feast apply silently ignores ttl updates when the new ttl is None or timedelta(0) #6703

Expected Behavior

Re-applying an existing FeatureView with ttl=None or ttl=timedelta(0) should clear the ttl in the registry. After feast apply, store.get_feature_view("foo_fv").ttl should reflect the new "no expiry" value.

Current Behavior

The old ttl is silently retained. store.get_feature_view("foo_fv").ttl still returns the previous finite value, with no warning or error, while feast apply keeps printing the same "changed" diff on every run.

Registry.apply_feature_view() routes ttl changes through _update_metadata_fields() instead of creating a new version, and that method gates the ttl copy behind a truthiness check on the raw timedelta:

if (
    hasattr(existing_proto.spec, "ttl")
    and hasattr(updated_fv, "ttl")
    and updated_fv.ttl          # falsy for both None and timedelta(0)
):

Since None and timedelta(0) are both falsy — and both are the documented way to express "no ttl" — the update is dropped exactly in the case it should support. get_ttl_duration() already handles None correctly; only the outer guard is wrong. Still present on master as of 2026-08.

Practical impact: a FeatureView intended to never expire silently kept a stale finite ttl, causing get_historical_features to drop ~11.6% of otherwise-valid rows from a point-in-time join, with no error or warning.

Steps to reproduce

from datetime import timedelta
from feast import Entity, FeatureView, Field, FileSource, Project
from feast.types import Float32

project = Project(name="ttl_repro")
entity = Entity(name="foo", join_keys=["foo_id"])
source = FileSource(
    name="foo_source",
    path="./data/foo.parquet",
    timestamp_field="event_timestamp",
)

# 1. Apply with a finite ttl -> fv.ttl == 10 days (correct)
fv = FeatureView(
    name="foo_fv",
    entities=[entity],
    ttl=timedelta(days=10),
    schema=[Field(name="bar", dtype=Float32)],
    source=source,
)

# 2. Change ttl to "no expiry" and `feast apply` again
fv_v2 = FeatureView(
    name="foo_fv",
    entities=[entity],
    ttl=None,  # or timedelta(days=0) - same result
    schema=[Field(name="bar", dtype=Float32)],
    source=source,
)

Then check store.get_feature_view("foo_fv").ttl — it is still timedelta(days=10).

Specifications

• Version: 0.65.0 (PyPI); also confirmed on master as of 2026-08
• Platform:
• Subsystem: Registry — sdk/python/feast/infra/registry/registry.py, Registry._update_metadata_fields()

Possible Solution

Drop the outer truthiness check and always sync ttl on update, relying on get_ttl_duration()'s existing None-handling and defaulting to an explicit zero Duration when unset:

if hasattr(existing_proto.spec, "ttl") and hasattr(updated_fv, "ttl"):
    if isinstance(updated_fv, FeatureView):
        existing_proto.spec.ttl.CopyFrom(updated_fv.get_ttl_duration() or Duration())
    elif isinstance(updated_fv, LabelView):
        ttl_duration = Duration()
        if updated_fv.ttl is not None:
            ttl_duration.FromTimedelta(updated_fv.ttl)
        existing_proto.spec.ttl.CopyFrom(ttl_duration)

Happy to open a PR with this plus a regression test (apply with finite ttl, re-apply with ttl=None, assert the registry reflects it)