Added StartupToken to idenitfy a process at startup #19014

czgdp1807 · 2021-09-30T17:31:43Z

Why are these changes needed?

The changes in this PR introduce StartupToken to uniquely identify a process at startup. This helps in avoiding conflicts due to differing PIDs on Windows due to wrapper processes. For more details please refer #18951 (comment)

After this PR, ray.init() completes without hanging in virtual and conda environments.

Related issue number

Closes #18951
Closes #18952
#15970

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 :(

cc: @dharhas @mattip

czgdp1807 · 2021-09-30T18:40:02Z

This branch builds successfully and behaving as desired with ray.init() on my Windows system. However it is failing on Github Actions. Anyways, I will be testing it tomorrow locally and will update here. Meanwhile if you have any comments/feedbacks please let me know. Thanks.

simon-mo · 2021-09-30T22:54:52Z

opps sorry I misclicked, please adjust to draft PR if not intended.

pcmoritz · 2021-10-01T05:27:55Z

src/ray/core_worker/core_worker.h

@@ -186,6 +187,7 @@ struct CoreWorkerOptions {
  int runtime_env_hash;
  /// The PID of the process for setup worker runtime env.
  pid_t worker_shim_pid;
+  StartupToken startup_token;


Please add a comment above describing what this does :)

I will add documentation comments at all the necessary places once I complete testing. There are some internal API changes so tests are required to be updated. Thanks.

pcmoritz · 2021-10-01T05:28:49Z

src/ray/raylet/worker.h

@@ -27,6 +27,12 @@
 #include "ray/rpc/worker/core_worker_client.h"
 #include "ray/util/process.h"

+#ifdef _WIN32


we should get rid of this typedef. Why not make it int64_t everywhere?

pcmoritz · 2021-10-01T05:30:36Z

src/ray/raylet/worker_pool.cc

@@ -370,7 +371,11 @@ Process WorkerPool::StartWorkerProcess(
                << " worker(s) with pid " << proc.GetId();
  MonitorStartingWorkerProcess(proc, language, worker_type);
  state.starting_worker_processes.emplace(
-      proc, StartingWorkerProcessInfo{workers_to_start, workers_to_start, worker_type});
+      startup_token_,
+      std::make_pair(


Can you put startup_token into the StartingWorkerProcessInfo? Then we don't need an additional level of indirections :)

The core of the fix is using startup_token as key. I can additionally add it in StartingWorkerProcessInfo and make necessary changes else where.

Ah sorry, I meant put proc into StartingWorkerProcessInfo and get rid of the std::make_pair (startup_token should be the key yes). It mainly seems odd to use both std::make_pair here and also have the struct. This will simplify the code elsewhere because then you don't need the additional first and second calls.

pcmoritz · 2021-10-01T05:31:56Z

src/ray/raylet/worker_pool.h

@@ -38,6 +39,12 @@ namespace ray {

 namespace raylet {

+#ifdef _WIN32


This shouldn't be needed since it is already in src/ray/raylet/worker.h. Just make sure it is in the right namespace there.

pcmoritz · 2021-10-01T05:33:52Z

src/ray/raylet_client/raylet_client.h

@@ -29,6 +29,12 @@
 #include "src/ray/protobuf/common.pb.h"
 #include "src/ray/protobuf/gcs.pb.h"

+#ifdef _WIN32


Actually maybe put it in ray/util/process.h that should be included by everybody who needs it :)

pcmoritz · 2021-10-01T05:35:11Z

Thanks a lot for fixing this @czgdp1807 :)

Can you shepherd this through the process (another round of reviews and making sure tests are fixed etc) @wuisawesome ?

src/ray/raylet/worker_pool_test.cc

czgdp1807 · 2021-10-03T13:28:20Z

Following tests fail on my branch,

FAIL: //:gcs_server_rpc_test 
FAIL: //:core_worker_test 
FAIL: //:event_test 
FAIL: //:gcs_server_test

The above tests fail on master as well with similar logs. So, I think that these failures are unrelated to my changes. I will add the documentation now. Some python tests failing on this branch are flaky I think. They end with RedisConnectionError. Moreover, this seems to be unrelated because if there would have been any error due to API changes then the syntax/semantic error should have been received from Python.

I have added documentation at necessary places in the diff. This PR is ready from my side. Please let me know if anything else is to be done. cc: @pcmoritz @wuisawesome @dharhas

wuisawesome · 2021-10-04T17:20:39Z

@czgdp1807 it looks like the build is failing on all non-windows platforms right now. Can you fix the initialization order?

