Skip to content

fix(evals): extract bucket name from eval_storage_uri when a path is given - #6882

Open
chelsealong wants to merge 1 commit into
google:mainfrom
chelsealong:fix-6881-gcs-eval-bucket-path-extraction
Open

fix(evals): extract bucket name from eval_storage_uri when a path is given#6882
chelsealong wants to merge 1 commit into
google:mainfrom
chelsealong:fix-6881-gcs-eval-bucket-path-extraction

Conversation

@chelsealong

Copy link
Copy Markdown
Contributor

Fixes #6881

Problem

create_gcs_eval_managers_from_uri()'s docstring says:

eval_storage_uri: The evals storage URI to use. Supported URIs: gs://<bucket name>. If a path is provided, the bucket will be extracted.

But the implementation never extracted a bucket from a path — it treated
everything after gs:// as the literal bucket name:

gcs_bucket = eval_storage_uri.split('://')[1]

gs://my-bucket/some/path produced the invalid bucket name
my-bucket/some/path (slashes included), which is then passed to
GcsEvalSetsManager/GcsEvalSetResultsManager and fails later with a
misleading "bucket does not exist" error, even when my-bucket itself
exists.

Fix

Split off the first path segment so only the bucket name is used:

gcs_bucket = eval_storage_uri.split('://')[1].split('/')[0]

This matches the docstring's documented behavior for gs://<bucket name>
URIs. As noted in the issue, threading the remaining path through as a
storage prefix is a separate, larger change (the eval history/sets
directory helpers currently hard-code their paths) and is left out of this
minimal fix.

Testing plan

Added test_create_gcs_eval_managers_from_uri_extracts_bucket_from_path to
tests/unittests/cli/utils/test_evals.py, asserting that
gs://test-bucket/some/path results in GcsEvalSetsManager and
GcsEvalSetResultsManager being constructed with bucket_name='test-bucket'.

Confirmed the new test fails against the old code:

FAILED tests/unittests/cli/utils/test_evals.py::test_create_gcs_eval_managers_from_uri_extracts_bucket_from_path - AssertionError: expected call not found.
Expected: GcsEvalSetsManager(bucket_name='test-bucket', project='test-project')
  Actual: GcsEvalSetsManager(bucket_name='test-bucket/some/path', project='test-project')

And passes with the fix, along with the full existing suite for this module:

$ pytest tests/unittests/cli/utils/ -q
421 passed, 2 xfailed, 143 warnings in 17.90s

Also verified pyink --check and isort --check are clean on both changed
files.

AI disclosure

This change was authored with the assistance of an AI coding agent (Claude),
with the diff reviewed before submission.

…given

create_gcs_eval_managers_from_uri() treated everything after `gs://` as
the literal bucket name, so `gs://my-bucket/some/path` produced the
invalid bucket name `my-bucket/some/path`, contradicting the docstring
which says the bucket will be extracted if a path is provided.

Fixes google#6881
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

create_gcs_eval_managers_from_uri doesn't extract bucket from path, contradicting its own docstring

2 participants