feat: Feast-MLflow Integration - #6235
Conversation
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
| with tempfile.TemporaryDirectory() as tmp_dir: | ||
| path = os.path.join(tmp_dir, "entity_df.parquet") | ||
| entity_df.to_parquet(path, index=False) | ||
| mlflow.log_artifact(path) |
There was a problem hiding this comment.
🟡 Mixing global mlflow.log_artifact() with explicit-URI client causes artifact/metadata to target different servers
In _auto_log_entity_df_info, tags and params are logged via a locally-created MlflowClient(tracking_uri=tracking_uri) (lines 304, 309-318), but the entity DataFrame artifact is uploaded via the global mlflow.log_artifact(path) (line 327). The global function uses the tracking URI set by mlflow.set_tracking_uri(), not the explicit tracking_uri from the config. If the global tracking URI is changed by another library in the process, or if _init_mlflow_tracking failed silently (caught by except Exception at sdk/python/feast/feature_store.py:249), the artifact would be uploaded to a different server than where the tags/params were logged, splitting metadata across two servers.
Was this helpful? React with 👍 or 👎 to provide feedback.
Sorry, something went wrong.
| try: | ||
| import mlflow | ||
|
|
||
| tracking_uri = mlflow_cfg.tracking_uri or "http://127.0.0.1:5000" |
There was a problem hiding this comment.
🔴 UI server ignores MLFLOW_TRACKING_URI env var, falls back to hardcoded localhost
In ui_server.py, both the /api/mlflow-runs and /api/mlflow-feature-models endpoints resolve the MLflow tracking URI using mlflow_cfg.tracking_uri or "http://127.0.0.1:5000". This reads the raw tracking_uri field from the config and falls back to a hardcoded localhost URL, completely bypassing the MLFLOW_TRACKING_URI environment variable. In contrast, the rest of the codebase (feature_store.py:243, feature_store.py:325, feature_store.py:1696, feature_store.py:2885) correctly calls mlflow_cfg.get_tracking_uri() which checks the env var via sdk/python/feast/mlflow_integration/config.py:19-29. When a user sets MLFLOW_TRACKING_URI without setting tracking_uri in YAML (which is a very common deployment pattern, and documented in the PR's own docs at docs/reference/mlflow.md:51), the UI endpoints will incorrectly connect to http://127.0.0.1:5000 instead of the env-var-specified server.
| tracking_uri = mlflow_cfg.tracking_uri or "http://127.0.0.1:5000" | |
| tracking_uri = mlflow_cfg.get_tracking_uri() or "http://127.0.0.1:5000" |
Was this helpful? React with 👍 or 👎 to provide feedback.
Sorry, something went wrong.
| try: | ||
| import mlflow | ||
|
|
||
| tracking_uri = mlflow_cfg.tracking_uri or "http://127.0.0.1:5000" |
There was a problem hiding this comment.
🔴 Second UI endpoint also ignores MLFLOW_TRACKING_URI env var
Same issue as in the /api/mlflow-runs endpoint: the /api/mlflow-feature-models endpoint at sdk/python/feast/ui_server.py:234 uses mlflow_cfg.tracking_uri or "http://127.0.0.1:5000" instead of mlflow_cfg.get_tracking_uri(), causing the MLFLOW_TRACKING_URI environment variable to be ignored.
Was this helpful? React with 👍 or 👎 to provide feedback.
Sorry, something went wrong.
| ) | ||
|
|
||
| # Emit MLflow event for materialization (Phase 7) | ||
| _mat_duration = time.monotonic() - _retrieval_start if '_retrieval_start' in dir() else 0 |
There was a problem hiding this comment.
🟡 Undefined _retrieval_start variable in materialize_incremental causes bogus duration or NameError
In materialize_incremental, the code at line 2115 references _retrieval_start which is never defined in that method scope. The expression '_retrieval_start' in dir() checks local variable names but dir() without arguments returns module-level names, not local variables -- so the check is unreliable. If _retrieval_start is not in the returned list, _mat_duration defaults to 0 (acceptable but meaningless). However, if it IS in the name list (e.g., a module-level _retrieval_start existed from another context), it would attempt time.monotonic() - _retrieval_start on a potentially unrelated value, resulting in a wrong duration or a NameError/TypeError at runtime.
| _mat_duration = time.monotonic() - _retrieval_start if '_retrieval_start' in dir() else 0 | |
| _mat_duration = 0 |
Was this helpful? React with 👍 or 👎 to provide feedback.