Use --sandbox_debug to see verbose messages from the sandbox
--
  | In file included from src/ray/raylet/worker_pool.cc:15:
  | bazel-out/k8-opt/bin/_virtual_includes/raylet_lib/ray/raylet/worker_pool.h: In constructor ‘ray::raylet::WorkerPool::WorkerPool(instrumented_io_context&, ray::NodeID, std::string, int, int, int, int, int, const std::vector<int>&, std::shared_ptr<ray::gcs::GcsClient>, const WorkerCommandMap&, std::function<void()>, int, std::function<double()>)’:
  | bazel-out/k8-opt/bin/_virtual_includes/raylet_lib/ray/raylet/worker_pool.h:608:21: error: ‘ray::raylet::WorkerPool::node_address_’ will be initialized after [-Werror=reorder]
  | 608 \|   const std::string node_address_;
  | \|                     ^~~~~~~~~~~~~
  | bazel-out/k8-opt/bin/_virtual_includes/raylet_lib/ray/raylet/worker_pool.h:141:16: error:   ‘ray::StartupToken ray::raylet::WorkerPool::startup_token_’ [-Werror=reorder]
  | 141 \|   StartupToken startup_token_;
  | \|                ^~~~~~~~~~~~~~
  | src/ray/raylet/worker_pool.cc:60:1: error:   when initialized here [-Werror=reorder]
  | 60 \| WorkerPool::WorkerPool(instrumented_io_context &io_service, const NodeID node_id,
  | \| ^~~~~~~~~~
  | cc1plus: all warnings being treated as errors

wuisawesome · 2021-10-04T17:22:13Z

src/ray/core_worker/core_worker.h

@@ -186,6 +187,9 @@ struct CoreWorkerOptions {
  int runtime_env_hash;
  /// The PID of the process for setup worker runtime env.
  pid_t worker_shim_pid;
+  /// The startup token of the process assgined to it
+  /// during startup via command line arguments.


Suggested change

/// during startup via command line arguments.

/// during startup via command line arguments. This is needed because the actual core worker process may not have the same pid as the process the worker pool starts (due to shim processes).

Should I change this everywhere or doing it only once here would suffice?

czgdp1807 · 2021-10-04T19:05:08Z

I have made the change @wuisawesome. Thanks for pointing out.

scv119 · 2021-10-14T17:34:41Z

src/ray/raylet/worker_pool.h

@@ -132,6 +133,12 @@ class Worker;
 /// The WorkerPool is responsible for managing a pool of Workers. Each Worker
 /// is a container for a unit of work.
 class WorkerPool : public WorkerPoolInterface, public IOWorkerPoolInterface {
+ protected:


any reason this should not be private? also as a general guideline, this should be declared after function declaration.

We need this value in WorkerPoolMock (inherits WorkerPool) in worker_pool_test.cc. The other way to expose this member is via a protected function in WorkerPool which can be called by WorkerPoolMock. I do not want to expose startup_token via public method. Making a protected data member appeared simpler to me instead of making a protected function and then using it everywhere else. Let me know if you have other ideas.

sounds reasonable.

scv119 · 2021-10-14T17:37:28Z

src/ray/raylet/worker_pool.h

+  /// Gloabl startup token variable. Incremented once assigned
+  /// to a worker process and is added to
+  /// state.starting_worker_processes.
+  StartupToken startup_token_;


nit: worker_startup_token_counter_;
startup_token_ is a bit confusing.

scv119 · 2021-10-14T17:49:17Z

src/ray/raylet/worker_pool.cc

@@ -559,9 +575,10 @@ Status WorkerPool::RegisterWorker(const std::shared_ptr<WorkerInterface> &worker
 void WorkerPool::OnWorkerStarted(const std::shared_ptr<WorkerInterface> &worker) {
  auto &state = GetStateForLanguage(worker->GetLanguage());
  const auto &shim_process = worker->GetShimProcess();
+  const StartupToken &worker_startup_token = worker->GetStartupToken();


doesn't need to be reference.

scv119 · 2021-10-14T17:54:00Z

src/ray/raylet/worker_pool.h

@@ -473,7 +484,8 @@ class WorkerPool : public WorkerPoolInterface, public IOWorkerPoolInterface {
    /// A map from the pids of this shim processes to the extra information of


this comment need to be updated.

