René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"OnDemandFeatureView.feature_transformation.infer_features does pass UDF outputs to python_type_to_feast_value_type","articleBody":"## Expected Behavior \r\n\r\n`OnDemandFeatureView.feature_transformation.infer_features` should be able to infer features from primitive python types for all supported feast data types, for all transformation backends.\r\n\r\n## Current Behavior\r\n\r\nAll on demand feature views are currently broken for list types, as there is no way to bypass schema inference.\r\n\r\n### Details\r\n\r\n`OnDemandFeatureView.feature_transformation.infer_features` can only infer features in the type map inside `python_type_to_feast_value_type`, _i.e._\r\n\r\n```Python\r\ntype_map = {\r\n    \"int\": ValueType.INT64,\r\n    \"str\": ValueType.STRING,\r\n    \"string\": ValueType.STRING,  # pandas.StringDtype\r\n    \"float\": ValueType.DOUBLE,\r\n    \"bytes\": ValueType.BYTES,\r\n    \"float64\": ValueType.DOUBLE,\r\n    \"float32\": ValueType.FLOAT,\r\n    \"int64\": ValueType.INT64,\r\n    \"uint64\": ValueType.INT64,\r\n    \"int32\": ValueType.INT32,\r\n    \"uint32\": ValueType.INT32,\r\n    \"int16\": ValueType.INT32,\r\n    \"uint16\": ValueType.INT32,\r\n    \"uint8\": ValueType.INT32,\r\n    \"int8\": ValueType.INT32,\r\n    \"bool\": ValueType.BOOL,\r\n    \"boolean\": ValueType.BOOL,\r\n    \"timedelta\": ValueType.UNIX_TIMESTAMP,\r\n    \"timestamp\": ValueType.UNIX_TIMESTAMP,\r\n    \"datetime\": ValueType.UNIX_TIMESTAMP,\r\n    \"datetime64[ns]\": ValueType.UNIX_TIMESTAMP,\r\n    \"datetime64[ns, tz]\": ValueType.UNIX_TIMESTAMP,\r\n    \"category\": ValueType.STRING,\r\n}\r\n```\r\n\r\nThis is because if the type _e.g._ `ValueType.FLOAT_LIST` doesn't have a mapping in the dictionary above, and `value is None`, then `isinstance(value, dtype)` checks will fall through to the `ValueError` in `python_type_to_feast_value_type`.\r\n\r\n## Steps to reproduce\r\n\r\nInitialize a new repository:\r\n\r\n```bash\r\nfeast init\r\n```\r\n\r\nModify the sample `on_demand_feature_view` to return an array of floats instead of just floats, _e.g._\r\n\r\n```diff\r\ndiff --git a/true_garfish/feature_repo/example_repo.py b/true_garfish/feature_repo/example_repo.py\r\nindex 1f5b946..59d4501 100644\r\n--- a/true_garfish/feature_repo/example_repo.py\r\n+++ b/true_garfish/feature_repo/example_repo.py\r\n@@ -16,7 +16,7 @@ from feast import (\r\n from feast.feature_logging import LoggingConfig\r\n from feast.infra.offline_stores.file_source import FileLoggingDestination\r\n from feast.on_demand_feature_view import on_demand_feature_view\r\n-from feast.types import Float32, Float64, Int64\r\n+from feast.types import Float32, Float64, Int64, Array\r\n \r\n # Define an entity for the driver. You can think of an entity as a primary key used to\r\n # fetch features.\r\n@@ -72,15 +72,16 @@ input_request = RequestSource(\r\n @on_demand_feature_view(\r\n     sources=[driver_stats_fv, input_request],\r\n     schema=[\r\n-        Field(name=\"conv_rate_plus_val1\", dtype=Float64),\r\n-        Field(name=\"conv_rate_plus_val2\", dtype=Float64),\r\n+        Field(name=\"conv_rate_plus_vals\", dtype=Array(Float64)),\r\n     ],\r\n )\r\n def transformed_conv_rate(inputs: pd.DataFrame) -\u003e pd.DataFrame:\r\n-    df = pd.DataFrame()\r\n-    df[\"conv_rate_plus_val1\"] = inputs[\"conv_rate\"] + inputs[\"val_to_add\"]\r\n-    df[\"conv_rate_plus_val2\"] = inputs[\"conv_rate\"] + inputs[\"val_to_add_2\"]\r\n-    return df\r\n+    result = {\"conv_rate_plus_vals\": []}\r\n+    for _, row in inputs.iterrows():\r\n+        result[\"conv_rate_plus_vals\"].append(\r\n+            [row[\"conv_rate\"] + row[\"val_to_add\"], row[\"conv_rate\"] + row[\"val_to_add_2\"]]\r\n+        )\r\n+    return pd.DataFrame(data=result)\r\n```\r\n\r\n3. Run `feast apply`, and you should get the following error:\r\n\r\n```bash\r\nTraceback (most recent call last):\r\n  File \"~/.../.venv/bin/feast\", line 