René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"Undefined features should be rejected when being fetched via `get_historical_features` / `get_online_features`","articleBody":"## Context\r\n\r\nI want to create versioned feature views. Through various versions, features could be added or removed.\r\n\r\n## Expected Behavior \r\n\r\nWhen doing `feast.get_historical_features`, features that are not defined should be rejected.\r\n\r\n## Current Behavior\r\n\r\nThe features get returned even though they have not been defined. \r\n\r\n## Steps to reproduce\r\n\r\n1. Initialize a new feast repository \r\n2. Define the features:\r\n\r\n```python\r\ndriver_hourly_stats = FileSource(\r\n    path=\"/home/benjamintan/workspace/feast-workflow-demo/feature_repo/data/driver_stats.parquet\",\r\n    timestamp_field=\"event_timestamp\",\r\n    created_timestamp_column=\"created\",\r\n)\r\n\r\ndriver = Entity(name=\"driver_id\", value_type=ValueType.INT64, description=\"driver id\",)\r\n\r\ndriver_hourly_stats_view_v1 = FeatureView(\r\n    name=\"driver_hourly_stats_v1\",\r\n    entities=[\"driver_id\"],\r\n    ttl=timedelta(days=1),\r\n    schema=[\r\n        Field(name=\"avg_daily_trips\", dtype=Int64),\r\n    ],\r\n    online=True,\r\n    batch_source=driver_hourly_stats,\r\n    tags={},\r\n)\r\n\r\ndriver_hourly_stats_view_v2 = FeatureView(\r\n    name=\"driver_hourly_stats_v2\",\r\n    entities=[\"driver_id\"],\r\n    ttl=timedelta(days=1),\r\n    schema=[\r\n        Field(name=\"conv_rate\", dtype=Float32),\r\n        Field(name=\"acc_rate\", dtype=Float32),\r\n        Field(name=\"avg_daily_trips\", dtype=Int64),\r\n    ],\r\n    online=True,\r\n    batch_source=driver_hourly_stats,\r\n    tags={},\r\n)\r\n```\r\n3. `feast apply`\r\n4. Querying Feast:\r\n\r\n```python\r\nfs = FeatureStore(repo_path='.')\r\n\r\nentity_df = pd.DataFrame(\r\n    {\r\n        \"event_timestamp\": [\r\n            pd.Timestamp(dt, unit=\"ms\", tz=\"UTC\").round(\"ms\")\r\n            for dt in pd.date_range(\r\n                start=datetime.now() - timedelta(days=3),\r\n                end=datetime.now(),\r\n                periods=3,\r\n            )\r\n        ],\r\n        \"driver_id\": [1001, 1002, 1003],\r\n    }\r\n)\r\n```\r\nI _do not_ expect the following to work:\r\n\r\n```python\r\n# THIS PART SHOULDN'T WORK\r\nfeatures_wrong = ['driver_hourly_stats_v1:conv_rate',  # doesn't exist in V1\r\n                              'driver_hourly_stats_v1:acc_rate',   # doesn't exist in V1\r\n                              'driver_hourly_stats_v1:avg_daily_trips',\r\n                             ]\r\n\r\nhist_features_wrong = fs.get_historical_features(\r\n    entity_df=entity_df,\r\n    features=features_wrong,\r\n)\r\n```\r\n\r\nBut I do get results:\r\n\r\n```\r\n                   event_timestamp  driver_id  ...  acc_rate  avg_daily_trips\r\n0 2022-04-17 09:35:35.658000+00:00       1001  ...  0.536431            742.0\r\n1 2022-04-18 21:35:35.658000+00:00       1002  ...  0.496901            678.0\r\n2 2022-04-20 09:35:35.658000+00:00       1003  ...       NaN     \r\n```\r\n\r\nI do not expect this to work because `driver_hourly_stats_v1:conv_rate` and `driver_hourly_stats_v1:acc_rate` were not defined in the `driver_hourly_stats_view_v1` FeatureView. \r\n\r\nAnd just to double check that `driver_hourly_stats_v1` only has `avg_daily_trips` defined:\r\n\r\n```\r\n➜ feast feature-views describe driver_hourly_stats_v1\r\nspec:\r\n  name: driver_hourly_stats_v1\r\n  entities:\r\n  - driver_id\r\n  features:\r\n  - name: avg_daily_trips\r\n    valueType: INT64\r\n  ttl: 86400s\r\n```\r\n\r\n### Specifications\r\n\r\n- Version: 0.20 (tested this on 0.19 and 0.18)\r\n- Platform: Linux\r\n- Subsystem: Ubuntu\r\n\r\n## Possible Solution\r\n\r\nThe list of features being passed in should be checked against the registry. Currently the feature view name and feature name pairs are not validated _together_. Here's an example that modifies `get_historical_features`:\r\n\r\n```python\r\n    @log_exceptions_and_usage\r\n    def get_historical_features(\r\n        self,\r\n        entity_df: Union[pd.DataFrame, str],\r\n        features: Union[List[str], FeatureService],\r\n        full_feature_names: bool = False,\r\n    ) -\u003e RetrievalJob:\r\n\r\n        # Build a dictionary of feature view names -\u003e feature names (not sure if this function already exists ...)\r\n        fv_name_features = dict([(fv.name, [f.name.split('-')[0] for f in fv.features]) for fv in self.list_feature_views()])\r\n        \r\n        # Check that input features are found in the `fv_name_features` dictionary\r\n        feature_views_not_found = []\r\n        for feature in features:\r\n            k, v = feature.split(\":\")\r\n            if v not in fv_name_features[k]:\r\n                feature_views_not_found.append(f'{k}:{v}')\r\n\r\n        if feature_views_not_found:\r\n            raise FeatureViewNotFoundException(', '.join(feature_views_not_found))\r\n```\r\n\r\nThis returns:\r\n\r\n```python\r\nfeast.errors.FeatureViewNotFoundException: Feature view driver_hourly_stats_v1:conv_rate, driver_hourly_stats_v1:acc_rate does not exist\r\n```\r\n\r\nThis doesn't handle the case when a `FeatureService` is passed in but it shouldn't be too hard.\r\n\r\n\r\nThis should also apply to `get_online_features`. \r\n","author":{"url":"https://github.com/benjamintanweihao","@type":"Person","name":"benjamintanweihao"},"datePublished":"2022-04-20T02:43:29.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":0},"url":"https://github.com/2576/feast/issues/2576"}