Dont retry on connection failure #103
Merged
Conversation
This file contains 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
jaideepr97
approved these changes
Aug 19, 2024
There was a problem hiding this comment.
Thanks @danmcp @booxter
I really like how the errors have been broken down by their codes - it helps to further my understanding of good error handling in general
we could probably start focusing on expanding the unit testing infrastructure on the project but that should probably be in a separate PR
danmcp
force-pushed
the
dont-retry-on-connection-failure
branch
from
a356211 to
1561eee
Compare
The `python` symlink may be missing on a system; or even point to py2. We should use `python3` to be sure. (It's ok to use `python` inside a virtualenv though.) Signed-off-by: Ihar Hrachyshka <[email protected]>
Before the patch, all errors from openai were handled by retrying up to API_MAX_RETRY times and returning $ERROR$ message at the last attempt. With this patch, if all attempts result in APIConnectionError, we raise a new EvalError exception. (If at least one of the previous attempts result in a different error, then we return $ERROR$ as usual.) Also, several errors are not expected to recover with a retry (400-404, 422). This patch makes them return $ERROR$ immediately without retrying. Closes: instructlab#77 Signed-off-by: Ihar Hrachyshka <[email protected]>
Before the patch, we were calculating them on every retry attempt. The function is pure, so there is no good reason to repeat the calculation. This also simplifies the function a bit. Signed-off-by: Ihar Hrachyshka <[email protected]>
Signed-off-by: Dan McPherson <[email protected]>
danmcp
force-pushed
the
dont-retry-on-connection-failure
branch
from
1561eee to
7fbd87e
Compare
nathan-weinberg
approved these changes
Aug 20, 2024
booxter
reviewed
Aug 21, 2024
| openai.PermissionDeniedError, # 403 | ||
| openai.NotFoundError, # 404 | ||
| # General catch-all | ||
| openai.OpenAIError, |
There was a problem hiding this comment.
So AFAIU all the explicit types listed above are for documentation purposes only? Because all of them are OpenAIErrors already.
There was a problem hiding this comment.
Yes. I thought they were really helpful to see elaborated. Although it's a great point that we should add a comment stating that logic explicitly.
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.
This is a continuation/replacement of #80
Resolves: #77
The approach of this separates errors into 3 buckets:
Fatal errors we retry and fail after max attempts (Ex: APIConnectionError, 404). These cases don't have much hope of fixing themselves.
Errors which we record after max attempts (Ex: RateLimitError). Failure on one try might not mean they are all going to fail.
Errors which we record after 1 occurrence (Ex: 400). Failure is request specific and retries aren't going to help.
Which allows the logic to fail fast when possible, not retry when it wouldn't help, and not catastrophically fail on isolated failures.