RequestSource can not be used with write_to_online_store

[HaoXuAI]

Expected Behavior

Use RequestSource with write_to_online_store

Current Behavior

Steps to reproduce

feature view has request source defined such as

 {
  "spec": {
    "name": "embedding_catalog",
    "features": [
      {
        "name": "embeddings",
        "valueType": "FLOAT_LIST"
      },
      {
        "name": "event_timestamp",
        "valueType": "UNIX_TIMESTAMP"
      }
    ],
    "ttl": "0s",
    "batchSource": {
      "type": "REQUEST_SOURCE",
      "dataSourceClassType": "feast.data_source.RequestSource",
      "requestDataOptions": {
        "schema": [
          {
            "name": "embeddings",
            "valueType": "FLOAT_LIST"
          },
          {
            "name": "event_timestamp",
            "valueType": "UNIX_TIMESTAMP"
          }
        ]
      },
      "name": "embedding_catalog"
    },
    "online": true
  },
  "meta": {
    "createdTimestamp": "2024-09-06T07:58:13.929715Z",
    "lastUpdatedTimestamp": "2024-09-06T07:58:13.929715Z"
  }
}

in the _convert_arrow_to_proto function it will fail at extracting the timestamp field:

    event_timestamps = [
        _coerce_datetime(val)
        for val in pd.to_datetime(
            table.column(feature_view.batch_source.timestamp_field).to_numpy(
                zero_copy_only=False
            )
        )
    ]

since feature_view.batch_source (request source) doesn't have timestamp_field.

Specifications

• Version:
• Platform:
• Subsystem:

Possible Solution

add an attributee for RequestSource for the user to specify the timestamp_field.

[franciscojavierarceo]

I'm doing some related work in my PR to create persistent on demand feature views #4418

[tokoko]

I'm a little lost here, what's the use for timestamp_field in RequestSource? RequestSources are only used in online odfv execution, you don't have a point-in-time join there.

[HaoXuAI]

I'm a little lost here, what's the use for timestamp_field in RequestSource? RequestSources are only used in online odfv execution, you don't have a point-in-time join there.

I'm not using it to get history features, just use it as a data source to be able to ingest dataframe to the feature view directly.