fix: Preserve millisecond precision for UNIX_TIMESTAMP sort keys

docs/reference/feature-store-yaml.md

@@ -24,7 +24,6 @@ online_store:
   * **path** \(a local filepath\) — Parameter for the sqlite online store. Defines the path to the SQLite database file.
   * **project\_id**  — Optional parameter for the datastore online store. Sets the GCP project id used by Feast, if not set Feast will use the default GCP project id in the local environment.
 * **project** — Defines a namespace for the entire feature store. Can be used to isolate multiple deployments in a single installation of Feast.
-
 ## Providers
 
 The `provider` field defines the environment in which Feast will execute data flows. As a result, it also determines the default values for other fields.

go/internal/feast/integration_tests/scylladb/feature_repo/regression_repo.py

@@ -0,0 +1,89 @@
+# Regression fixture for EAPC-22316 follow-up: a SortedFeatureView whose
+# UNIX_TIMESTAMP sort key is requested as a feature must retain millisecond
+# precision for rows that differ by less than one second (Go read-side fix in
+# InterfaceToProtoValue, go/types/typeconversion.go).
+
+from datetime import timedelta
+
+from feast import Entity, Field, FileSource, SortedFeatureView
+from feast.protos.feast.core.SortedFeatureView_pb2 import SortOrder
+from feast.sort_key import SortKey
+from feast.types import String, UnixTimestamp
+from feast.value_type import ValueType
+
+tags = {"team": "Feast"}
+owner = "test@test.com"
+
+sub_second_entity: Entity = Entity(
+    name="sub_second_entity",
+    join_keys=["sub_second_entity_id"],
+    value_type=ValueType.STRING,
+    tags=tags,
+    owner=owner,
+)
+
+sub_second_source: FileSource = FileSource(
+    path="sub_second_data.parquet", timestamp_field="event_timestamp"
+)
+
+sub_second_sort_key_view: SortedFeatureView = SortedFeatureView(
+    name="sub_second_sort_key_view",
+    entities=[sub_second_entity],
+    ttl=timedelta(days=0),
+    source=sub_second_source,
+    tags=tags,
+    description="Regression fixture: sort key rows <1s apart for the same entity",
+    owner=owner,
+    online=True,
+    sort_keys=[
+        SortKey(
+            name="event_timestamp",
+            value_type=ValueType.UNIX_TIMESTAMP,
+            default_sort_order=SortOrder.DESC,
+        )
+    ],
+    schema=[
+        Field(name="value", dtype=String),
+        Field(name="event_timestamp", dtype=UnixTimestamp),
+    ],
+)
+
+# Second regression fixture: the sort key is a secondary feature with a name
+# other than "event_timestamp" (mirroring the real customer schema, where the
+# sort key was named "viewed_at" and was distinct from the source's own
+# ingestion timestamp field). This guards against the fix accidentally being
+# tied to the "event_timestamp" column name rather than working for any
+# UNIX_TIMESTAMP column declared as a sort key.
+custom_sortkey_entity: Entity = Entity(
+    name="custom_sortkey_entity",
+    join_keys=["custom_sortkey_entity_id"],
+    value_type=ValueType.STRING,
+    tags=tags,
+    owner=owner,
+)
+
+sub_second_custom_sortkey_source: FileSource = FileSource(
+    path="sub_second_custom_sortkey_data.parquet", timestamp_field="event_timestamp"
+)
+
+sub_second_custom_sortkey_view: SortedFeatureView = SortedFeatureView(
+    name="sub_second_custom_sortkey_view",
+    entities=[custom_sortkey_entity],
+    ttl=timedelta(days=0),
+    source=sub_second_custom_sortkey_source,
+    tags=tags,
+    description="Regression fixture: sort key named 'viewed_at' (not 'event_timestamp'), rows <1s apart",
+    owner=owner,
+    online=True,
+    sort_keys=[
+        SortKey(
+            name="viewed_at",
+            value_type=ValueType.UNIX_TIMESTAMP,
+            default_sort_order=SortOrder.DESC,
+        )
+    ],
+    schema=[
+        Field(name="value", dtype=String),
+        Field(name="viewed_at", dtype=UnixTimestamp),
+    ],
+)

go/internal/feast/integration_tests/scylladb/http/http_integration_test.go

@@ -215,7 +215,7 @@ func TestGetOnlineFeaturesRange_Http_withOnlyEqualsFilter(t *testing.T) {
 	  "sort_key_filters": [
 		{
 		  "sort_key_name": "event_timestamp",
-		  "equals": 1744769171
+		  "equals": 1744769171919
 		}
 	  ],
 	  "limit": 10

go/internal/feast/integration_tests/scylladb/http/valid_equals_response.json

@@ -357,7 +357,7 @@
     {
       "values": [
         [
-          "2025-04-16 02:06:11Z"
+          "2025-04-16 02:06:11.919Z"
         ]
       ]
     }

