Fix infinite retry in Push function. #2133

guoyuhong · 2018-05-24T04:35:17Z

Hello!
This is Yuhong Guo from Ant Financial. I'm trying to do some small code change to go through the full code review process.

What do these changes do?

Remove the infinite call in Push function.
Fix some comment mismatches.

Related issue number

N/A

AmplabJenkins · 2018-05-24T05:07:46Z

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5613/
Test FAILed.

robertnishihara

Thanks for the fixes, this looks good to me. It'd be great if @elibol or @stephanie-wang could comment on the retry design.

robertnishihara · 2018-05-24T07:13:21Z

src/ray/object_manager/object_manager.cc

+    if (retry < 0) {
+      retry = config_.max_retries;
+    } else if (retry == 0) {
+      RAY_LOG(ERROR) << "Invalid Push request ObjectID: " << object_id.hex()


remove .hex(), just object_id is fine because it automatically converts to hex

robertnishihara · 2018-05-24T07:20:03Z

src/ray/object_manager/object_manager.cc

 }

-ray::Status ObjectManager::Push(const ObjectID &object_id, const ClientID &client_id) {
+ray::Status ObjectManager::Push(const ObjectID &object_id, const ClientID &client_id,


@elibol, it looks like what will typically happen is all max_retries Push calls will happen in quick succession (with essentially no delay between them unless the manager is busy with other events), is that right? Is that how we want to do this? An alternative approach is to keep a list of object/client pairs that we want to push and then do any push events that can be done when a new object becomes available.

Not suggesting that for this PR.

Yes, the pushes will happen in quick succession. The post call will queue the retry attempt on the main thread. Yes, I think the more long-term solution is to maintain a mapping from one object to many clients, and push to those clients when those objects become available locally. This mapping would be periodically cleared.

robertnishihara · 2018-05-24T07:21:16Z

src/common/state/ray_config.h

    return object_manager_max_receives_;
  }

+  int object_manager_max_retries() const { return object_manager_max_retries_; }


I think this should be more specific, e.g., object_manager_max_retries -> object_manager_max_push_retries

robertnishihara · 2018-05-24T07:21:40Z

src/ray/object_manager/object_manager.cc

      connection_pool_() {
  RAY_CHECK(config_.max_sends > 0);
  RAY_CHECK(config_.max_receives > 0);
+  RAY_CHECK(config_.max_retries > 0);


max_retries -> max_push_retries

robertnishihara · 2018-05-24T07:23:35Z

jenkins retest this please

raulchen · 2018-05-24T08:10:54Z

Instead of busy retrying without delay, I think it might make more sense to add the id in a waiting list and re-trigger the push when the object is added (in NotifyDirectoryObjectAdd function)?

Also, I'm more curios about what could make this situation happen. The only reason I can think of is that the object was evicted after Pull, and before Push (pls correct me if this's wrong). In this case, we'll also need to reconstruct the object before retrying.

AmplabJenkins · 2018-05-24T08:28:20Z

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5616/
Test PASSed.

…nfiniteRetry

AmplabJenkins · 2018-05-24T10:19:56Z

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5619/
Test FAILed.

elibol · 2018-05-24T16:38:19Z

@raulchen yes, this is rare. Your proposed solution sounds right. We would also want to periodically clear the (object_id, client_id) wait list.

But this may make sense coupled with a change to the code at task forwarding time (in the node manager): If we push task dependencies (the objects) to the node to which a task is forwarded without checking whether the object is local, some of those objects may never be local on the node that is forwarding the task. Additionally, with such changes to task forwarding, some objects may appear locally after the call to push: If we invoke an actor method, the task may be forwarded before the object store notification is received by the node manager. We currently check that an object is local before pushing, which sometimes fails to push objects for which an add notification hasn't yet been received. See https://github.com/ray-project/ray/blob/master/src/ray/raylet/node_manager.cc#L821.

elibol

These changes cap the number of retry attempts for a particular push. A longer term solution is to use notifications from the object store instead of repeatedly invoking push. Thanks for your contribution! I will approve this PR once the changes I've requested are made, and merge once tests are passing.

@guoyuhong if you're interested, I would also be happy to review a subsequent PR that improves on your solution by using object store notifications (e.g. by adding a check for ObjectID in a map from one object_id to many client_ids in the object manager handler for SubscribeObjAdded).

elibol · 2018-05-24T19:12:27Z

src/common/state/ray_config.h

  /// Maximum number of concurrent receives allowed by the object manager.
  int object_manager_max_receives_;

+  /// Maximum push retry times allowed by the object manager.


