When a task failed, it'll wait for the death info and then fail. The waiting is 1s and the checking is every 1s. This is good for usability, but it causes issues for some cases because it'll delay the task return at most 2s and at least 1s.
This PR introduce an early cut where when the timeout is set to be 0, it'll just return immediately. The semantics doesn't change and for most users they are still going to get the message.
## Why are these changes needed?
When GCS restarts, sometimes, raylet needs a while to reconnect to the GCS, for example, in k8s env, it needs a while to move GSC to the service. This PR try to fix this by allowing a longer timeout for the first ping when GCS restarts.
Once GCS get the first ping, it'll just use the regular timeout instead.
Currently, trainables will try to sync up/down temporary checkpoints from cloud storage, leading to errors. These erros come up e.g. with PBT, which heavily uses saving/restoring from objects.
Instead, we should not sync these temporary checkpoints up at all, and we should generally not sync down if a local checkpoint directory exists, which will prevent us also from trying to sync down non-existent temporary checkpoint directories.
See #26714
Signed-off-by: Kai Fricke <kai@anyscale.com>
Calling e.g. `os.path.exists(checkpoint)` currently raises an TypeError, but we should make it more explicit and guide users towards the correct API.
Signed-off-by: Kai Fricke <kai@anyscale.com>
This PR just applies the changes from the following PRs:
[Datasets] Automatically cast tensor columns when building Pandas blocks. #26684
reverted by Revert "[Datasets] Automatically cast tensor columns when building Pandas blocks." #26921
[AIR - Datasets] Fix TensorDtype construction from string and fix example. #26904
This fixes the test failures introduced in the originally reverted PRs.
Update cluster_activities endpoint to use pydantic so we have better data validation.
Make timestamp a required field.
Add pydantic to ray[default] requirements
Why are these changes needed?
Resubmitting #26869.
This PR was reverted due to failing tests; however, those failures were actually due to a dependency: #26950
Signed-off-by: Matthew Deng matt@anyscale.com
Note: This aims to mitigate the errors of the failing tests, but a follow-up is needed for a long-term solution.
Why are these changes needed?
A bunch of CI tests started failing on 7/23.
Quick sanity check shows only werkzeug was upgraded from 2.1.2 to 2.2.0. The new version was released on 7.23.
Verified that running pip install -U Werkzeug==2.1.2 fixes (at least) test_dataset_formats.
This PR does 3 things.
1. Warn if callsite is disabled when `ray list objects` and `ray summary objects`
2. Decode owner_id for ray list actors
3. Support raise_on_missing_output
Signed-off-by: Yi Cheng <chengyidna@gmail.com>
## Why are these changes needed?
Right now, only cpp layer in ray is connecting to redis which means we don't need pip redis to connect to a redis db.
The blocking part is that we are doing some sharding in redis right now. But this feature is not actually used and the shard is always 1. So to make things simple, this feature is just disabled.
Test is added to make sure we can start ray with a redis db without pip redis.
Splitting up #26884: This PR includes changes to use Tuner() instead of tune.run() for all examples included in python/ray/tune/examples
Signed-off-by: xwjiang2010 <xwjiang2010@gmail.com>
Signed-off-by: Kai Fricke <kai@anyscale.com>
Co-authored-by: xwjiang2010 <xwjiang2010@gmail.com>
Co-authored-by: Richard Liaw <rliaw@berkeley.edu>
Signed-off-by: Kai Fricke coding@kaifricke.com
Why are these changes needed?
Splitting up #26884: This PR includes changes to use Tuner() instead of tune.run() for most docs files (rst and py), and a change to move reuse_actors to the TuneConfig
Signed-off-by: Yi Cheng <chengyidna@gmail.com>
## Why are these changes needed?
When actor died, it'll send notification to core workers. Right now, sometimes, core worker will queue the task waiting for actor death info and pop it up for better usability. But in async cases, this is going to cause issues.
The callback might submit tasks which require holding the lock. But it's already being held. This is going to cause a deadlock.
This PR fixed this by moving the failure part out of the lock.
Ray 2.0 API deprecation for:
ray.remote(): placement_group
ray.remote(): placement_group_bundle_index
ray.remote(): placement_group_capture_child_tasks
ray.get_dashboard_url()
ray.get_resource_ids()
ray.disconnect()
ray.connect()
ray.util.ActorGroup
ray.util.ActorPool
Add get_xx_id() to return hex (rather than object), and then deprecate the xx_id() (which returns Cython object): the xx here can be node, task etc.
ray start: --plasma-store-socket-name
ray start: --raylet-socket-name
Why are these changes needed?
Introduce a stable version of split with hints with a stable equalizing algorithm:
use the greedy algorithm to generate the initial unbalanced splits.
for each splits, first shave them so the number for rows are below the target_size
based on how many rows needed for each split, do a one time split_at_index to the left over blocks.
merge the shaved splits with the leftover splits.
The guarantee of this algorithm is we at most need to split O(split) number of blocks.
Why are these changes needed?
Consumers (e.g. Train) may expect generated batches to be of the same size. Prior to this change, the default behavior would be for each batch to be one block, which may be of different sizes.
Changes
Set default batch_size to 256. This was chosen to be a sensible default for training workloads, which is intentionally different from the existing default batch_size value for Dataset.map_batches.
Update docs for Dataset.iter_batches, Dataset.map_batches, and DatasetPipeline.iter_batches to be consistent.
Updated tests and examples to explicitly pass in batch_size=None as these tests were intentionally testing block iteration, and there are other tests that test explicit batch sizes.
ray.init() will currently start a new Ray instance even if one is already existing, which is very confusing if you are a new user trying to go from local development to a cluster. This PR changes it so that, when no address is specified, we first try to find an existing Ray cluster that was created through `ray start`. If none is found, we will start a new one.
This makes two changes to the ray.init() resolution order:
1. When `ray start` is called, the started cluster address was already written to a file called `/tmp/ray/ray_current_cluster`. For ray.init() and ray.init(address="auto"), we will first check this local file for an existing cluster address. The file is deleted on `ray stop`. If the file is empty, autodetect any running cluster (legacy behavior) if address="auto", or we will start a new local Ray instance if address=None.
2. When ray.init(address="local") is called, we will create a new local Ray instance, even if one is already existing. This behavior seems to be necessary mainly for `ray.client` use cases.
This also surfaces the logs about which Ray instance we are connecting to. Previously these were hidden because we didn't set up the log until after connecting to Ray. So now Ray will log one of the following messages during ray.init:
```
(Connecting to existing Ray cluster at address: <IP>...)
...connection...
(Started a local Ray cluster.| Connected to Ray Cluster.)( View the dashboard at <URL>)
```
Note that this changes the dashboard URL to be printed with `ray.init()` instead of when the dashboard is first started.
Co-authored-by: Eric Liang <ekhliang@gmail.com>