Sorry, something went wrong.
| if tracking_uri: | ||
| mlflow.set_tracking_uri(tracking_uri) | ||
|
|
||
| experiment_name = f"{project}{ops_experiment_suffix}" | ||
| mlflow.set_experiment(experiment_name) |
There was a problem hiding this comment.
🔴 log_apply_to_mlflow and log_materialize_to_mlflow mutate global MLflow tracking URI and experiment, corrupting concurrent user runs
Both log_apply_to_mlflow (sdk/python/feast/mlflow_integration/logger.py:185-189) and log_materialize_to_mlflow (sdk/python/feast/mlflow_integration/logger.py:251-255) call mlflow.set_tracking_uri(tracking_uri) and mlflow.set_experiment(experiment_name) globally. These mutate process-wide global state. If a user has an active MLflow run in a different experiment (e.g. during feast apply inside a training script), these calls redirect subsequent MLflow operations to the ops experiment. While the functions try to restore mlflow.set_experiment(project) afterward (line 224, 274), the tracking URI is never restored. In the exception path, if mlflow.start_run() raises (line 202/258), the experiment remains set to the ops experiment. This corrupts any concurrent or subsequent user operations in the same process.
Prompt for agentsThe functions log_apply_to_mlflow and log_materialize_to_mlflow in sdk/python/feast/mlflow_integration/logger.py mutate process-wide global state by calling mlflow.set_tracking_uri() and mlflow.set_experiment(). This corrupts any concurrent user MLflow runs in the same process. The fix should save and restore the original tracking URI and experiment before/after the ops logging, or preferably use the MlflowClient API exclusively (which already takes tracking_uri as a parameter) instead of the global mlflow module-level functions. The MlflowClient.create_run() method allows creating runs in specific experiments without mutating global state. This pattern is already used in log_feature_retrieval_to_mlflow which correctly uses client = mlflow.MlflowClient(tracking_uri=tracking_uri) without touching global state.
Was this helpful? React with 👍 or 👎 to provide feedback.
Sorry, something went wrong.
| def mlflow(self): | ||
| """Get the MLflow configuration.""" | ||
| if not self._mlflow: | ||
| if isinstance(self.mlflow_config, Dict): | ||
| from feast.mlflow_integration.config import MlflowConfig | ||
|
|
||
| self._mlflow = MlflowConfig(**self.mlflow_config) | ||
| elif self.mlflow_config: | ||
| self._mlflow = self.mlflow_config | ||
| return self._mlflow |
There was a problem hiding this comment.
🟡 mlflow property returns falsy None on subsequent calls when MlflowConfig(enabled=False) is set
The mlflow property in RepoConfig (sdk/python/feast/repo_config.py:497-506) uses if not self._mlflow: as the guard. When the config is a valid MlflowConfig object with enabled=False, self._mlflow will be a truthy MlflowConfig instance (Pydantic models are truthy), so this works correctly on first access. However, if mlflow_config is None (the default), self._mlflow stays None forever and the property returns None -- which is fine. The real issue is: if mlflow_config is an empty dict {}, MlflowConfig(**{}) creates a valid config with enabled=False, and self._mlflow becomes truthy. But if mlflow_config is set to some falsy-like Pydantic object (unlikely but possible with custom subclasses), the if not self._mlflow guard would re-create it every time. This is the same pattern as openlineage property so it's consistent, but not actually a bug for standard usage.
Was this helpful? React with 👍 or 👎 to provide feedback.
Sorry, something went wrong.
| - **Online feature retrieval** -- `get_online_features()` tags the run with the same metadata | ||
| - **Entity DataFrame archival** -- optionally saves the training entity DataFrame as an MLflow artifact for full reproducibility | ||
| - **Execution context tagging** -- tags runs with where they ran (workbench, KFP pipeline, feature server, or standalone) | ||
| - **Operation logging** -- optionally logs `feast apply` and `feast materialize` to a separate MLflow experiment |
There was a problem hiding this comment.
not sure this is actually useful
Sorry, something went wrong.
