Currently, when the GCS RPC failed with gRPC unavailable error because the GCS is dead, it will retry forever.
b3a9d4d87d/src/ray/rpc/gcs_server/gcs_rpc_client.h (L57)
And it takes about 10 minutes to detect the GCS server failure, meaning if GCS is dead, users will notice in 10 minutes.
This can easily cause confusion that the cluster is hanging (since users are not that patient). Also, since GCS is not fault tolerant in OSS now, 10 minutes are too long timeout to detect GCS death.
This PR changes the value to 60 seconds, which I believe is much more reasonable (since this is the same value as our blocking RPC call timeout).
This fixes the bug where empty line is printed to the driver when multi threads are used.
e.g.,
```
2021-12-12 23:20:06,876 INFO worker.py:853 -- Connecting to existing Ray cluster at address: 127.0.0.1:6379
(TESTER pid=12344)
(TESTER pid=12348)
```
## How does the current log work?
After actor initialization method is done, it prints ({repr(actor)}) to the stdout and stderr, which will be read by log monitor. Log monitor now can parse the actor repr and starts publishing log messages to the driver with the repr name.
## Problem
If actor init method contains a new background thread, when we call print({repr(actor)}), it seems flush() is not happening atomically. Based on my observation, it seems to flush the repr(actor) first, and then we flush the end="\n" (the default end parameter of print function) after.
Since log monitor never closes the file descriptor, it is possible it reads the log line before the end separator "\n" is flushed. That says, this sort of scenario can happen.
Expected
- `"repr(actor)\n"` is written to the log file. (which includes the default print end separator `"\n"`).
- Log monitor reads `repr(actor)\n` and publishes it.
Edge case
- `"repr(actor)"` is written to the log
- Log monitor publishes `repr(actor)`.
- `"\n"` is written to the log (end separator).
- Log monitor publishes `"\n"`.
Note that this is only happening when we print the actor repr "after" actor initialization. I think since new thread is running in the background, GIL comes in, and it creates the gap between `flush(repr(actor))` and `flush("\n")`, which causes the issue.
I verified the is fixed when I add the separator ("\n") directly to the string, and make the end separator as an empty string.
Alternative fix is to file lock the log file whenever log monitor reads / core worker writes, but it seems to be too heavy solution compared to its benefit.
Currently, there's no way to specify grpc port for dashboard agents (only http ports are allowed to be configured). This PR allows users to do that.
This also add more features;
- Format the port conflict error message better. I found the default worker ports are from 10002-19999, which spams the terminal. It is now formatted as 10002-19999.
- Add dashboard http port to the port conflict list.
`test_multi_node_3` failed because we kill the raylet before the cluster is up which leads the raylet to become a zombie process. This fix wait until the cluster up and kill it.
Resubmit the PR https://github.com/ray-project/ray/pull/19936
I've figure out that the test case `//rllib:tests/test_gpus::test_gpus_in_local_mode` failed due to deadlock in local mode.
In local mode, if the user code submits another task during the executing of current task, the `CoreWorker::actor_task_mutex_` may cause deadlock.
The solution is quite simple, release the lock before executing task in local mode.
In the commit 7c2f61c76c:
1. Release the lock in local mode to fix the bug. @scv119
2. `test_local_mode_deadlock` added to cover the case. @rkooo567
3. Left a trivial change in `rllib/tests/test_gpus.py` to make the `RAY_CI_RLLIB_DIRECTLY_AFFECTED ` to take effect.
This is the first PR of a series of Ray Dataset and Pandas integration improvements.
This PR refactors `ArrowRow`, `ArrowBlockBuilder`, `ArrowBlockAccessor` by extracting base classes `TableRow`, `TableBlockBuilder`, `TableBlockAccessor`, which can then be inherited for pandas DataFrame support in the next PR.
In our code, we wish to have control over when GC runs. We do this by `gc.disable()` and then manually triggering GC at moments that work for us. This does not work if Ray re-enables GC.
Co-authored-by: hauntsaninja <>
PR #19014 introduced the idea of a StartupToken to uniquely identify a worker via a counter. This PR:
- returns the Process and the StartupToken from StartWorkerProcess (previously only Process was returned)
- Change the starting_workers_to_tasks map to index via the StartupToken, which seems to fix the windows failures.
- Unskip the windows tests in test_basic_2.py
It seems once a fix to PR #18167 goes in, the starting_workers_to_tasks map will be removed, which should remove the need for the changes to StartWorkerProcess made in this PR.
Currently if local lease request fails due to raylet death, direct_task_transport.cc will retry forever for driver.
With this PR, we treat grpc unavailable as non-retryable error (the assumption is that local grpc is always reliable and grpc unavailable error indicates that server is dead) and will just fail the task.
Note: this PR doesn't try to address a bigger problem: don't crash driver when local raylet dies. We have multiple places in the code that assumes the local raylet never fail and have CHECK_STATUS_OK for that. All these places need to be changed so we can properly propagate failures to the user.
A worker can crash right after putting its return values into the object store. Then, the owner will receive the worker crashed error, but the return objects will still be in the remote object store. Later, if the task is retried, the worker will crash on [this line](https://github.com/ray-project/ray/blob/master/src/ray/core_worker/transport/direct_actor_transport.cc#L105) because the object already exists.