scv119 · 2021-10-14T17:59:02Z

src/ray/raylet/worker_pool.cc

@@ -516,6 +525,7 @@ Status WorkerPool::RegisterWorker(const std::shared_ptr<WorkerInterface> &worker
  }
  auto process = Process::FromPid(pid);
  worker->SetProcess(process);
+  worker->SetStartupToken(worker_startup_token);


if we can move

worker->SetProcess(process); worker->SetStartupToken(worker_startup_token);

to worker's constructor and make them const member variables we could just get rid of a category of errors

scv119

Thanks a lot for helping out on the Windows issues!

scv119 · 2021-10-14T18:02:43Z

src/ray/core_worker/core_worker.h

+  /// This is needed because the actual core worker process
+  /// may not have the same pid as the process the worker pool
+  /// starts (due to shim processes).
+  StartupToken startup_token{-1};


is there any special meaning for -1? maybe worth documenting if that's the case.

The startup tokens usually start from 0. So, set it to -1 so that first process created gets 0 as the startup token.

scv119 · 2021-10-14T18:03:53Z

src/ray/raylet/worker_pool.cc

-      proc, StartingWorkerProcessInfo{workers_to_start, workers_to_start, worker_type});
+      startup_token_,
+      StartingWorkerProcessInfo{workers_to_start, workers_to_start, worker_type, proc});
+  startup_token_ += 1;


imo this is a bit fragile... maybe have a helper function called get_next_worker_startup_token

scv119 · 2021-10-14T18:04:25Z

src/ray/raylet/worker_pool.cc

@@ -360,6 +366,12 @@ Process WorkerPool::StartWorkerProcess(
    env.insert({"SPT_NOENV", "1"});
  }

+  if (language == Language::PYTHON) {
+    worker_command_args.push_back("--startup-token=" + std::to_string(startup_token_));


a bit sad we have these slight difference between different languanges

scv119

overall looks good to me. @wuisawesome takes another look before we merge this? :)

scv119 · 2021-10-15T12:39:34Z

src/ray/raylet/worker.cc

 void Worker::SetProcess(Process proc) {
  RAY_CHECK(proc_.IsNull());  // this procedure should not be called multiple times
  proc_ = std::move(proc);
 }

+void Worker::SetStartupToken(StartupToken startup_token) {


is this only used for test? we can make it private and use FRIEND_TEST/friend to access it.

Ah! right. This can be protected as it isn't required to be exposed publicly. What do you say?

davidberenstein1957 · 2021-10-25T10:56:10Z

@pcmoritz
This still seems to result in some process id issues for the agent_manager.
After deployment of the ray cluster locally within a poetry environment.
Example:
import ray import os ray.init(num_cpus=os.cpu_count(), namespace=namespace, include_dashboard=False) serve.start(detached=False)
Error:
(pid=None) [2021-10-25 12:41:57,798 E 5088 10652] agent_manager.cc:90: [ Event 19616925b8e76a75abc202154fcbf045c675 {"ip":"10.110.71.210","node_id":"84a2e03c8733f7ad48b0aa4e7a84911cd937d785fa7c829175cfd08d","pid":1148} ] Agent process with pid 3464 has not registered, restart it.

czgdp1807 · 2021-10-25T11:10:17Z

Hi @davidberenstein1957. I never saw this error before. Can I try re-producing it on my system and investigate if it's somehow related to the changes we have made here. AFAIK, the changes here only deal with worker processes and not agent processes. There are two things that I will do for re-producing this error - i) check on master build and ii) on some previous version before this PR to make sure if this is a result of any ripple effects. In any case, I am happy to investigate it further. Let me know @pcmoritz @davidberenstein1957

davidberenstein1957 · 2021-10-25T11:54:32Z

@czgdp1807
For the entire reproduction, see the following environment and script.
ENVIRONMENT

[tool.poetry]
name = "pandora.nlp.gpt"
version = "0.1.0"
description = ""
authors = ["David Berenstein"]

[tool.poetry.dependencies]
python = "3.8.2"
fastapi = "0.62.0"
spacy = "2.3.2"
spacy-langdetect = "0.1.2"
torch = "1.8.1"
fast-bert = "^1.9.8"
uvicorn = "^0.13.4"
pytz = "2020.5"
scikit-learn = "^0.24.2"
transformers = "^4.9.2"

[tool.poetry.dev-dependencies]
flake8 = "^3.9.2"
black = {version = "^21.5b1", allow-prereleases = true}
autopep8 = "^1.5.7"

[build-system]
requires = ["poetry-core>=1.0.0"]
build-backend = "poetry.core.masonry.api"