There was a problem hiding this comment.
actually, this is not updated one, thought of this one for backtracking from mllfow run to the workbench experiment but now it is omitted
Sorry, something went wrong.
| - **Entity DataFrame archival** -- optionally saves the training entity DataFrame as an MLflow artifact for full reproducibility | ||
| - **Execution context tagging** -- tags runs with where they ran (workbench, KFP pipeline, feature server, or standalone) | ||
| - **Operation logging** -- optionally logs `feast apply` and `feast materialize` to a separate MLflow experiment | ||
| - **Model-to-Feature resolution** -- map any MLflow model URI back to its Feast feature service |
There was a problem hiding this comment.
nice
Sorry, something went wrong.
| - **Execution context tagging** -- tags runs with where they ran (workbench, KFP pipeline, feature server, or standalone) | ||
| - **Operation logging** -- optionally logs `feast apply` and `feast materialize` to a separate MLflow experiment | ||
| - **Model-to-Feature resolution** -- map any MLflow model URI back to its Feast feature service | ||
| - **Training reproducibility** -- reconstruct the exact entity DataFrame from a past MLflow run |
There was a problem hiding this comment.
👏
Sorry, something went wrong.
| - **Operation logging** -- optionally logs `feast apply` and `feast materialize` to a separate MLflow experiment | ||
| - **Model-to-Feature resolution** -- map any MLflow model URI back to its Feast feature service | ||
| - **Training reproducibility** -- reconstruct the exact entity DataFrame from a past MLflow run | ||
| - **Training-to-prediction linkage** -- `FeastMlflowClient.load_model()` links prediction runs back to their training runs |
There was a problem hiding this comment.
very nice
Sorry, something went wrong.
| - **Model-to-Feature resolution** -- map any MLflow model URI back to its Feast feature service | ||
| - **Training reproducibility** -- reconstruct the exact entity DataFrame from a past MLflow run | ||
| - **Training-to-prediction linkage** -- `FeastMlflowClient.load_model()` links prediction runs back to their training runs | ||
| - **Feast MLflow Client** -- a thin wrapper that eliminates direct `import mlflow` in user code |
There was a problem hiding this comment.
makes sense
Sorry, something went wrong.
| _consecutive_failures = 0 | ||
|
|
||
|
|
||
| def log_feature_retrieval_to_mlflow( |
There was a problem hiding this comment.
i guess if we have this enabled why do we need to even have the wrapper? can't we full handle this for the user?
Sorry, something went wrong.
There was a problem hiding this comment.
I guess my question is, can't we just hide all of the mlflow client usage to the user?
Sorry, something went wrong.
There was a problem hiding this comment.
the auto-logging in logger.py fully handled for users who already use mlflow.start_run() or client.start_run() considering our FeastMlflowClient wrapper which is for users who want to avoid import mlflow entirely and get additional Feast-specific behavior on top ex: logging a model, registering it, loading it for inference etc soFeastMlflowClient is for users who want the extra lineage features without touching MLflow directly
Sorry, something went wrong.
There was a problem hiding this comment.
I would rather push all these functions to integration client as everything here is a part of mlflow feast integration?
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
jyejare
left a comment
There was a problem hiding this comment.
Comments inline:
- We also need to enable mlflow configuratins from feast operator else admin has to do it explicitly.
- Also, we need to remove duplicate mechanisms for initiating/using mlflow client and we should provide a clean way that is store.mlflow whenever possible.
Sorry, something went wrong.
| 2. `MLFLOW_TRACKING_URI` environment variable | ||
| 3. MLflow's default (`./mlruns` local directory) | ||
|
|
||
| This means you can omit `tracking_uri` from the YAML and set `MLFLOW_TRACKING_URI` in your environment instead. |
There was a problem hiding this comment.
| This means you can omit `tracking_uri` from the YAML and set `MLFLOW_TRACKING_URI` in your environment instead. | |
| This means you can omit `tracking_uri` from the YAML and set `MLFLOW_TRACKING_URI` in your environment instead or it would be pulled from /.mlruns automatically when both are not set. |
wdyt ?
Sorry, something went wrong.
|
|
||
| | Artifact | Description | | ||
| |----------|-------------| | ||
| | `required_features.json` | JSON list of feature references the model was trained on | |
There was a problem hiding this comment.
