## Why are these changes needed?
This PR aims to port concurrency groups functionality with asyncio for Python.
### API
```python
@ray.remote(concurrency_groups={"io": 2, "compute": 4})
class AsyncActor:
def __init__(self):
pass
@ray.method(concurrency_group="io")
async def f1(self):
pass
@ray.method(concurrency_group="io")
def f2(self):
pass
@ray.method(concurrency_group="compute")
def f3(self):
pass
@ray.method(concurrency_group="compute")
def f4(self):
pass
def f5(self):
pass
```
The annotation above the actor class `AsyncActor` defines this actor will have 2 concurrency groups and defines their max concurrencies, and it has a default concurrency group. Every concurrency group has an async eventloop and a pythread to execute the methods which is defined on them.
Method `f1` will be invoked in the `io` concurrency group. `f2` in `io`, `f3` in `compute` and etc.
TO BE NOTICED, `f5` and `__init__` will be invoked in the default concurrency.
The following method `f2` will be invoked in the concurrency group `compute` since the dynamic specifying has a higher priority.
```python
a.f2.options(concurrency_group="compute").remote()
```
### Implementation
The straightforward implementation details are:
- Before we only have 1 eventloop binding 1 pythread for an asyncio actor. Now we create 1 eventloop binding 1 pythread for every concurrency group of the asyncio actor.
- Before we have 1 fiber state for every caller in the asyncio actor. Now we create a FiberStateManager for every caller in the asyncio actor. And the FiberStateManager manages the fiber states for concurrency groups.
## Related issue number
#16047
## Why are these changes needed?
It looks like the metrics set on server side are wrong. The time the query is constructed sometimes is not the time we get the query. This PR fixed this.
## Related issue number
Why are these changes needed?
Right now the failure signal handler registered in Python worker is skipped on crashes like segfault, because C++ core worker overrides the failure signal handler here and does not call the previously registered handler. This prevents Python stack trace from being printed on crashes. The fix is to make the C++ fault signal handler to call the previous signal handler registered in Python. For example with the script below which segfaults,
import ray
ray.init()
@ray.remote
def f():
import ctypes;
ctypes.string_at(0)
ray.get(f.remote())
Ray currently only prints the following stack trace:
(pid=26693) *** SIGSEGV received at time=1634418743 ***
(pid=26693) PC: @ 0x7fff203d9552 (unknown) _platform_strlen
(pid=26693) [2021-10-16 14:12:23,331 E 26693 12194577] logging.cc:313: *** SIGSEGV received at time=1634418743 ***
(pid=26693) [2021-10-16 14:12:23,331 E 26693 12194577] logging.cc:313: PC: @ 0x7fff203d9552 (unknown) _platform_strlen
With this change, Python stack trace will be printed in addition to the stack trace above:
(pid=26693) Fatal Python error: Segmentation fault
(pid=26693)
(pid=26693) Stack (most recent call first):
(pid=26693) File "/Users/mwtian/opt/anaconda3/envs/ray/lib/python3.7/ctypes/__init__.py", line 505 in string_at
(pid=26693) File "stack.py", line 7 in f
(pid=26693) File "/Users/mwtian/work/ray-project/ray/python/ray/worker.py", line 425 in main_loop
(pid=26693) File "/Users/mwtian/work/ray-project/ray/python/ray/workers/default_worker.py", line 212 in <module>
This should make debugging crashes in Python worker easier, for users and Ray devs.
Also, try to initialize symbolizer in GCS, Raylet and core worker. This is a no-op on MacOS and some Linux environments (e.g. Ray on Ubuntu 20.04 already produces symbolized stack traces), but should make Ray more likely to have symbolized stack traces on other platforms.
## Why are these changes needed?
There are some issues left from previous PRs.
- Put the gcs_actor_scheduler_mock_test back
- Add comment for named actor creation behavior
- Fix the comment for some flags.
## Related issue number
Why are these changes needed?
