Repository navigation
e2e: give first test the full RUNPOD_INIT_TIMEOUT as cold-start buffer - #347
Merged
Merged
Conversation
The per-test deadline was timeout(300s) + 600s buffer = 900s, while workers are allowed 1600s to initialize (RUNPOD_INIT_TIMEOUT, added in #341). When cold start exceeded 15 minutes (fresh 9GB image pull + 61GB gpt-oss-120b weight download + CUDA graph capture), basic_inference_test gave up before the worker came up; the following test then passed on the warm worker, so the suite failed despite a healthy image. Default the buffer to the full init timeout (first-test deadline 1900s) and derive both values from one constant so they cannot drift apart.
colin99d
approved these changes
Sep 25, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
The per-test deadline in
scripts/serverless_e2e_test.pywastimeout (300s) + cold-start buffer (600s)= 900s, while workers are allowed 1600s to initialize (RUNPOD_INIT_TIMEOUT, added in #341). When cold start exceeded 15 minutes — fresh 9GB image pull, 61GB gpt-oss-120b weight download at runtime, ~90s of CUDA graph capture — the first test gave up before the worker came up, and the following test then passed on the warm worker, failing the suite despite a healthy image.Observed on #345 (
feature/bump-vllm-v0.30.0, run 35910590767):basic_inference_testtimed out at exactly the 900s mark,openai_messages_testpassed ~7.5 min later. Reproduced independently with the same CI image and script: cold start was ~9.4 min and both tests passed — i.e. the failure is a marginal timeout, not a worker defect.Change
RUNPOD_INIT_TIMEOUT_SECONDS = 1600as a single constant used for both the template env var and the buffer default, so the two can't drift apart again.--cold-start-buffer-secondsfrom 600 → 1600 (first-test deadline 900s → 1900s = init window + test timeout). CLI override unchanged.Verified with
python3 scripts/serverless_e2e_test.py --helpand a live end-to-end run of the harness against the CI-built v0.30.0 image.Test plan
--helpcheck