorkerDirectActorTaskSubmitter::inflight_task_callbacks.
For debugging client environments, it is helpful to print the installed pip packages.
Additionally, a fix for the environment of the ml_user_tune_rllib_connect_test is added. Additionally, anyscale import errors are reported verbosely to help debug missing packages.
The prefetch_blocks implementation doesn't work as expected. Due to ray.wait() doesn't given us fine grained control, today we block waiting any of the block returns. As I read the code, it may or may not actually fetching all the blocks.
A better way to ensure prefetching not blocking is to use ray remote function call, which is not blocking and ensures the blocks are fetched eventually.
Adds a fast file metadata provider that trades comprehensive file metadata collection for speed of metadata collection, and which also disabled directory path expansion which can be very slow on some cloud storage service providers. This PR also refactors the Parquet datasource to be able to take advantage of both these changes and the content-type agnostic partitioning support from #23624.
This is the second PR of a series originally proposed in #23179.
Adds a Categorizer preprocessor to automatically set the Categorical dtype on a dataset. This is useful for eg. LightGBM, which has build-in support for features with that dtype.
Depends on #24144.
After https://github.com/ray-project/ray/pull/24066, some release tests are running into:
```
ModuleNotFoundError: No module named 'ray.train.impl'
```
This PR simply adds a `__init__.py` file to resolve this.
We also add a 5 wecond delay for client runners in release test to give clusters a bit of slack to come up (and avoid ray client connection errors)
#24311 added the `test_update_num_replicas_anonymous_namespace` unit test to check for replica leaks in anonymous namespaces. This change adds this test to the master branch.
The postprocess checkpoint method was introduced to be able to add data to function runner checkpoint directories before they are uploaded to external (cloud) storage. Instead, we should just use the existing separation of `save_checkpoint()` and `save()`.
The simple shuffle currently implemented in Datasets does not reliably scale past 1000+ partitions due to metadata and I/O overhead.
This PR adds an experimental shuffle implementation for a "push-based shuffle", as described in this paper draft. This algorithm should see better performance at larger data scales. The algorithm works by merging intermediate map outputs at the reducer side while other map tasks are executing. Then, a final reduce task merges these merged outputs.
Currently, the PR exposes this option through the DatasetContext. It can also be set through a hidden OS environment variable (RAY_DATASET_PUSH_BASED_SHUFFLE). Once we have more comprehensive benchmarks, we can better document this option and allow the algorithm to be chosen at run time.
Redo for #23758 to fix CI.
For tasks with node affinity scheduling strategy, the resource demands shouldn't create new nodes. This PR achieves this by not reporting demand to autoscaler. In the future, we will explore sending scheduling strategy information to autoscaler.
To add basic plotting feature for Ray DAGs.
`ray.experimental.dag.plot(dag: DAGNode, to_file=None)`
### Behavior
1. dump the dag plot (Dot) to file.
2. also render the image whenever possible. E.g. if running in Jupyter notebook, the image will not only be saved, but also rendered in the notebook.
3. when to_file is not set (i.e. None), it will be saved to a tempfile for rendering purpose only. This is common when users plot DAGs in notebook env to explore the DAG structure without wanting to save it to a file.
Refactors _get_unique_value_indices (used in Encoder preprocessors) for much improved performance with multiple columns. Also uses the same, more robust intermediary dataset format in _get_most_frequent_values (Imputers).
The existing unit tests pass, and no functionality has been changed.
RayDP has updated their code and tests can be re-enabled now.
In addition, we want to support ray client in raydp dataset operation. Right now, if users want to do dataset.to_spark(spark) while using ray client, it will immediately fail because the local ray worker is not connected. By wrapping it in a function decorated with @client_mode_wrap, It works well no matter ray client is used or not.
Lineage-based serialization isn't supported for fan-in operations such as unions and zips. This PR adds documentation indicating as much, and ensures that a good error message is raised.
When using the actor compute model for batch mapping (e.g. in batch inference), map tasks are often blocked waiting for their dependencies to be fetched since we submit one actor task at a time. This commit changes the default behavior of the actor compute model to have up to two actor tasks in flight for each actor in order to better pipeline task dependency fetching with the actual compute.
This "max tasks in flight per actor worker" is also made configurable, in case a particular use case warrants more aggressive pipelining (e.g. big blocks and/or fast maps) or more conservative pipelining (e.g. small data or slow maps).
Refactors Dataset splitting to make it less hacky and address the TODO. Also makes Dataset ingest in general configurable for Ray Train. This is an internal only change for now, but will set the stage for the proposed ingest API
Customizable ingest for GBDT Trainers is out of scope for this PR.
