One GC test has unnecessary sleeps which are quite expensive due to the parametrization (2 x 2 x 2 = 8 iterations). They are unnecessary because they check that garbage collection of runtime env URIs doesn't occur after a certain time, but garbage collection isn't time-based. This PR removes the sleeps.
This PR is just to fix CI; a followup PR will make the test more effective by attempting to trigger GC in a more targeted way (by starting multiple tasks with different runtime_env resources. GC is only triggered upon *creation* of a new resource that causes the cache size to be exceeded.)
It's still not clear what exactly caused the test suite to start taking longer recently, but it might be due to some change elsewhere in Ray, since there were no runtime_env related commits in that time period.
This PR ccf4116 makes cluster_utils.add_node take about 1 more second because of the raylet start path refactoring.
It seems like as a result, test_placement_group_3 has an occasional timeout. I extended the test to be large. Let's see if this fixes the issue.
Why are these changes needed?
The Ray-level OOM killer preemptively kills a worker process when the node is under memory pressure. This PR leverages the memory monitor from #27017 and supersedes #26962 to kill worker processes when the system is running low on memory. The node manager implements the callback in the memory monitor and kills the worker process with the newest task. It evicts only one worker at a time and enforces that by tracking the last evicted worker. If the eviction is still in progress it will not evict another worker even if the memory usage is above the threshold.
This PR is a no-op since the monitor is disabled by default.
In the previously merged pr(https://github.com/ray-project/ray/pull/22726/commits), java serve's support for python deployment was not implemented. This PR is used to implement this feature.
Co-authored-by: nanqi.yxf <nanqi.yxf@antgroup.com>
This change makes us report placement resources for actor creation tasks. Essentially, the resource model here is that a placement resource/actor creation task is a task that runs very quickly.
Closes#26806
Co-authored-by: Alex <alex@anyscale.com>
This PR fixes a reference counting bug for borrowed objects sent to an actor creation task that is then cancelled.
Before this PR, when actor creation is cancelled before the creation task has been scheduled, the GCS-based actor manager would would destroy the actor without replying to the task submission RPC from the actor creating worker, resulting in the reference counts on that worker to never get cleaned up. This caused us to leak borrowed objects when such cancellation-before-scheduling happened for actors.
This PR fixes this by ensuring that the task submission RPC receives a reply indicating that the actor creation task has been cancelled, at which point the submitting worker will run through the same reference counting cleanup as is done for normal task cancellation.
Updates jobs api
Updates snapshot api
Updates state api
Increases jobs api version to 2
Signed-off-by: Alan Guo aguo@anyscale.com
Why are these changes needed?
follow-up for #25902 (comment)
Signed-off-by: Nikita Vemuri nikitavemuri@gmail.com
Why are these changes needed?
Support printing a Ray dashboard URL that the user specifies through environment variable. This can be helpful if the Ray dashboard is hosted externally.
When reopening a file due to an inode change, we weren't seeking back to the right location. Now we are (with a unit test).
Closes (but not really until it's cherry-picked) #27507
Co-authored-by: Alex <alex@anyscale.com>
Co-authored-by: Eric Liang <ekhliang@gmail.com>
This reverts commit cf7305a, and unreverts #25896.
This was reverted due to a failing Windows test: #26287
We can merge once the failing Windows test (and all other relevant tests) pass.
Datasets currently eagerly kicks off all read tasks when truncating a dataset immediately after a read via ray.data.read_*().limit(); this results in a lot of wasted computation and unnecessary object store bloat, especially when trying to poke at a very small subset of the data.
This PR avoids these unnecessary reads by truncating the blocklist to the minimum number of blocks needed to meet the row limit before doing the actual block splitting, thereby avoiding materialization of unnecessary read tasks in the common splitting path.
Integration between Ray Serve and Gradio. Users of Gradio can wrap their Gradio app in a Serve deployment by using `GradioIngress`, and scale it up through more replicas or more CPU/GPU resources.
Previous PR that adds in lightweight config updates: https://github.com/ray-project/ray/pull/27000. It only tracks the config options for `deployments` (bumps version if certain deployment options are changed, but otherwise keeps versions the same). However we should bump the versions of all deployments if `import_path` or `runtime_env` is changed.
Failed node launch can lead to an extra unexpected error in the node launcher due to the definition of a mock prometheus metric method.
This failure leads to a permanently hanging autoscaler with "launching nodes" never cleared out and the autoscaler unable to proceed to launch nodes.
This PR fixes the method signature leading to the unexpected failure.
Signed-off-by: Alex Wu <alex@anyscale.io>
This is a minor QoL improvement to bump the hardcoded limit for number of aws keys per account. The limit is arbitrary and has been bumped before. AFAICT the fundamental aws limit is a 5000 key per region limit which we are not close to.
This PR adds a user guide to AIR for using Ray Train. It provides a high level overview of the trainers and removes redundant sections.
The main file to review is here: doc/source/ray-air/trainer.rst.
Signed-off-by: xwjiang2010 <xwjiang2010@gmail.com>
Signed-off-by: Richard Liaw <rliaw@berkeley.edu>
Signed-off-by: Kai Fricke <kai@anyscale.com>
Co-authored-by: Richard Liaw <rliaw@berkeley.edu>
Co-authored-by: Kai Fricke <kai@anyscale.com>
1. If a user reads a folder with grayscale and color images, ImageFolderDatasource errors.
2. There's no way to retain image shapes.
Co-authored-by: Clark Zinzow <clarkzinzow@gmail.com>
Converting a Pandas DataFrame column to an ndarray (e.g. via df[col].values) can often result in a full copy of the column in order to construct the ndarray due to Pandas' 2D block management. This PR ports tensor extension type checking to checking the dtype, which is always an O(1) check.
Signed-off-by: Clark Zinzow <clarkzinzow@gmail.com>
We decided to allow escaping the parent pg via `PlacementGroupSchedulingStrategy(placement_group=None)` instead of using "DEFAULT". Our doc is updated with that but in the code it's still not allowed.
