feast apply broken with 0.15.* if the registry already exists

[MattDelac]

Expected Behavior

Running feast apply twice in a row should work as expected

Current Behavior

The second time we run feast apply, it will break with feast v0.15.*

Steps to reproduce

• install Feast 0.15.1
• Add a local config

project: local
registry: tmp/registry.db
provider: local
offline_store:
    type: bigquery
    dataset: scratch
online_store:
    type: sqlite
    path: tmp/database.db

• feast apply ✅ if the registry does not exist
• feast apply 🔴

From Slack > And interestingly, it does not always fail on the same FeatureView being applied > And I can successfully print feature_view_proto.spec.entities right before it fails

Specifications

• Version:
• Platform:
• Subsystem:

Possible Solution

[adchia]

There was a fix shipped to better handle entityless feature views, but seems like we're having trouble reproducing this bug.

Can you try again at HEAD and see if this still reproduces?

[MattDelac]

There was a fix shipped to better handle entityless feature views, but seems like we're having trouble reproducing this bug.

Can you try again at HEAD and see if this still reproduces?

I get another issue that does not seem relevant to this ticket

Was using the following commit 54d0f3a1842c7613aaffcd67512e08ab9b688740

"feast[gcp] @ git+https://github.com/feast-dev/feast.git@54d0f3a1842c7613aaffcd67512e08ab9b688740#subdirectory=sdk/python&egg=feast",

[felixwang9817]

hey @MattDelac, this issue looks like it might be related to #2038 and #2063

maybe try pip installing tensorflow-metadata and trying again?

[MattDelac]

hey @MattDelac, this issue looks like it might be related to #2038 and #2063

maybe try pip installing tensorflow-metadata and trying again?

It was already installed tensorflow-metadata==1.4.0, so it did not change anything

[MattDelac]

hey @MattDelac, this issue looks like it might be related to #2038 and #2063

maybe try pip installing tensorflow-metadata and trying again?

I tried all versions of tensorflow-metadata from 1.2.0 to 1.5.0

Always get the same error. Would today be a good time to pair on this issue ?

[MattDelac]

So the problem is that I create a child class from the Feast FeatureView class. I do so in order to add Shopify specifics.

Since feast==0.15.0, the cls argument passed to the classmethod from_proto() was the child class instead of the parent class. Thus, the from_proto() method was broken when it was trying to create a new FeatureView object.

The solution is to override this method class and explicitly return the one coming from the parent class (see code below)

from feast.feature_view import FeatureView as FeastFeatureView

class FeatureView(FeastFeatureView):
    def __init__(
        self,
        name: str,
        entity: str,
        owner: str,
        features: List[Feature],
        batch_source: CustomSource,
        ttl: Optional[timedelta] = None,
    ):
        super().__init__(
            name=name,
            entities=[entity],
            features=features,
            batch_source=batch_source,
            ttl=ttl,
            tags={...}
        )

    @classmethod
    def from_proto(self, feature_view_proto):
        return FeastFeatureView.from_proto(feature_view_proto)