Could we rename to relate it with feast better and so that it would be self describing in MLFLow ecosystem, something like:
required_feast_features.json
OR
feast_features.json
Sorry, something went wrong.
| With the configuration above, feature metadata is logged whenever there is an active MLflow run: | ||
|
|
||
| ```python | ||
| import mlflow |
There was a problem hiding this comment.
I think explicit import is no more required as feast module gives you the ability to:
store.mlflow.start_run(....
Sorry, something went wrong.
| from feast import FeatureStore | ||
|
|
||
| store = FeatureStore(".") | ||
| client = store.get_mlflow_client() |
There was a problem hiding this comment.
Even this is changed now. Lets keep only store.mlflow. examples.
Sorry, something went wrong.
| | Condition | Error | | ||
| |-----------|-------| | ||
| | No `feature_store.yaml` in cwd and no store created | `RuntimeError` with guidance to call `feast.mlflow.init(store)` | | ||
| | `mlflow.enabled` is `false` or missing | `RuntimeError` with guidance to set `mlflow.enabled=true` | |
There was a problem hiding this comment.
Even if user wants to set it to false ?
Sorry, something went wrong.
| invalid, resolution fails, or validation against the store fails. | ||
| """ | ||
| try: | ||
| import mlflow |
There was a problem hiding this comment.
Same here. Create a class and initiate in integration object/class that takes store object as param.
Sorry, something went wrong.
| if mlflow_cfg is None or not mlflow_cfg.enabled: | ||
| return | ||
|
|
||
| import mlflow |
There was a problem hiding this comment.
The store object is available to you here as self, so why not :
self.mlflow.set_tracking_uri(....
Sorry, something went wrong.
| client = MlflowClient(tracking_uri=tracking_uri) | ||
| refs = ["z_view:f1", "a_view:f2", "z_view:f3", "a_view:f4"] | ||
|
|
||
| with mlflow.start_run() as run: |
There was a problem hiding this comment.
The tests should also be focusing only on store.mlflow... culture.
Sorry, something went wrong.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
This comment was marked as resolved.
| ops_suffix = mlflow_cfg.ops_experiment_suffix | ||
|
|
||
| effective_uri = self._tracking_uri or self._mlflow.get_tracking_uri() | ||
| client = self._mlflow.MlflowClient(tracking_uri=effective_uri) |
There was a problem hiding this comment.
instead of creating new client everytime, can we initialise once and reuse ?
Sorry, something went wrong.
| # Initialize feature service cache for performance optimization | ||
| self._feature_service_cache = {} | ||
|
|
||
| # Cache for _resolve_feature_service_name lookups |
There was a problem hiding this comment.
the MLflow auto-log could resolve to a deleted or renamed feature service. _fs_name_cache and _fs_name_index caches (used by _resolve_feature_service_name) are never cleared when refresh_registry() or _clear_feature_service_cache() is called
Sorry, something went wrong.
| self._fs_name_cache = {} | ||
| self._fs_name_index_ts = time.monotonic() | ||
|
|
||
| def _resolve_feature_service_name(self, feature_refs: List[str]) -> Optional[str]: |
There was a problem hiding this comment.
can't we use existing _feature_service_cache from FeatureStore?
Sorry, something went wrong.
There was a problem hiding this comment.
_feature_service_cache lookup is : service_name → [feature_refs]and is populated lazily on demand
_fs_name_index lookup is reverse {feature_refs} → service_name, populated from all registry services needed for subset matching when users pass raw refs instead of a FeatureService
Sorry, something went wrong.
|
|
||
| try: | ||
| if self.mlflow is not None and self.config.mlflow.auto_log: | ||
| import asyncio |
There was a problem hiding this comment.
already imported at top
Sorry, something went wrong.
| self._fs_name_index_ts: float = -self._FS_NAME_INDEX_TTL_SECONDS | ||
|
|
||
| self._mlflow_client: Optional[Any] = None | ||
| self._init_mlflow() |
There was a problem hiding this comment.
I think better to do lazy loading here, first check the config and then load module.
Sorry, something went wrong.
| self._default_experiment = store.config.project | ||
|
|
||
| if self._tracking_uri: | ||
| _mlflow_mod.set_tracking_uri(self._tracking_uri) |
There was a problem hiding this comment.
remove set_tracking_uri() from __init__ and instead pass tracking_uri explicitly when needed (the MlflowClient already receives it)
Sorry, something went wrong.
| ) | ||
|
|
||
| @app.get("/api/mlflow-runs") | ||
| def get_mlflow_runs(max_results: int = 50): |
There was a problem hiding this comment.
should this be a registry API? Not a blocker for this PR, but architecture wise, registry is source of truth
Sorry, something went wrong.