SCRIPT:

from fastapi import FastAPI
from ray import serve
import ray
import os
from starlette.middleware.cors import CORSMiddleware
from transformers import pipeline

app = FastAPI(
    title="GPT",
    description="GPT API - includes classification endpoints.",
    version="1.0.0",
    docs_url="/",
)

ray.init(num_cpus=os.cpu_count()-1, namespace=namespace, include_dashboard=False)
serve.start(detached=detached)

origins = [
    "http://localhost",
    "http://localhost:88",
]

@serve.deployment(num_replicas=1, max_concurrent_queries=100, ray_actor_options={'num_cpus': 0.25, 'num_gpus': 0, 'memory':1024*1024*1000})
class GPT2:
        def __init__(self):
            self.nlp_model = pipeline("text-generation", model="gpt2")
        async def __call__(self, request):
            return self.nlp_model(request.data, max_length=50)

GPT2..options(
    name='gpt2_test'
).deploy()

app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

@main_router.post("/predict", tags=["Model"])
async def predict(query:str):
    handle = serve.get_deployment(name).get_handle(sync=False)
    return ray.get(await handle.remote(**kwargs))

RUN APP
``
uvicorn main:app --port 88

davidberenstein1957 · 2021-10-26T08:15:50Z

I get a comparable message when enabling GPU without full 'ray[default]' installment, even though I disabled dashboard usage. Same code as above, but with a initialization with GPU.

ray.init(num_cpus=os.cpu_count()-1, num_gpus=1, namespace=namespace, include_dashboard=False)
  @serve.deployment(num_replicas=1, max_concurrent_queries=100, 
    ray_actor_options={'num_cpus': 0.25, 'num_gpus': 1, 'memory':1024*1024*1000})

(pid=None) C:\Users\david\Documents\programming\Pandora.NLP.Classification\.venv\lib\site-packages
\ray\dashboard\modules\reporter\reporter_agent.py:37: 
    UserWarning: `gpustat` package is not installed. GPU monitoring is not available. 
    To have full functionality of the dashboard please install `pip install ray[default]`.)

mattip · 2021-10-26T09:07:42Z

I reformatted your error message to make the error clearer. You should install gpustat >= 1.0.0b1 for windows.

pcmoritz · 2021-11-12T03:47:29Z

@davidberenstein1957 Thanks for reporting! I just went through your script. There were a few typos, but after fixing them (see attached), on my machine the problem was #19948, and after running pip install aiohttp==3.7 the error message was gone. We have a final fix for this in #20261, which will be part of the next release!

In the case that only ray and not ray[default] is installed, you will still get the UserWarning but that's expected :)

Glad to know the issue is unrelated to this PR though (which makes sense, since the PID of the agent is actually handled correctly and doesn't suffer from the issue that was fixed in this PR, as it calls os.pid() directly in the relevant process instead of trying to get the PID of the child).

For anybody curious how the aiohttp issue could cause this: Basically ray._private.utils.check_dashboard_dependencies_installed() returns False even if ray[default] is installed because of aiohttp.signals is not found.

Please let me know if the problem persists for you even after pinning aiohttp!

from fastapi import FastAPI
from ray import serve
import ray
import os
from starlette.middleware.cors import CORSMiddleware
from transformers import pipeline
 
app = FastAPI(
    title="GPT",
    description="GPT API - includes classification endpoints.",
    version="1.0.0",
    docs_url="/",
)
 
ray.init(num_cpus=os.cpu_count()-1, namespace="test", include_dashboard=False)
serve.start(detached=False)
 
origins = [
    "http://localhost",
    "http://localhost:88",
]
 
@serve.deployment(num_replicas=1, max_concurrent_queries=100, ray_actor_options={'num_cpus': 0.25, 'num_gpus': 0, 'memory':1024*1024*1000})
class GPT2:
        def __init__(self):
            self.nlp_model = pipeline("text-generation", model="gpt2")
        async def __call__(self, request):
            return self.nlp_model(request, max_length=50)
 
GPT2.options(
    name='gpt2_test'
).deploy()
 
app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)
 
@app.post("/predict", tags=["Model"])
async def predict(query:str):
    handle = serve.get_deployment("gpt2_test").get_handle(sync=False)
    return ray.get(await handle.remote(query))

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.

