[tune] Evict object cache for actor re-use when search ended - #33593
Merged
Conversation
Signed-off-by: Kai Fricke <kai@anyscale.com>
gjoliver
approved these changes
Mar 22, 2023
Member
|
thanks for the quick fix. |
Signed-off-by: Kai Fricke <kai@anyscale.com>
elliottower
pushed a commit
to elliottower/ray
that referenced
this pull request
Apr 22, 2023
…ject#33593) Similar regression that was originally fixed in ray-project#31974, but re-surfaced after ray-project#33045. With actor re-use, we speculatively keep one cached actor around in case it is needed when new trials are added (e.g. if we add trials one-by-one). However, we should only do this when the search has not ended, as otherwise we keep an extra actor until the end of the experiment, wasting resources. This leads the tune_scale_up_down release test fail. In our refactor to ray-project#33045 we generalized object caching, but the eviction logic here explicitly keeps one object in cache. Our previous implementation in RayTrialExecutor relied on not calling the eviction function at all when trials were still coming up, essentially not adjusting the number of cached actors down. This is rarely a problem in practice, but does not make a clean contract when separated out into a component. In this PR, we change the logic as follows: When the search ended, no trials are pending execution, and we don't want to explicitly cache an actor, we force eviction of all cached objects. We are refactoring our execution backend, thus I believe it's sufficient to keep the release test to catch this regression. In the new backend we can add light weight unit tests to capture this behavior. Signed-off-by: Kai Fricke <kai@anyscale.com> Signed-off-by: elliottower <elliot@elliottower.com>
ProjectsByJackHe
pushed a commit
to ProjectsByJackHe/ray
that referenced
this pull request
May 4, 2023
…ject#33593) Similar regression that was originally fixed in ray-project#31974, but re-surfaced after ray-project#33045. With actor re-use, we speculatively keep one cached actor around in case it is needed when new trials are added (e.g. if we add trials one-by-one). However, we should only do this when the search has not ended, as otherwise we keep an extra actor until the end of the experiment, wasting resources. This leads the tune_scale_up_down release test fail. In our refactor to ray-project#33045 we generalized object caching, but the eviction logic here explicitly keeps one object in cache. Our previous implementation in RayTrialExecutor relied on not calling the eviction function at all when trials were still coming up, essentially not adjusting the number of cached actors down. This is rarely a problem in practice, but does not make a clean contract when separated out into a component. In this PR, we change the logic as follows: When the search ended, no trials are pending execution, and we don't want to explicitly cache an actor, we force eviction of all cached objects. We are refactoring our execution backend, thus I believe it's sufficient to keep the release test to catch this regression. In the new backend we can add light weight unit tests to capture this behavior. Signed-off-by: Kai Fricke <kai@anyscale.com> Signed-off-by: Jack He <jackhe2345@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why are these changes needed?
Similar regression that was originally fixed in #31974, but re-surfaced after #33045.
With actor re-use, we speculatively keep one cached actor around in case it is needed when new trials are added (e.g. if we add trials one-by-one). However, we should only do this when the search has not ended, as otherwise we keep an extra actor until the end of the experiment, wasting resources. This leads the
tune_scale_up_downrelease test fail.In our refactor to #33045 we generalized object caching, but the eviction logic here explicitly keeps one object in cache. Our previous implementation in RayTrialExecutor relied on not calling the eviction function at all when trials were still coming up, essentially not adjusting the number of cached actors down. This is rarely a problem in practice, but does not make a clean contract when separated out into a component.
In this PR, we change the logic as follows: When the search ended, no trials are pending execution, and we don't want to explicitly cache an actor, we force eviction of all cached objects.
We are refactoring our execution backend, thus I believe it's sufficient to keep the release test to catch this regression. In the new backend we can add light weight unit tests to capture this behavior.
Related issue number
Checks
git commit -s) in this PR.scripts/format.shto lint the changes in this PR.method in Tune, I've added it in
doc/source/tune/api/under thecorresponding
.rstfile.