There was a problem hiding this comment.
Yes, will add this one will open to other benefits too !
Sorry, something went wrong.
There was a problem hiding this comment.
@Vperiodt going to handle this in this PR or different one?
Sorry, something went wrong.
There was a problem hiding this comment.
yess, will address this in a follow up PR
Sorry, something went wrong.
|
linter is failing |
Sorry, something went wrong.
fixed |
Sorry, something went wrong.
There was a problem hiding this comment.
include mlflow here in ci dependencies
Sorry, something went wrong.
| <EuiFlexItem> | ||
| <EuiStat title={`${numOfFs}`} description="Consuming Services" /> | ||
| </EuiFlexItem> | ||
| <EuiFlexItem> |
There was a problem hiding this comment.
This should be conditionally rendered, users who don't use MLflow will wonder what these stats mean and why they're there
Sorry, something went wrong.
| const usage: FeatureUsageEntry | undefined = | ||
| data.feature_usage?.[featureViewName]; | ||
|
|
||
| if (!usage || usage.run_count === 0) { |
There was a problem hiding this comment.
same here, if mlflow is not configured, this needs to be hidden
Sorry, something went wrong.
|
@Vperiodt Guard the UI components behind a check on whether MLflow data is available. The simplest approach is to have the backend return a mlflow_enabled: false flag in the API responses (or a new lightweight /api/mlflow-status endpoint), and conditionally render MLflow components only when enabled. |
Sorry, something went wrong.
hello ! @ntkathole, Addressed all reviews |
Sorry, something went wrong.
|
@Vperiodt need to update pixi.lock file, also please rebase |
Sorry, something went wrong.
Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED refine MLflow integration, UI runs endpoint, and lockfiles Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED fix issues Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED handle fallback Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED added model to features table Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED address CI Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED add docs Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED fix-lint Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED update docs Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED update-logic Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED add-integration-test Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED mlflowclient wrapper Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED address-reviews Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED update docs Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED fix-lint Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED fix-auto-log Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED update feastMlflowClient Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED changing init Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED fixing docs Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED remove doc Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED update format Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED fix-test Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED update docs Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED addressed reviews Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED fix: update .secrets.baseline to remove pragma-allowlisted entry Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED ui-fallback Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED update requirements Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED
Signed-off-by: Vanshika Vanshika <vvanshik@redhat.com> rh-pre-commit.version: 2.3.2 rh-pre-commit.check-secrets: ENABLED
|
lgtm |
Sorry, something went wrong.