Added StartupToken to idenitfy a process at startup

d31e556

czgdp1807 added the windows label Sep 30, 2021

czgdp1807 requested review from pcmoritz and wuisawesome September 30, 2021 17:31

Applied linting formats

d0982e5

richardliaw assigned pcmoritz and wuisawesome Sep 30, 2021

simon-mo marked this pull request as ready for review September 30, 2021 22:54

pcmoritz reviewed Oct 1, 2021

View reviewed changes

Addressed reviews

80ab85b

czgdp1807 force-pushed the fix_ray_init branch from 5b1f8a5 to da4bf7e Compare October 1, 2021 18:41

czgdp1807 commented Oct 1, 2021

View reviewed changes

src/ray/raylet/worker_pool_test.cc Outdated Show resolved Hide resolved

Fixing worker_pool_test

8045363

czgdp1807 force-pushed the fix_ray_init branch from da4bf7e to 8045363 Compare October 1, 2021 18:44

Fixed worker_pool_test

2a46f62

czgdp1807 added 3 commits October 4, 2021 16:45

Applied linting formatting

6af350a

Added documentation for StartupToken

f26d416

Fixed linting

161a4f4

wuisawesome reviewed Oct 4, 2021

View reviewed changes

Reordered initialisation of WorkerPool members

99d3f6c

scv119 self-assigned this Oct 14, 2021

scv119 reviewed Oct 14, 2021

View reviewed changes

czgdp1807 added 3 commits October 15, 2021 13:22

Addressed reviews and applied scripts/format.sh

be78bd5

Merge branch 'master' into fix_ray_init

128c08a

Applied linting script from ci/travis

8698663

scv119 reviewed Oct 15, 2021

View reviewed changes

czgdp1807 added 2 commits October 15, 2021 22:43

Removed unrequired methods from public scope

eeccf6c

Applied linting

e10c5ee

pcmoritz approved these changes Oct 15, 2021

View reviewed changes

pcmoritz merged commit d226cbf into ray-project:master Oct 15, 2021

czgdp1807 deleted the fix_ray_init branch October 15, 2021 22:15

This was referenced Oct 18, 2021

No output in windows machine. #15970

Closed

Cannot get ray to start on Windows (stuck on init) #16137

Closed

This comment has been minimized.

Sign in to view

This was referenced Dec 6, 2021

Move the worker_startup_token_counter_ to the Process for process ha… #20898

Closed

Startup token ownership #20937

Closed

Starting workers map #20986

Merged

pcmoritz mentioned this pull request Dec 13, 2021

Remove hash function from Process class #21045

Closed

6 tasks

czgdp1807 restored the fix_ray_init branch February 24, 2022 09:54

czgdp1807 deleted the fix_ray_init branch May 7, 2022 06:30

		@@ -38,6 +39,12 @@ namespace ray {

		namespace raylet {

		#ifdef _WIN32

	/// during startup via command line arguments.
	/// during startup via command line arguments. This is needed because the actual core worker process may not have the same pid as the process the worker pool starts (due to shim processes).

		@@ -473,7 +484,8 @@ class WorkerPool : public WorkerPoolInterface, public IOWorkerPoolInterface {
		/// A map from the pids of this shim processes to the extra information of

Added StartupToken to idenitfy a process at startup #19014

Added StartupToken to idenitfy a process at startup #19014

Conversation

czgdp1807 commented Sep 30, 2021 • edited Loading

Why are these changes needed?

Related issue number

Checks

czgdp1807 commented Sep 30, 2021

simon-mo commented Sep 30, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

pcmoritz commented Oct 1, 2021

czgdp1807 commented Oct 3, 2021 • edited Loading

wuisawesome commented Oct 4, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

czgdp1807 commented Oct 4, 2021

Choose a reason for hiding this comment

czgdp1807 Oct 14, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

scv119 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

scv119 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

davidberenstein1957 commented Oct 25, 2021 • edited Loading

czgdp1807 commented Oct 25, 2021

davidberenstein1957 commented Oct 25, 2021 • edited Loading

davidberenstein1957 commented Oct 26, 2021 • edited by mattip Loading

mattip commented Oct 26, 2021

This comment has been minimized.

pcmoritz commented Nov 12, 2021 • edited Loading

czgdp1807 commented Sep 30, 2021 •

edited

Loading

czgdp1807 commented Oct 3, 2021 •

edited

Loading

czgdp1807 Oct 14, 2021 •

edited

Loading

davidberenstein1957 commented Oct 25, 2021 •

edited

Loading

davidberenstein1957 commented Oct 25, 2021 •

edited

Loading

davidberenstein1957 commented Oct 26, 2021 •

edited by mattip

Loading

pcmoritz commented Nov 12, 2021 •

edited

Loading