go/internal/feast/integration_tests/scylladb/http/valid_response.json

@@ -70,6 +70,6 @@
   }, {
     "values" : [ [ null, null, null, null, null, null, null, null, null, null ], [ null, null, null, null, null, null, null, null, null, null ], [ null, null, null, null, null, null, null, null, null, null ] ]
   }, {
-    "values" : [ [ "2025-04-16 02:06:10Z", "2025-04-16 02:06:09Z", "2025-04-16 02:06:08Z", "2025-04-16 02:06:07Z", "2025-04-16 02:06:06Z", "2025-04-16 02:06:05Z", "2025-04-16 02:06:04Z", "2025-04-16 02:06:03Z", "2025-04-16 02:06:02Z", "2025-04-16 02:06:01Z" ], [ "2025-04-16 02:06:11Z", "2025-04-16 02:06:10Z", "2025-04-16 02:06:09Z", "2025-04-16 02:06:08Z", "2025-04-16 02:06:07Z", "2025-04-16 02:06:06Z", "2025-04-16 02:06:05Z", "2025-04-16 02:06:04Z", "2025-04-16 02:06:03Z", "2025-04-16 02:06:02Z" ], [ "2025-04-16 02:06:11Z", "2025-04-16 02:06:10Z", "2025-04-16 02:06:09Z", "2025-04-16 02:06:08Z", "2025-04-16 02:06:07Z", "2025-04-16 02:06:06Z", "2025-04-16 02:06:05Z", "2025-04-16 02:06:04Z", "2025-04-16 02:06:03Z", "2025-04-16 02:06:02Z" ] ]
+    "values" : [ [ "2025-04-16 02:06:10.919Z", "2025-04-16 02:06:09.919Z", "2025-04-16 02:06:08.919Z", "2025-04-16 02:06:07.919Z", "2025-04-16 02:06:06.919Z", "2025-04-16 02:06:05.919Z", "2025-04-16 02:06:04.919Z", "2025-04-16 02:06:03.919Z", "2025-04-16 02:06:02.919Z", "2025-04-16 02:06:01.919Z" ], [ "2025-04-16 02:06:11.919Z", "2025-04-16 02:06:10.919Z", "2025-04-16 02:06:09.919Z", "2025-04-16 02:06:08.919Z", "2025-04-16 02:06:07.919Z", "2025-04-16 02:06:06.919Z", "2025-04-16 02:06:05.919Z", "2025-04-16 02:06:04.919Z", "2025-04-16 02:06:03.919Z", "2025-04-16 02:06:02.919Z" ], [ "2025-04-16 02:06:11.919Z", "2025-04-16 02:06:10.919Z", "2025-04-16 02:06:09.919Z", "2025-04-16 02:06:08.919Z", "2025-04-16 02:06:07.919Z", "2025-04-16 02:06:06.919Z", "2025-04-16 02:06:05.919Z", "2025-04-16 02:06:04.919Z", "2025-04-16 02:06:03.919Z", "2025-04-16 02:06:02.919Z" ] ]
   } ]
 }

go/internal/feast/integration_tests/scylladb/scylladb_integration_test.go

@@ -129,7 +129,7 @@ func TestGetOnlineFeaturesRange_withOnlyEqualsFilter(t *testing.T) {
 			{
 				SortKeyName: "event_timestamp",
 				Query: &serving.SortKeyFilter_Equals{
-					Equals: &types.Value{Val: &types.Value_UnixTimestampVal{UnixTimestampVal: 1744769171}},
+					Equals: &types.Value{Val: &types.Value_UnixTimestampVal{UnixTimestampVal: 1744769171919}},
 				},
 			},
 		},
@@ -370,6 +370,181 @@ func TestGetOnlineFeaturesRange_withFeatureViewThrowsError(t *testing.T) {
 		"Expected error message for non-existent sorted feature view")
 }
 