The theory around #19270 is there are two create actor requests sent to the same threaded actor due to retry logic. Specifically:
the first request comes and calls CoreWorkerDirectTaskReceiver::HandleTask, it's queued to be executed by thread pool;
then the second request comes and calls CoreWorkerDirectTaskReceiver::HandleTask again, before first request being executed and calls worker_context_.SetCurrentTask;
this fails the current dedupe logic and leads to SetMaxActorConcurrency be called twice, which fails the RAY_CHECK.
In this PR, we fix the dedupe logic by adding SetCurrentActorId and calling it in the task execution thread. this ensures the dedupe logic works for threaded actor.
we also noticed that the WorkerContext is actually not thread safe in threaded actors, thus make it thread safe in this PR as well.
Related issue number
Closes#19270
Checks
I've run scripts/format.sh to lint the changes in this PR.
I've included any doc changes needed for https://docs.ray.io/en/master/.
I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
Testing Strategy
Unit tests
Release tests
This PR is not tested :(
## Why are these changes needed?
Before this PR, there is a race condition where:
- job register starts
- driver start to launch actor
- gcs register actor ===> crash
- job register ends
Actor registration should be forced to be after driver registration. This PR enforces that.
## Related issue number
Closes#19172
* Update build rules and patches for darwin_arm64 platform.
Changes include:
Update nelhage/rules_boost package from current version (08/5/2020) to 5/27/2021 version.
Remove rules_boost-undefine-boost_fallthrough.patch, since BOOST_FALLTHROUGH seems to be defined now.
Minor changes to rules_boost-windows-linkopts.patch to use default condition to add -lpthread flag for all platforms.
Add darwin_arm64 config to BUILD files for lib civetweb pulled in via prometheu dependency.
* upgrade boost to 1.74.0 from 1.71.0 to match the udpated build file for windows.
* Fix ray_cpp_pkg
* Use boost/bind/bind.hpp
boost/bind.hpp and global namespace placeholders are deprecated.
* lint
* Use absl::bind_front when possible. Otherwise, NOLINT
* lint
* lint
* lint
* lint
* more lint
* final lint
* trigger build
In general, broadcasting changes to the replicas via the LongPollClient is hard to reason about (it circumvents our versioning semantics as there's no rolling update). Ideally we would only be using the LongPollClient to broadcast replica membership and nothing else.
* Convert worker pool to queue
* Start up to backlog size more workers
* fixes
* Prestart workers according to num available CPUs
* lint
* x
* Update src/ray/raylet/worker_pool.h
Co-authored-by: Eric Liang <ekhliang@gmail.com>
* Update src/ray/raylet/worker_pool.h
Co-authored-by: Eric Liang <ekhliang@gmail.com>
* dedicated workers
* Fix tests
* x
* fix
* asan
* asan
* Workers can only exec tasks with same job ID
* size_t for runtime env hash, fix unit tests
* include job ID in runtime env hash, remove from worker registration msg
* x
* conflict
* debug
* Schedule and dispatch periodically, skip if no new tasks
* Update src/ray/common/task/task_spec.h
Co-authored-by: Eric Liang <ekhliang@gmail.com>
* Update src/ray/raylet/scheduling/cluster_task_manager.h
Co-authored-by: Eric Liang <ekhliang@gmail.com>
* Update src/ray/raylet/worker_pool.h
Co-authored-by: Eric Liang <ekhliang@gmail.com>
Co-authored-by: Eric Liang <ekhliang@gmail.com>
* exp backoff
* up
* format
* up
* up
* up
* up
* up
* format
* fix
* up
* format
* adjust ordering
* up
* Revert "[tune] Cache unstaged placement groups for potential re-use (#18706)"
This reverts commit 2e99fb215f.
* up
* update
* format
* up
* format
* fix
* Revert "Revert "[tune] Cache unstaged placement groups for potential re-use (#18706)""
This reverts commit 93425fdb986059e53699623a0fc8590c062e139b.
* up
* format
* fix lint
* up
* up
* up
* up
* check
* add test1
* format
* up
* add test
* up
* up
* up
* fix
* up
* up
* up
* add test
* format
* up
* up
* fix lint
* format
* fix
* format
* fix
* up