Change to "Maximum push retries allowed by the object manager."

elibol · 2018-05-24T21:07:58Z

src/ray/object_manager/object_manager.h

  /// Push an object to to the node manager on the node corresponding to client id.
  ///
  /// \param object_id The object's object id.
  /// \param client_id The remote node's client id.


Document retry.

That would be interesting. In my observation in the object_manager_stress_test, this retry mechanism is triggered, so a long-term solution is necessary.

By my understanding, we can first check the object_id in local_objects_ in the Push function. If it exists, the Push function continues. Otherwise, we can add this request to a map (as you described), and in NotifyDirectoryObjectAdd, the Push request will be re-triggered. It that correct?

After talking to @raulchen and @eric-jj , I decide to change the code to waiting-list version.

This is correct.

AmplabJenkins · 2018-05-25T01:42:30Z

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5628/
Test FAILed.

guoyuhong · 2018-05-25T02:01:13Z

@robertnishihara Hi Robert, could you please re-trigger the jenkins test. Only one test passed in previous tests, the rest tests failed due to time out... Is that expected?

robertnishihara · 2018-05-25T07:04:06Z

jenkins, retest this please

AmplabJenkins · 2018-05-25T08:10:47Z

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/5631/
Test PASSed.

comments addressed

* master: [autoscaler] GCP node provider (ray-project#2061) [xray] Evict tasks from the lineage cache (ray-project#2152) [ASV] Add ray.init and simple Ray benchmarks (ray-project#2166) Re-encrypt key for uploading to S3 from travis to use travis-ci.com. (ray-project#2169) [rllib] Fix A3C PyTorch implementation (ray-project#2036) [JavaWorker] Do not kill local-scheduler-forked workers in RunManager.cleanup (ray-project#2151) Update Travis CI badge from travis-ci.org to travis-ci.com. (ray-project#2155) Implement Python global state API for xray. (ray-project#2125) [xray] Improve flush algorithm for the lineage cache (ray-project#2130) Fix support for actor classmethods (ray-project#2146) Add empty df test (ray-project#1879) [JavaWorker] Enable java worker support (ray-project#2094) [DataFrame] Fixing the code formatting of the tests (ray-project#2123) Update resource documentation (remove outdated limitations). (ray-project#2022) bugfix: use array redis_primary_addr out of its scope (ray-project#2139) Fix infinite retry in Push function. (ray-project#2133) [JavaWorker] Changes to the directory under src for support java worker (ray-project#2093) Integrate credis with Ray & route task table entries into credis. (ray-project#1841)

Yuhong Guo added 3 commits May 24, 2018 12:09

Remove the Infinite retry of Push in ObjectManager.

6c92888

Fix some comment mismatchs.

de47a52

Fix test fail and fix lint.

7729265

robertnishihara reviewed May 24, 2018

View reviewed changes

Yuhong Guo added 2 commits May 24, 2018 16:54

Merge branch 'master' of https://github.com/ray-project/ray into FixI…

667aded

…nfiniteRetry

Change max_retries to max_push_retries

5dcc6e4

elibol previously requested changes May 24, 2018

View reviewed changes

Fix according to CR comment.

7f67e60

robertnishihara mentioned this pull request May 25, 2018

[dataframes] Travis test failure in test_groupby.py. #2136

Closed

robertnishihara approved these changes May 25, 2018

View reviewed changes

robertnishihara merged commit a8517cc into ray-project:master May 25, 2018

guoyuhong deleted the FixInfiniteRetry branch May 27, 2018 03:25

Fix infinite retry in Push function. #2133

Fix infinite retry in Push function. #2133

Uh oh!

Conversation

guoyuhong commented May 24, 2018

What do these changes do?

Related issue number

Uh oh!

AmplabJenkins commented May 24, 2018

Uh oh!

robertnishihara left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

robertnishihara commented May 24, 2018

Uh oh!

raulchen commented May 24, 2018

Uh oh!

AmplabJenkins commented May 24, 2018

Uh oh!

AmplabJenkins commented May 24, 2018

Uh oh!

elibol commented May 24, 2018

Uh oh!

elibol left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AmplabJenkins commented May 25, 2018

Uh oh!

guoyuhong commented May 25, 2018

Uh oh!

robertnishihara commented May 25, 2018

Uh oh!

AmplabJenkins commented May 25, 2018

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

elibol left a comment •

edited

Loading