Can you review this PR and add any suggestions @cburroughs

Maintainer action needed per the Feast development guide: please add the kind/feature label and approve/run the fork PR workflows (or apply ok-to-test if that is the preferred process). I do not have repository permissions to add labels or approve Actions runs from the fork.

@XuananLe please check the AI review results. :). Also the Lint issues.

Addressed the Copilot AI review items in b3731e6:

• Build a typed PyFlink Schema for pandas entity DataFrames instead of passing list[str] as the schema.
• Stream FlinkOutputNode writes through Table.execute().collect() into Arrow batches, so materialization writes no longer collect the full table via to_pandas() first.
• Track and drop Feast-created Flink temporary views after retrieval/materialization execution.
• Render TTL lower-bound SQL as repeated interval subtraction (entity_ts - INTERVAL ... - INTERVAL ...) instead of subtracting a parenthesized sum of intervals.

Also added focused unit coverage for those cases and reran:

• uv run python -m pytest sdk/python/tests/unit/infra/compute_engines/flink/test_flink_compute_engine.py -q
• uv run ruff check sdk/python/feast/infra/compute_engines/flink sdk/python/tests/unit/infra/compute_engines/flink/test_flink_compute_engine.py
• scoped mypy for the Flink files/test with --follow-imports=skip --check-untyped-defs

All passed locally.

Follow-up for the registration check failures: both failed during pixi install -e registration-tests --locked with lock-file not up-to-date with the workspace.

Refreshed pixi.lock in 2aae600 and verified locally with the same CI setup command:

• /tmp/pixi-v0.63.1/pixi install -e registration-tests --locked --manifest-path pyproject.toml

That command now passes locally.

Follow-up on the lint concern: local full lint found mypy errors in flink/feature_builder.py after the Copilot fixes. Addressed them in 37accc6 by typing the current DAG node as the base DAGNode and widening FlinkOutputNode to accept FeatureView like the other compute-engine write nodes.

Verified locally:

• make lint-python passed end-to-end (ruff check, ruff format --check, full mypy feast).
• uv run python -m pytest sdk/python/tests/unit/infra/compute_engines/flink/test_flink_compute_engine.py -q passed.
• Scoped Flink mypy/ruff also passed.

The pull_request workflow checks shown as failing/action-required in the UI still need maintainer approval to run on the fork; they have not actually executed on the latest commit yet.

Overall looks good to me.
Thanks for the contribution.
Might be better to refactor some functions that can be used across engines by pyspark stream as well in the future

Thanks, agreed. I added a narrow refactor in c99bc7f to centralize the compute-engine entity timestamp alias, entity row id, and timestamp-column resolution helpers in feast.infra.compute_engines.utils. Flink now uses those helpers, and Spark/Local/Ray use the shared timestamp alias/resolver where applicable, so future PySpark streaming work can reuse the same Feast point-in-time join semantics without duplicating them again.

Verified locally:

• uv run python -m pytest sdk/python/tests/unit/infra/compute_engines/flink/test_flink_compute_engine.py -q
• make lint-python

@shuchu I addressed the AI review/lint fallout and pushed 6ba7a5305.

What changed:

• Added the missing flink expectations to the transformation factory/mode unit tests.
• Confirmed the prior unit failures were from those missing expectations; local Spark-related failures required Java 17 locally, not code changes.

Local verification:

• JAVA_HOME=/usr/lib/jvm/jdk-17.0.12-oracle-x64 PATH=/usr/lib/jvm/jdk-17.0.12-oracle-x64/bin:$PATH make test-python-unit -> 1934 passed, 15 skipped
• make lint-python -> passed
• uv run python -c "from feast.cli import cli" -> passed

GitHub checks that ran on the latest commit are green:

• DCO
• GitBook / docs.feast.dev
• pr-registration-integration-tests
• pr-integration-tests (integration-test-python and mcp-feature-server-runtime)

The remaining pull_request workflows (unit-tests, linter, DuckDB/Ray/local integration, REST API, security, etc.) are currently action_required for the new fork commit, not failed. Could you please approve/run the workflows again when you get a chance?

@shuchu I pushed 4bf1433 to fix the macOS 3.11 unit failure.

The failure was isolated to the feature server performance microbenchmark: CI measured 1.44x against a hard 1.5x threshold. The patch keeps 1.5x on non-macOS and uses 1.3x on macOS to avoid runner timing flakiness while still checking for a meaningful speedup.

Local checks passed:

• Targeted failed test
• Full test_feature_server_utils.py file: 55 passed
• Ruff check and format check

Could you please approve/re-run CI for the new fork commit?

approved and rerun now

@HaoXuAI I pushed 7aebf78 to fix the new unit failure.

The failure was in test_image_utils.py: the unit tests were loading a real resnet model from the model hub. I changed the tests to mock timm model loading with a deterministic dummy model, so they no longer depend on network/cache state.

Local checks passed:

• pytest sdk/python/tests/unit/test_image_utils.py -q: 16 passed
• ruff check and format check on the file

Could you please approve/re-run CI for the new fork commit?

@HaoXuAI I pushed ebbc86b to harden the stale remote RBAC TLS failure.

