[YAshhh29]

Why

The Google provider enabled google-cloud-aiplatform[evaluation] for every installation, even though its evaluation dependency stack is only needed by Vertex AI generative-model evaluation APIs. Users of unrelated Google hooks and operators therefore received unnecessary ML and evaluation dependencies.

What changed

• Keep google-cloud-aiplatform>=1.164.0 in the base provider dependencies without enabling evaluation.
• Add a vertex-eval provider extra that installs google-cloud-aiplatform[evaluation]>=1.164.0.
• Load vertexai.preview.evaluation only when an evaluation API is called and provide an actionable installation error when it is unavailable.
• Update generated dependency tables and add a breaking-change migration note.
• Add regressions for the dependency metadata split and lazy evaluation import.

Migration

Users of GenerativeModelHook.get_eval_task, GenerativeModelHook.run_evaluation, or RunEvaluationOperator must install:

pip install 'apache-airflow-providers-google[vertex-eval]'

Validation

• Dependency metadata regression: 1 passed; the same assertions fail against upstream/main.
• uv lock --check: passed, resolving 965 packages.
• ruff format and ruff check: passed.
• git diff --check: passed.
• GitHub mergeability check: passed after rebasing onto current main.

Full provider tests were not run locally because Airflow does not support native Windows; CI provides the Linux test result.

related: #69323

---

Was generative AI tooling used to co-author this PR?

• Yes - GitHub Copilot

Generated-by: GitHub Copilot following the guidelines

---

Drafted-by: GitHub Copilot (no human review before posting)

[YAshhh29]

Rebased onto latest main and fixed the Static Checks failure: the sync-provider-readme hook wanted to update the Requirements table in README.rst from google-cloud-aiplatform[evaluation] → google-cloud-aiplatform to match the new pyproject.toml base dep. I'd missed that during the initial edit — the docs index.rst had the right value but the README's Requirements table hadn't been re-synced.

The remaining core-DB test failures on the previous run (MySQL/Postgres/Sqlite/SQLAlchemy variants) don't touch anything in this diff and looked like a main-wide flake; the rebase should clear them.

---

Drafted-by: GitHub Copilot (Claude Opus 4.6); reviewed by @YAshhh29 before posting

[YAshhh29]

Rebased onto latest main - the earlier core-DB failures were a stale-base flake and are now cleared, so all checks are green.
This is a small, self-contained dependency change (moves the heavy google-cloud-aiplatform[evaluation] stack behind an opt-in vertex-eval extra).
Would appreciate a review from a Google-provider maintainer when convenient. Thanks!

[github-actions]

uv.lock on main just moved via #71924 ("[main] Upgrade important CI environment"), commit 4a247b9 and this PR currently conflicts.

Quickest fix:

git fetch upstream main && git rebase upstream/main
rm uv.lock && uv lock
git add uv.lock && git rebase --continue
git push --force-with-lease

Automated nudge — ignore if you're not ready to rebase. This comment is updated in place on future uv.lock bumps.

[YAshhh29]

Quick update: I rebased this PR onto the latest main and resolved the changelog conflict.

While rebasing, I also tightened the regression coverage around the behavior this PR changes.
The tests now verify that Vertex AI evaluation dependencies remain opt-in through vertex-eval, and that the hook can still be imported without the evaluation stack installed.

uv lock --check, Ruff, and the focused metadata test all pass. GitHub now marks the PR as mergeable.
A fresh review would be appreciated.

[potiuk]

Hello @YAshhh29 - thank you for your contributions to Apache Airflow!

The Airflow community has introduced a limit of 5 open pull requests at a time for contributors without write access to the repository. You currently have 6 open pull requests, so - as a one-time step of introducing the limit - we closed the ones where maintainers have not engaged yet:

• Make Vertex AI evaluation stack an opt-in provider extra #69385 - Make Vertex AI evaluation stack an opt-in provider extra
• Add QdrantSearchOperator for vector similarity search #69673 - Add QdrantSearchOperator for vector similarity search
• Make generated AIP playbooks portable and verifiable #70477 - Make generated AIP playbooks portable and verifiable
• Fix upstream failure branch in task lifecycle diagram #71101 - Fix upstream failure branch in task lifecycle diagram

These pull requests stay open because maintainers are already engaged in them - they count towards your limit:

• Support Pydantic structured outputs in OpenAIResponseOperator #69812 - Support Pydantic structured outputs in OpenAIResponseOperator
• Fingerprint durable agent steps that hold sets or values that are not JSON #71575 - Do not replay cached AI steps that cannot be verified

This is not a judgement of you or of your changes. We never told contributors before that opening many pull requests at once was a problem, so there is nothing to feel bad about - and nothing is lost: your branches, commits and the review history stay where they are.

What we ask you to do is to make your first prioritization decision: choose which of the pull requests above matter most to you, and reopen them (up to 5 open at a time, including the ones still open) with the "Reopen pull request" button or gh pr reopen <PR_NUMBER> --repo apache/airflow. Reopen the ones you are ready to follow through - keep them rebased, respond to review comments and fix failing checks.

While your pull requests are waiting for review, the most valuable thing you can do is help in other ways - reviewing other contributors' pull requests, helping with issues, and taking part in the discussions on the devlist and Slack.

Why we introduced the limit, what it means for you and how to reopen or restore a pull request is explained in https://github.com/apache/airflow/blob/main/contributing-docs/32_open_pull_request_limit.rst.

---

Drafted-by: Claude Code (Opus 5); reviewed by @potiuk before posting