+// TestGetOnlineFeaturesRange_SubSecondSortKeyValuesRemainDistinct is a
+// regression test for the EAPC-22316 follow-up: InterfaceToProtoValue in
+// go/types/typeconversion.go was truncating UNIX_TIMESTAMP sort key values
+// read back from Cassandra to whole seconds via time.Time.Unix(), causing
+// rows for the same entity that differ by less than a second to appear
+// identical when the sort key is requested as a feature value. The fixture
+// (sub_second_sort_key_view, seeded from sub_second_data.parquet) has three
+// rows for one entity: two only 18ms apart within the same second, and one
+// two seconds later - mirroring the real-world repro that caught this bug.
+func TestGetOnlineFeaturesRange_SubSecondSortKeyValuesRemainDistinct(t *testing.T) {
+	entities := make(map[string]*types.RepeatedValue)
+	entities["sub_second_entity_id"] = &types.RepeatedValue{
+		Val: []*types.Value{
+			{Val: &types.Value_StringVal{StringVal: "entity-1"}},
+		},
+	}
+
+	request := &serving.GetOnlineFeaturesRangeRequest{
+		Kind: &serving.GetOnlineFeaturesRangeRequest_Features{
+			Features: &serving.FeatureList{
+				Val: []string{
+					"sub_second_sort_key_view:event_timestamp",
+					"sub_second_sort_key_view:value",
+				},
+			},
+		},
+		Entities: entities,
+		SortKeyFilters: []*serving.SortKeyFilter{
+			{
+				SortKeyName: "event_timestamp",
+				Query: &serving.SortKeyFilter_Range{
+					Range: &serving.SortKeyFilter_RangeQuery{
+						RangeStart: &types.Value{Val: &types.Value_UnixTimestampVal{UnixTimestampVal: 0}},
+					},
+				},
+			},
+		},
+		Limit:           10,
+		IncludeMetadata: true,
+	}
+
+	response, err := client.GetOnlineFeaturesRange(ctx, request)
+	require.NoError(t, err)
+	require.NotNil(t, response)
+	require.Len(t, response.Results, 2, "expected event_timestamp and value results")
+
+	eventTimestampIdx := -1
+	for i, name := range response.Metadata.FeatureNames.Val {
+		if name == "event_timestamp" {
+			eventTimestampIdx = i
+		}
+	}
+	require.NotEqual(t, -1, eventTimestampIdx, "expected to find event_timestamp in the response")
+
+	eventTimestampVector := response.Results[eventTimestampIdx]
+	require.Len(t, eventTimestampVector.Values, 1, "expected results for the single requested entity")
+
+	timestamps := eventTimestampVector.Values[0].Val
+	require.Len(t, timestamps, 3, "expected all 3 sub-second-apart rows to be returned")
+
+	seen := make(map[int64]bool)
+	for _, v := range timestamps {
+		ts := v.GetUnixTimestampVal()
+		assert.False(t, seen[ts], "sort key value %d was returned more than once; sub-second rows collapsed", ts)
+		seen[ts] = true
+	}
+	assert.Len(t, seen, 3, "expected 3 distinct millisecond-precision sort key values")
+
+	expected := map[int64]bool{
+		1717244257886: true, // 2024-06-01 12:17:37.886
+		1717244257904: true, // 2024-06-01 12:17:37.904
+		1717244259035: true, // 2024-06-01 12:17:39.035
+	}
+	for ts := range seen {
+		assert.True(t, expected[ts], "unexpected sort key value %d", ts)
+	}
+
+	// EventTimestamps come from the _ts:<fv> hash field (a timestamppb.Timestamp).
+	// grpc_server calls GetTimestampSeconds, so values must be whole-second Unix timestamps.
+	require.Len(t, eventTimestampVector.EventTimestamps, 1, "expected EventTimestamps for 1 entity")
+	require.Len(t, eventTimestampVector.EventTimestamps[0].Val, 3, "expected EventTimestamp for each of the 3 rows")
+	for _, tsVal := range eventTimestampVector.EventTimestamps[0].Val {
+		secs := tsVal.GetUnixTimestampVal()
+		assert.True(t, secs > 1_000_000_000 && secs < 2_000_000_000,
+			"EventTimestamp must be seconds-precision (~2001–2033), got %d", secs)
+	}
+}
+
+// TestGetOnlineFeaturesRange_CustomNamedSortKeyValuesRemainDistinct guards
+// against the EAPC-22316 fix being tied to the "event_timestamp" column name
+// rather than working for any UNIX_TIMESTAMP column declared as a sort key.
+// The fixture (sub_second_custom_sortkey_view, seeded from
+// sub_second_custom_sortkey_data.parquet) declares "viewed_at" - a name
+// distinct from the source's own "event_timestamp" watermark field - as the
+// sort key, mirroring the real customer schema. Same three-row shape: two
+// rows 18ms apart within the same second, one two seconds later.
+func TestGetOnlineFeaturesRange_CustomNamedSortKeyValuesRemainDistinct(t *testing.T) {
+	entities := make(map[string]*types.RepeatedValue)
+	entities["custom_sortkey_entity_id"] = &types.RepeatedValue{
+		Val: []*types.Value{
+			{Val: &types.Value_StringVal{StringVal: "entity-1"}},
+		},
+	}
+
+	request := &serving.GetOnlineFeaturesRangeRequest{
+		Kind: &serving.GetOnlineFeaturesRangeRequest_Features{
+			Features: &serving.FeatureList{
+				Val: []string{
+					"sub_second_custom_sortkey_view:viewed_at",
+					"sub_second_custom_sortkey_view:value",
+				},
+			},
+		},
+		Entities: entities,
+		SortKeyFilters: []*serving.SortKeyFilter{
+			{
+				SortKeyName: "viewed_at",
+				Query: &serving.SortKeyFilter_Range{
+					Range: &serving.SortKeyFilter_RangeQuery{
+						RangeStart: &types.Value{Val: &types.Value_UnixTimestampVal{UnixTimestampVal: 0}},
+					},
+				},
+			},
+		},
+		Limit:           10,
+		IncludeMetadata: true,
+	}
+
+	response, err := client.GetOnlineFeaturesRange(ctx, request)
+	require.NoError(t, err)
+	require.NotNil(t, response)
+	require.Len(t, response.Results, 2, "expected viewed_at and value results")
+
+	viewedAtIdx := -1
+	for i, name := range response.Metadata.FeatureNames.Val {
+		if name == "viewed_at" {
+			viewedAtIdx = i
+		}
+	}
+	require.NotEqual(t, -1, viewedAtIdx, "expected to find viewed_at in the response")
+
+	viewedAtVector := response.Results[viewedAtIdx]
+	require.Len(t, viewedAtVector.Values, 1, "expected results for the single requested entity")
+
+	timestamps := viewedAtVector.Values[0].Val
+	require.Len(t, timestamps, 3, "expected all 3 sub-second-apart rows to be returned")
+
+	seen := make(map[int64]bool)
+	for _, v := range timestamps {
+		ts := v.GetUnixTimestampVal()
+		assert.False(t, seen[ts], "sort key value %d was returned more than once; sub-second rows collapsed", ts)
+		seen[ts] = true
+	}
+	assert.Len(t, seen, 3, "expected 3 distinct millisecond-precision sort key values")
+
+	expected := map[int64]bool{
+		1717244257886: true, // 2024-06-01 12:17:37.886
+		1717244257904: true, // 2024-06-01 12:17:37.904
+		1717244259035: true, // 2024-06-01 12:17:39.035
+	}
+	for ts := range seen {
+		assert.True(t, expected[ts], "unexpected sort key value %d", ts)
+	}
+
+	// EventTimestamps come from the _ts:<fv> hash field (a timestamppb.Timestamp).
+	// grpc_server calls GetTimestampSeconds, so values must be whole-second Unix timestamps.
+	require.Len(t, viewedAtVector.EventTimestamps, 1, "expected EventTimestamps for 1 entity")
+	require.Len(t, viewedAtVector.EventTimestamps[0].Val, 3, "expected EventTimestamp for each of the 3 rows")
+	for _, tsVal := range viewedAtVector.EventTimestamps[0].Val {
+		secs := tsVal.GetUnixTimestampVal()
+		assert.True(t, secs > 1_000_000_000 && secs < 2_000_000_000,
+			"EventTimestamp must be seconds-precision (~2001–2033), got %d", secs)
+	}
+}
+
 func assertResponseData(t *testing.T, response *serving.GetOnlineFeaturesRangeResponse, featureNames []string, entitiesRequested int, includeMetadata bool) {
 	assert.NotNil(t, response)
 	assert.Equal(t, 1, len(response.Entities), "Should have 1 list of entity")