# [0.64.0](v0.63.0...v0.64.0) (2026-06-13) ### Bug Fixes * Add async_supported property to RedisOnlineStore ([9b088fe](9b088fe)) * Add missing feast init templates to operator CRD and enhance persistence documentation ([1941d4d](1941d4d)) * Allow to publish from reference branch ([5458ec8](5458ec8)) * API calls list ([4203eb7](4203eb7)) * **bigquery:** Enable list inference for parquet loads in offline_write_batch ([9243497](9243497)), closes [#5845](#5845) * Bump grpcio dependencies ([07b4782](07b4782)) * **compute-engine/local:** Honor field_mapping on join keys in dedup + join nodes ([#6395](#6395)) ([bd01824](bd01824)) * **dynamodb:** Avoid tag race condition by using diff-based tag updates ([#6479](#6479)) ([bad2b7d](bad2b7d)), closes [#6418](#6418) * **dynamodb:** Fix mypy type for _build_projection_expression return ([217b4da](217b4da)) * Fix intermittent async test failures for DynamoDB and Redis ([63c5eb1](63c5eb1)) * Fix mongodb blog title ([57d28d4](57d28d4)) * Fix shared SQL registry crash - avoid unnecessary UDF deserialization in proto cache building ([ac588d7](ac588d7)) * Fix SparkRetrievalJob.persist() failing for SparkSource ([209d7cd](209d7cd)) * Fixed formatting and image for mongo blog ([#6377](#6377)) ([f8389fb](f8389fb)) * Fixes for ray source ([7f592a4](7f592a4)) * **go:** skip registry refresh when cache_ttl_seconds <= 0 ([97ed40c](97ed40c)) * Handle array of strings columns in Athena materialization ([#6324](#6324)) ([4ed0278](4ed0278)) * make milvus VARCHAR max_length configurable, remove hardcoded 512 limit ([3b98c22](3b98c22)) * **operator:** Set appProtocol: grpc on registry gRPC Service ([#6367](#6367)) ([c9ae2b4](c9ae2b4)) * PyJWT 2.10+ added validation that rejects empty HMAC keys ([e756ffe](e756ffe)) * RemoteOnlineStore sends all features in a single HTTP request ([8f187dd](8f187dd)) * Remove registry proto dump to enforce RBAC and add permission checks to Commit/Refresh RPCs ([328431f](328431f)) * Remove selector migration job - no longer needed ([51c325e](51c325e)) * replace broken .claude skill symlink with correct relative path ([4541690](4541690)) * Replace selector label strip patch with migration Job for upgrade-safe selector uniqueness ([00dea50](00dea50)) * Scope feature view name conflict check to current project in file-based registry ([#6369](#6369)) ([a4fde83](a4fde83)), closes [#6209](#6209) * **snowflake:** Stop double-quoting connection identifiers ([#6462](#6462)) ([e914d59](e914d59)) * **spark:** S3/GCS PyArrow filesystem resolution for staging paths ([#6442](#6442)) ([ae50414](ae50414)) * **trino:** Clean up temporary entity tables after retrieval ([#6381](#6381)) ([d86b13d](d86b13d)), closes [#6306](#6306) * Update go-feature-server base image to Go 1.25 and fix operator Dockerfile COPY permissions ([86ef0bc](86ef0bc)) ### Features * [Backend] Data Quality Monitoring with native compute, multi-backend support, REST API, CLI ([#6202](#6202)) ([5458c37](5458c37)) * Add apache flink compute engine ([#6476](#6476)) ([9636d6a](9636d6a)) * Add demo noteboooks for users ([e362173](e362173)) * Add enabled/disabled toggle for feature views ([#6401](#6401)) ([5f1fa0d](5f1fa0d)), closes [#6395](#6395) * Add Label View to init template ([ec272d5](ec272d5)) * Add mTLS support to remote registry gRPC client ([#6474](#6474)) ([c9602d8](c9602d8)) * Add Prometheus gauges for FeatureStore installation telemetry ([#6354](#6354)) ([1b681b7](1b681b7)) * Adds registry REST API endpoints for managing entities, data sources, and feature views ([#6413](#6413)) ([f77bd1d](f77bd1d)) * Allow CRUD on entities, data sources, and feature views from UI ([#6412](#6412)) ([2321c07](2321c07)) * Allow default openlineage configuration ([#6467](#6467)) ([276b6df](276b6df)) * **bigquery:** Support DATE-type event timestamp columns ([#6362](#6362)) ([753dee5](753dee5)), closes [#2530](#2530) * **cli:** Add `feast projects delete` command (closes [#5095](#5095)) ([#6318](#6318)) ([1a4b96c](1a4b96c)) * Data Quality Monitoring added in feast UI ([#6422](#6422)) ([fa271be](fa271be)) * **dynamodb:** Use ProjectionExpression when requested_features is set ([0adc906](0adc906)), closes [#6058](#6058) * Enhance