fix: UDF/ODFV source rehydrate (+ Postgres / online cache)

[aniketpalu]

Summary

Fixes three issues found while testing BatchFeatureView UDFs and OnDemandFeatureViews with Spark transformations (SparkApplication / SparkComputeEngine path) against a remote feature server.

All three blocked a clean UDF/ODFV materialize → online path. This PR is scoped to that story. Shared helper feast.transformation.udf_rehydrate is introduced so the same contract can later be wired into RayTransformation.from_proto (intentionally not changed here).

1. Spark BFV UDF — prefer source over dill (exit 139)

Problem: Materializing a BatchFeatureView with a Python UDF on Spark could segfault the driver (exit 139). Registry stores both dill (body) and source (body_text / udf_string); the Spark path executed the dill-restored callable, which is unsafe with DataFrame ops on our Spark stack.

Change: SparkTransformationNode resolves the UDF via shared resolve_udf: exec trusted udf_string when present; fall back to the existing callable / dill. SparkFeatureBuilder passes udf_string through.

Files: spark/nodes.py, spark/feature_builder.py, transformation/udf_rehydrate.py

2. ODFV / local transformations — source-first from_proto

Problem: ODFV serve (and feast-apply rehydrate from SQL registry) used dill.loads(body) only. That fails across Python minors (e.g. apply on Spark driver 3.10, serve on feature-server 3.12) and when body_text includes @on_demand_feature_view(...) (NameError / CrashLoop if source is exec'd naively).

Change: PandasTransformation.from_proto and PythonTransformation.from_proto use the same resolve_udf (strip leading decorators, then exec; dill fallback).

Out of scope (follow-up): RayTransformation.from_proto remains dill.loads for now.

Files: pandas_transformation.py, python_transformation.py, udf_rehydrate.py

3. Postgres empty feature_name_columns → SELECT *

Problem: BatchFeatureView python/pandas mode signals "all columns" with empty feature_name_columns. Postgres offline store still projected only join keys + timestamps, starving the UDF of input features.

Change: When feature_name_columns is empty, pull_latest / pull_all use a.* / alias.* (aligned with Ray offline-store semantics).

Files: postgres_offline_store/postgres.py + unit tests

4. Fresh online registry reads

Problem: After materialize completes, online serving could briefly (or stickily) gate on cached MATERIALIZING / stale FV metadata.

Change: Online request path uses allow_cache=False when resolving features / feature services (utils._get_online_request_context, feature server _get_features).

Files: utils.py, feature_server.py

Out of scope

• RayTransformation.from_proto / Ray compute-engine dill paths (follow-up using udf_rehydrate)
• Remote materialization async failure → GENERATED wipe (separate remote-mat PR)
• Redis same-timestamp idempotency

Which issue(s) this PR fixes

Found during BYOS Spark UDF/ODFV E2E; no single upstream issue required. Related discussions: Spark UDF materialize reliability, ODFV cross-image Python skew.

Checks

• I've made sure the tests are passing.
• My commits are signed off (git commit -s)
• My PR title follows conventional commits format

Testing Strategy

• Unit tests
• Integration tests
• Manual tests
• Testing is not required for this change

Misc

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 66.43357% with 48 lines in your changes missing coverage. Please review.
✅ Project coverage is 46.42%. Comparing base (6019bd6) to head (fd2178d).
⚠️ Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
sdk/python/feast/transformation/udf_rehydrate.py	64.65%	27 Missing and 14 partials ⚠️
.../python/feast/infra/compute_engines/spark/nodes.py	50.00%	3 Missing ⚠️
...ast/infra/compute_engines/spark/feature_builder.py	0.00%	2 Missing ⚠️
..._stores/contrib/postgres_offline_store/postgres.py	84.61%	1 Missing and 1 partial ⚠️
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6655      +/-   ##
==========================================
+ Coverage   46.36%   46.42%   +0.05%     
==========================================
  Files         414      415       +1     
  Lines       50072    50206     +134     
  Branches     7154     7190      +36     
==========================================
+ Hits        23218    23307      +89     
- Misses      25231    25262      +31     
- Partials     1623     1637      +14     

Flag	Coverage Δ
go-feature-server	30.58% <ø> (ø)
python-unit	47.72% <66.43%> (+0.05%)	⬆️

Files with missing lines	Coverage Δ
...thon/feast/transformation/pandas_transformation.py	73.68% <100.00%> (+0.95%)	⬆️
...thon/feast/transformation/python_transformation.py	87.50% <100.00%> (+0.46%)	⬆️
...ast/infra/compute_engines/spark/feature_builder.py	0.00% <0.00%> (ø)
..._stores/contrib/postgres_offline_store/postgres.py	46.13% <84.61%> (+1.05%)	⬆️
.../python/feast/infra/compute_engines/spark/nodes.py	26.36% <50.00%> (+0.29%)	⬆️
sdk/python/feast/transformation/udf_rehydrate.py	64.65% <64.65%> (ø)

---

Continue to review full report in Codecov by Harness.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6019bd6...fd2178d. Read the comment docs.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[franciscojavierarceo]

Two blocking concerns from the remote diff:

1. feature_server.py and _get_online_request_context() now pass allow_cache=False for feature-service and feature-view resolution on every online request. For SQL-backed registries this moves backing-registry I/O into the serving hot path, which can materially regress latency and registry load. Please refresh/invalidate the affected cached definitions when UDF state changes (or isolate and justify this serving behavior separately) instead of globally bypassing the cache.
2. CodeQL's open finding in transformation/udf_rehydrate.py is valid: the decorator-stripping regex has nested unbounded repetition and processes registry-provided source text. Please replace it with a linear parser/line-based approach and add an adversarial regression case.

[ntkathole]

@aniketpalu Please fix CI