8, in \u003cmodule\u003e\r\n    sys.exit(cli())\r\n             ^^^^^\r\n  File \"~/.../.venv/lib/python3.12/site-packages/click/core.py\", line 1157, in __call__\r\n    return self.main(*args, **kwargs)\r\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n  File \"~/.../.venv/lib/python3.12/site-packages/click/core.py\", line 1078, in main\r\n    rv = self.invoke(ctx)\r\n         ^^^^^^^^^^^^^^^^\r\n  File \"~/.../.venv/lib/python3.12/site-packages/click/core.py\", line 1688, in invoke\r\n    return _process_result(sub_ctx.command.invoke(sub_ctx))\r\n                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n  File \"~/.../.venv/lib/python3.12/site-packages/click/core.py\", line 1434, in invoke\r\n    return ctx.invoke(self.callback, **ctx.params)\r\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n  File \"~/.../.venv/lib/python3.12/site-packages/click/core.py\", line 783, in invoke\r\n    return __callback(*args, **kwargs)\r\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n  File \"~/.../.venv/lib/python3.12/site-packages/click/decorators.py\", line 33, in new_func\r\n    return f(get_current_context(), *args, **kwargs)\r\n           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n  File \"~/.../.venv/lib/python3.12/site-packages/feast/cli.py\", line 506, in apply_total_command\r\n    apply_total(repo_config, repo, skip_source_validation)\r\n  File \"~/.../.venv/lib/python3.12/site-packages/feast/repo_operations.py\", line 347, in apply_total\r\n    apply_total_with_repo_instance(\r\n  File \"~/.../.venv/lib/python3.12/site-packages/feast/repo_operations.py\", line 299, in apply_total_with_repo_instance\r\n    registry_diff, infra_diff, new_infra = store.plan(repo)\r\n                                           ^^^^^^^^^^^^^^^^\r\n  File \"~/.../.venv/lib/python3.12/site-packages/feast/feature_store.py\", line 745, in plan\r\n    self._make_inferences(\r\n  File \"~/.../.venv/lib/python3.12/site-packages/feast/feature_store.py\", line 640, in _make_inferences\r\n    odfv.infer_features()\r\n  File \"~/.../.venv/lib/python3.12/site-packages/feast/on_demand_feature_view.py\", line 521, in infer_features\r\n    inferred_features = self.feature_transformation.infer_features(\r\n                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n  File \"~/....venv/lib/python3.12/site-packages/feast/transformation/pandas_transformation.py\", line 47, in infer_features\r\n    python_type_to_feast_value_type(f, type_name=str(dt))\r\n  File \"~/.../.venv/lib/python3.12/site-packages/feast/type_map.py\", line 215, in python_type_to_feast_value_type\r\n    raise ValueError(\r\nValueError: Value with native type object cannot be converted into Feast value type\r\n```\r\n\r\nAdding some debug statements inside `python_type_to_feast_value_type`, we get the following locals before the error was raised:\r\n\r\n```\r\nname='conv_rate_plus_vals'\r\nvalue=None\r\nrecurse=True\r\ntype_name='object'\r\ntype(value)=\u003cclass 'NoneType'\u003e\r\n``` \r\n\r\nAs mentioned before this is because all transformation backends don't pass values to the type mapper, _e.g._ the [pandas backend in this case](https://github.com/feast-dev/feast/blob/master/sdk/python/feast/transformation/pandas_transformation.py#L47) \r\n\r\n### Specifications\r\n\r\n- Version: 0.39.0\r\n- Platform: arm64\r\n- Subsystem: MacOS\r\n\r\n## Possible Solution\r\n\r\n- Pass the sample values generated for type inference through to the type mapper\r\n- Update the type mapper to handle lists that are two levels deep. This is because primitive UDF outputs are wrapped in either a `np.array` or `list` of length 1, so therefore lists should be two levels deep with the inner list being the list of feature values.\r\n","author":{"url":"https://github.com/alexmirrington","@type":"Person","name":"alexmirrington"},"datePublished":"2024-06-24T08:49:49.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":1},"url":"https://github.com/4308/feast/issues/4308"}