go/internal/feast/integration_tests/valkey/feature_repo/regression_repo.py

@@ -0,0 +1,47 @@
+# Regression fixture for EAPC-22316: UNIX_TIMESTAMP sort keys must retain
+# millisecond precision for rows that differ by less than one second.
+
+from datetime import timedelta
+
+from feast import Entity, Field, FileSource, SortedFeatureView
+from feast.protos.feast.core.SortedFeatureView_pb2 import SortOrder
+from feast.sort_key import SortKey
+from feast.types import String, UnixTimestamp
+from feast.value_type import ValueType
+
+tags = {"team": "Feast"}
+owner = "test@test.com"
+
+sub_second_entity: Entity = Entity(
+    name="sub_second_entity",
+    join_keys=["sub_second_entity_id"],
+    value_type=ValueType.STRING,
+    tags=tags,
+    owner=owner,
+)
+
+sub_second_source: FileSource = FileSource(
+    path="sub_second_data.parquet", timestamp_field="event_timestamp"
+)
+
+sub_second_sort_key_view: SortedFeatureView = SortedFeatureView(
+    name="sub_second_sort_key_view",
+    entities=[sub_second_entity],
+    ttl=timedelta(days=0),
+    source=sub_second_source,
+    tags=tags,
+    description="Regression fixture: sort key rows <1s apart for the same entity",
+    owner=owner,
+    online=True,
+    sort_keys=[
+        SortKey(
+            name="event_timestamp",
+            value_type=ValueType.UNIX_TIMESTAMP,
+            default_sort_order=SortOrder.DESC,
+        )
+    ],
+    schema=[
+        Field(name="value", dtype=String),
+        Field(name="event_timestamp", dtype=UnixTimestamp),
+    ],
+)