The failed old run was on the previous SHA and hit RemoteOfflineTlsStore with an Arrow Flight SSL handshake wrong-version error against 0.0.0.0:. The test helpers were binding servers to 0.0.0.0 and also using 0.0.0.0 as the client/config host. This patch keeps the server bind address as 0.0.0.0 but uses localhost for client URLs, readiness checks, and remote offline store configs.

Local checks passed:

• ruff check sdk/python/tests/universal/feature_repos/universal/data_sources/file.py
• ruff format --check sdk/python/tests/universal/feature_repos/universal/data_sources/file.py
• git diff --check

Could you please approve/re-run CI for the new fork commit?

@HaoXuAI I pushed 71f5728fd to fix the latest macOS 3.12 unit failure.

The failing test was the feature server serializer microbenchmark again. Instead of lowering the threshold further, this patch changes the benchmark from wall-clock timing to process CPU timing, which removes GitHub-hosted macOS/xdist scheduler noise while restoring the original 1.5x speedup threshold for every OS.

Local verification on Python 3.12 CI deps passed:

• uv run python -m pytest sdk/python/tests/unit/test_feature_server_utils.py::TestPerformance::test_faster_than_message_to_dict -q -s
• uv run python -m pytest sdk/python/tests/unit/test_feature_server_utils.py -q
• uv run python -m pytest -n 8 sdk/python/tests/unit/test_feature_server_utils.py -q
• uv run ruff check sdk/python/tests/unit/test_feature_server_utils.py
• uv run ruff format --check sdk/python/tests/unit/test_feature_server_utils.py

Could you please approve/re-run CI for the new fork commit?

@HaoXuAI I pushed 77463caf4 to address the latest remote-rbac-integration-tests-python failure.

The failed GitHub job had 167 passed and only errored during teardown of test_historical_features_with_no_ttl for RemoteOfflineOidcAuthStore: the remote offline server port never closed after proc.kill(). The job cleanup also showed many orphan feast processes.

The patch starts feast serve_offline test servers in their own process group, sends stdout/stderr to DEVNULL, and tears them down by terminating/killing the whole process group and waiting for the process before checking the port closure.

Local verification passed:

• uv run ruff check sdk/python/tests/universal/feature_repos/universal/data_sources/file.py
• uv run ruff format --check sdk/python/tests/universal/feature_repos/universal/data_sources/file.py
• scoped mypy for the touched file
• exact OIDC failed case: 2 passed
• CI-shaped -n 8 --dist loadgroup run for plain/TLS/OIDC test_historical_features_with_no_ttl: 6 passed

Could you please approve/re-run CI for the new fork commit?

@HaoXuAI @franciscojavierarceo could you help us understand the current CI state? We have been fixing the failures as they appear, but CI keeps surfacing different failures after reruns and the latest visible failure does not look related to the Flink compute engine changes.

Current failing check: integration-test-python (3.11, ubuntu-latest) on run 27094711100 / job 79964631819. The job ended with 306 passed, 50 skipped, 1 error. The error is from sdk/python/tests/integration/online_store/test_mongodb_vector_search.py::TestMongoDBVectorSearch::test_write_and_retrieve_round_trip, where PyMongo raises OperationFailure while collection.list_search_indexes() is connecting to the MongoDB Atlas Local Search Index Management service.

I have not patched MongoDB/vector-search code in this Flink PR because that seems outside the scope of this contribution. Could you advise whether this should be rerun/treated as an infra flake, or whether there is a preferred separate fix for that test?

thanks @XuananLe! can you resolve the conflicts? Once CI is green let's get this shipped for the next release!

@franciscojavierarceo conflicts are resolved now. I pushed 9a71db163, merging the latest origin/master into the PR branch and regenerating pixi.lock with Pixi 0.63.1.

Local checks run:

• uv run ruff check on the Flink-related Python files passed
• uv run python -m pytest sdk/python/tests/unit/infra/compute_engines/flink/test_flink_compute_engine.py -q passed (21 passed)
• pixi lock --check passed

GitHub now shows the PR as no longer dirty; CI is running again.

@franciscojavierarceo I pushed a66fe10d8 addressing the two Flink blockers:

• Empty pandas entity_df is now treated as a supplied entity dataframe, so historical retrieval keeps the entity join and returns an empty result instead of feature rows. Added unit coverage: test_flink_historical_retrieval_with_empty_entity_df_returns_empty_result.
• FlinkSourceReadNode no longer requires a nonstandard RetrievalJob.to_flink_table(table_env). It still uses that native path if present, but now falls back to the standard production RetrievalJob.to_arrow() path and converts the Arrow table to a Flink table. Added unit coverage: test_flink_source_read_node_converts_arrow_retrieval_jobs.

I also merged latest origin/master, so the branch is up to date. Local validation passed:

• uv run ruff check sdk/python/feast/infra/compute_engines/flink/feature_builder.py sdk/python/feast/infra/compute_engines/flink/nodes.py sdk/python/tests/unit/infra/compute_engines/flink/test_flink_compute_engine.py
• uv run ruff format --check ...
• uv run python -m pytest sdk/python/tests/unit/infra/compute_engines/flink/test_flink_compute_engine.py -q (22 passed)

CI is running again on the new head.

@franciscojavierarceo all CI is green now: 30/30 checks passed, and the branch has no merge conflicts. Could you please take another look when you have a chance?