DataSource and FeatureView modals with error handling and submission states ([96d7169](96d7169)) * Expose registry endpoints on feature server for MCP access ([f77981c](f77981c)) * Feast First-Class LabelView Implementation ([#6292](#6292)) ([c0e7e5d](c0e7e5d)) * Feast-MLflow Integration ([#6235](#6235)) ([7279c75](7279c75)) * Operational metrics for offline store and SOX metrics for both ([#6340](#6340)) ([65b1b80](65b1b80)) * Pre-compute feature service ([8011550](8011550)) * REST API-backed UI for RBAC compatibility and per-page lazy loading ([#6414](#6414)) ([6ae80af](6ae80af)) * Support non-string map key types ([#6382](#6382)) ([#6383](#6383)) ([728aa2e](728aa2e)) * Update FeatureStore CRD with DRA Fields ([01241e4](01241e4)) ### Performance Improvements * Cache feature view resolution in get_online_features to reduce per-request overhead ([55c2f18](55c2f18)) * Optimize feature serving latency with batched async Redis, cached checks fix ([103809a](103809a)) * Replace MessageToDict with optimized custom dict builder ([#6015](#6015)) ([9902064](9902064))
# [0.64.0](feast-dev/feast@v0.63.0...v0.64.0) (2026-06-13) ### Bug Fixes * Add async_supported property to RedisOnlineStore ([9b088fe](feast-dev@9b088fe)) * Add missing feast init templates to operator CRD and enhance persistence documentation ([1941d4d](feast-dev@1941d4d)) * Allow to publish from reference branch ([5458ec8](feast-dev@5458ec8)) * API calls list ([4203eb7](feast-dev@4203eb7)) * **bigquery:** Enable list inference for parquet loads in offline_write_batch ([9243497](feast-dev@9243497)), closes [feast-dev#5845](feast-dev#5845) * Bump grpcio dependencies ([07b4782](feast-dev@07b4782)) * **compute-engine/local:** Honor field_mapping on join keys in dedup + join nodes ([feast-dev#6395](feast-dev#6395)) ([bd01824](feast-dev@bd01824)) * **dynamodb:** Avoid tag race condition by using diff-based tag updates ([feast-dev#6479](feast-dev#6479)) ([bad2b7d](feast-dev@bad2b7d)), closes [feast-dev#6418](feast-dev#6418) * **dynamodb:** Fix mypy type for _build_projection_expression return ([217b4da](feast-dev@217b4da)) * Fix intermittent async test failures for DynamoDB and Redis ([63c5eb1](feast-dev@63c5eb1)) * Fix mongodb blog title ([57d28d4](feast-dev@57d28d4)) * Fix shared SQL registry crash - avoid unnecessary UDF deserialization in proto cache building ([ac588d7](feast-dev@ac588d7)) * Fix SparkRetrievalJob.persist() failing for SparkSource ([209d7cd](feast-dev@209d7cd)) * Fixed formatting and image for mongo blog ([feast-dev#6377](feast-dev#6377)) ([f8389fb](feast-dev@f8389fb)) * Fixes for ray source ([7f592a4](feast-dev@7f592a4)) * **go:** skip registry refresh when cache_ttl_seconds <= 0 ([97ed40c](feast-dev@97ed40c)) * Handle array of strings columns in Athena materialization ([feast-dev#6324](feast-dev#6324)) ([4ed0278](feast-dev@4ed0278)) * make milvus VARCHAR max_length configurable, remove hardcoded 512 limit ([3b98c22](feast-dev@3b98c22)) * **operator:** Set appProtocol: grpc on registry gRPC Service ([feast-dev#6367](feast-dev#6367)) ([c9ae2b4](feast-dev@c9ae2b4)) * PyJWT 2.10+ added validation that rejects empty HMAC keys ([e756ffe](feast-dev@e756ffe)) * RemoteOnlineStore sends all features in a single HTTP request ([8f187dd](feast-dev@8f187dd)) * Remove registry proto dump to enforce RBAC and add permission checks to Commit/Refresh RPCs ([328431f](feast-dev@328431f)) * Remove selector migration job - no longer needed ([51c325e](feast-dev@51c325e)) * replace broken .claude skill symlink with correct relative path ([4541690](feast-dev@4541690)) * Replace selector label strip patch with migration Job for upgrade-safe selector uniqueness ([00dea50](feast-dev@00dea50)) * Scope feature view name conflict check to current project in file-based registry ([feast-dev#6369](feast-dev#6369)) ([a4fde83](feast-dev@a4fde83)), closes [feast-dev#6209](feast-dev#6209) * **snowflake:** Stop double-quoting connection identifiers ([feast-dev#6462](feast-dev#6462)) ([e914d59](feast-dev@e914d59)) * **spark:** S3/GCS PyArrow filesystem resolution for staging paths ([feast-dev#6442](feast-dev#6442)) ([ae50414](feast-dev@ae50414)) * **trino:** Clean up temporary entity tables after retrieval ([feast-dev#6381](feast-dev#6381)) ([d86b13d](feast-dev@d86b13d)), closes [feast-dev#6306](feast-dev#6306) * Update go-feature-server base image to Go 1.25 and fix operator Dockerfile COPY permissions ([86ef0bc](feast-dev@86ef0bc)) ### Features * [Backend] Data Quality Monitoring with native compute, multi-backend support, REST API, CLI ([feast-dev#6202](feast-dev#6202)) ([5458c37](feast-dev@5458c37)) * Add apache flink compute engine ([feast-dev#6476](feast-dev#6476)) ([9636d6a](feast-dev@9636d6a)) * Add demo noteboooks for users ([e362173](feast-dev@e362173)) * Add enabled/disabled toggle for feature views ([feast-dev#6401](feast-dev#6401)) ([5f1fa0d](feast-dev@5f1fa0d)), closes [feast-dev#6395](feast-dev#6395) * Add Label View to init template ([ec272d5](feast-dev@ec272d5)) * Add mTLS support to remote registry gRPC client ([feast-dev#6474](feast-dev#6474)) ([c9602d8](feast-dev@c9602d8)) * Add Prometheus gauges for FeatureStore installation telemetry ([feast-dev#6354](feast-dev#6354)) ([1b681b7](feast-dev@1b681b7)) * Adds registry REST API endpoints for managing entities, data sources, and feature views ([feast-dev#6413](feast-dev#6413)) ([f77bd1d](feast-dev@f77bd1d)) * Allow CRUD on entities, data sources, and feature views from UI ([feast-dev#6412](feast-dev#6412)) ([2321c07](feast-dev@2321c07)) * Allow default openlineage configuration ([feast-dev#6467](feast-dev#6467)) ([276b6df](feast-dev@276b6df)) * **bigquery:** Support DATE-type event timestamp columns ([feast-dev#6362](feast-dev#6362)) ([753dee5](feast-dev@753dee5)), closes [feast-dev#2530](feast-dev#2530) * **cli:** Add `feast projects delete` command (closes [feast-dev#5095](feast-dev#5095)) ([feast-dev#6318](feast-dev#6318)) ([1a4b96c](feast-dev@1a4b96c)) * Data Quality Monitoring added in feast UI ([feast-dev#6422](feast-dev#6422)) ([fa271be](feast-dev@fa271be)) * **dynamodb:** Use ProjectionExpression when requested_features is set ([0adc906](feast-dev@0adc906)), closes [feast-dev#6058](feast-dev#6058) * Enhance DataSource and FeatureView modals with error handling and submission states ([96d7169](feast-dev@96d7169)) * Expose registry endpoints on feature server for MCP access ([f77981c](feast-dev@f77981c)) * Feast First-Class LabelView Implementation ([feast-dev#6292](feast-dev#6292)) ([c0e7e5d](feast-dev@c0e7e5d)) * Feast-MLflow Integration ([feast-dev#6235](feast-dev#6235)) ([7279c75](feast-dev@7279c75)) * Operational metrics for offline store and SOX metrics for both ([feast-dev#6340](feast-dev#6340)) ([65b1b80](feast-dev@65b1b80)) * Pre-compute feature service ([8011550](feast-dev@8011550)) * REST API-backed UI for RBAC compatibility and per-page lazy loading ([feast-dev#6414](feast-dev#6414)) ([6ae80af](feast-dev@6ae80af)) * Support non-string map key types ([feast-dev#6382](feast-dev#6382)) ([feast-dev#6383](feast-dev#6383)) ([728aa2e](feast-dev@728aa2e)) * Update FeatureStore CRD with DRA Fields ([01241e4](feast-dev@01241e4)) ### Performance Improvements * Cache feature view resolution in get_online_features to reduce per-request overhead ([55c2f18](feast-dev@55c2f18)) * Optimize feature serving latency with batched async Redis, cached checks fix ([103809a](feast-dev@103809a)) * Replace MessageToDict with optimized custom dict builder ([feast-dev#6015](feast-dev#6015)) ([9902064](feast-dev@9902064))
What this PR does / why we need it:
final_mlflow_demo.mp4Native MLflow integration for Feast that provides automatic feature lineage tracking alongside ML experiments. When enabled via feature_store.yaml, every feature retrieval is logged to the active MLflow run.
Key Capabilities
Demo
Full end-to-end demo notebook with all capabilities: integration-example
Which issue(s) this PR fixes:
Checks
Testing Strategy
Misc