Another way this can happen is if a task has multiple return values, and one of those return values is transferred to another node. If the task is later re-executed on that node, the task will fail because of the same error.
This PR fixes the crash so that:
1. If an object already exists, we try to pin that copy. Ideally, we should destroy the old copy and create the new one to make sure that metadata like the owner address is in sync, but this is pretty complicated to do right now.
2. If the pinning fails, we store an OBJECT_LOST error to throw to the application.
3. On the raylet, we check whether we already have the object pinned, and only subscribe to the owner's eviction message if the object is not pinned.
4. Also fixes bugs in the analogous case for `ray.put` (previously this would hang, now the application will receive an error if a `ray.put` object already exists).
Dashboard contains resource reporter and actor subscribers. Dashboard agent has resource report publisher. So GCS pubsub needs to support these channel types.
Also refactor GCS AIO subscribers to have each subscriber per channel. This matches the API of GCS sync subscribers, and make subscribing with multiple channels easier.
Bazel has a rule for enforcing the version so we can just reuse that.
This redundant bazel version check logic in setup.py is also causing issue when building conda package, because conda has its own version of bazel and it doesn't support `--version`.
This PR adds RecordMetrics and DebugString to all raylet components.
Some of methods are probably empty now. They are going to be supported in the next PR
Before this PR, then raylet notices there is something wrong with the worker starting, it'll start a new worker but not kill the old one. If the old one is hanging, it'll lead to resource waste.
This PR killed the failed worker if it's still alive and also print useful logs
This adds an initial Dataset.stats() framework for debugging dataset performance. At a high level, execution stats for tasks (e.g., CPU time) are attached to block metadata objects. Datasets have stats objects that hold references to these stats and parent dataset stats (this avoids stats holding references to parent datasets, allowing them to be gc'ed). Similarly, DatasetPipelines hold stats from recently computed datasets.
Currently only basic ops like map / map_batches are instrumented. TODO placeholders are left for future PRs.
This PR centralizes all existing metrics to `metric_defs.h`.
Previously, each file relies on implicit import of metric_def.h within the stats module. After this PR we only precisely import `metric_defs.h` for each file.
This PR adds an ability to set a random seed/numpy random generator in BasicVariantGenerator, allowing for reproducibility across separate runs. All the changes are fully backwards compatible.
Co-authored-by: Kai Fricke <krfricke@users.noreply.github.com>
This PR adds per task/actor scheduling strategy and currently the only strategy are PlacementGroupSchedulingStrategy and DefaultSchedulingStrategy.
Going forward, people should use `scheduling_strategy=PlacementGroupSchedulingStrategy` to define placement group for actor/task. The old way will be deprecated.
There is a race condition between `DestoryActor` (due to handle out of scope) and `CreateActor`. If `DestoryActor` happens first, then `CreateActor` will fail complaining about not being able to find the registered actor.
As described in #18884, reconfiguration will mutate state mid-query. I try to solve this problem by adding read/write lock to each replica.
Co-authored-by: yuzihao.2001 <yuzihao.2001@bytedance.com>
We can just pickle task options instead of json so that we don't need to write custom `to_dict` and `from_dict` methods for complex python option objects (e.g. PlacementGroup).
Partly addresses #20774 by registering node launcher failures in driver logs, via the event summarizer.
This way, users can tell that the launch failed from the driver logs.
Also pushes the node creation exception to driver logs, but only once per 60 minutes.
Right now in ray, a lot of edge cases related to grpc are not tested. This PR is just a simple try to give the developer some way to delay grpc request. It could be used with manual testing and also e2e test since it's supporting delay for specific grpc method.
To use this feature, just simple set os env `RAY_TESTING_ASIO_DELAY_US="method1=10:20,method2=20:30,*=200:200"`
This means, for `method1` it'll delay 10-20us, for method2 it'll delay 20-30us. For all the rest, it'll delay 200us.
Arrow byte-packs boolean arrays, with 8 entries per byte, and both Arrow and NumPy utilize bit-based offsets for indexing into data buffers. This PR adds support for properly indexing into boolean tensor columns by using bit-based offsets for such columns.
We shouldn't ray.get() all the blocks immediately during the to_pandas call, it's better to do it one by one. That's a little slower but to_pandas() isn't expected to be fast anyways.
This PR addresses two issues an issue with the Session.get_next docstring:
It's unclear whether the tense should be imperative or non-imperative. The Ray documentation states that we use Google style (which is non-imperative), but we are formatting using PEP8 (which is imperative). Moreover, we use both imperative and non-imperative summaries across the Ray Train code. I've stuck with a non-imperative summary for consistency with the rest of the Session class.
The docstring doesn't describe the conditions under which the function returns None.
It's useful when measuring autoscaler performance to know how long the autoscaler takes for update iterations.
This PR adds a prometheus metric for that.
The bucket ticks for the histogram are arranged in powers of ten:
[.001, .01, .1, 1, 10, 100, 1000]. Depending on the situation, we've seen the update time range from .1 second to a few minutes.
