feat: Faiss and In memory store (feast-dev#4464)

* add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add faiss & in memory online store Signed-off-by: cmuhao <sduxuhao@gmail.com> * add dependency Signed-off-by: cmuhao <sduxuhao@gmail.com> * update package name Signed-off-by: cmuhao <sduxuhao@gmail.com> --------- Signed-off-by: cmuhao <sduxuhao@gmail.com>
lokeshrangineni · Oct 23, 2024 · a1ff129 · a1ff129
1 parent 9ca1452
commit a1ff129
Show file tree

Hide file tree

Showing 2 changed files with 240 additions and 0 deletions.
diff --git a/sdk/python/feast/infra/online_stores/contrib/faiss_online_store.py b/sdk/python/feast/infra/online_stores/contrib/faiss_online_store.py
@@ -0,0 +1,236 @@
+import logging
+from datetime import datetime
+from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple
+
+import faiss
+import numpy as np
+from google.protobuf.timestamp_pb2 import Timestamp
+
+from feast import Entity, FeatureView, RepoConfig
+from feast.infra.key_encoding_utils import serialize_entity_key
+from feast.infra.online_stores.online_store import OnlineStore
+from feast.protos.feast.types.EntityKey_pb2 import EntityKey
+from feast.protos.feast.types.Value_pb2 import Value
+from feast.repo_config import FeastConfigBaseModel
+
+
+class FaissOnlineStoreConfig(FeastConfigBaseModel):
+    dimension: int
+    index_path: str
+    index_type: str = "IVFFlat"
+    nlist: int = 100
+
+
+class InMemoryStore:
+    def __init__(self):
+        self.feature_names: List[str] = []
+        self.entity_keys: Dict[str, int] = {}
+
+    def update(self, feature_names: List[str], entity_keys: Dict[str, int]):
+        self.feature_names = feature_names
+        self.entity_keys = entity_keys
+
+    def delete(self, entity_keys: List[str]):
+        for entity_key in entity_keys:
+            if entity_key in self.entity_keys:
+                del self.entity_keys[entity_key]
+
+    def read(self, entity_keys: List[str]) -> List[Optional[int]]:
+        return [self.entity_keys.get(entity_key) for entity_key in entity_keys]
+
+    def teardown(self):
+        self.feature_names = []
+        self.entity_keys = {}
+
+
+class FaissOnlineStore(OnlineStore):
+    _index: Optional[faiss.IndexIVFFlat] = None
+    _in_memory_store: InMemoryStore = InMemoryStore()
+    _config: Optional[FaissOnlineStoreConfig] = None
+    _logger: logging.Logger = logging.getLogger(__name__)
+
+    def _get_index(self, config: RepoConfig) -> faiss.IndexIVFFlat:
+        if self._index is None or self._config is None:
+            raise ValueError("Index is not initialized")
+        return self._index
+
+    def update(
+        self,
+        config: RepoConfig,
+        tables_to_delete: Sequence[FeatureView],
+        tables_to_keep: Sequence[FeatureView],
+        entities_to_delete: Sequence[Entity],
+        entities_to_keep: Sequence[Entity],
+        partial: bool,
+    ):
+        feature_views = tables_to_keep
+        if not feature_views:
+            return
+
+        feature_names = [f.name for f in feature_views[0].features]
+        dimension = len(feature_names)
+
+        self._config = FaissOnlineStoreConfig(**config.online_store.dict())
+        if self._index is None or not partial:
+            quantizer = faiss.IndexFlatL2(dimension)
+            self._index = faiss.IndexIVFFlat(quantizer, dimension, self._config.nlist)
+            self._index.train(
+                np.random.rand(self._config.nlist * 100, dimension).astype(np.float32)
+            )
+            self._in_memory_store = InMemoryStore()
+
+        self._in_memory_store.update(feature_names, {})
+
+    def teardown(
+        self,
+        config: RepoConfig,
+        tables: Sequence[FeatureView],
+        entities: Sequence[Entity],
+    ):
+        self._index = None
+        self._in_memory_store.teardown()
+
+    def online_read(
+        self,
+        config: RepoConfig,
+        table: FeatureView,
+        entity_keys: List[EntityKey],
+        requested_features: Optional[List[str]] = None,
+    ) -> List[Tuple[Optional[datetime], Optional[Dict[str, Value]]]]:
+        if self._index is None:
+            return [(None, None)] * len(entity_keys)
+
+        results: List[Tuple[Optional[datetime], Optional[Dict[str, Any]]]] = []
+        for entity_key in entity_keys:
+            serialized_key = serialize_entity_key(
+                entity_key, config.entity_key_serialization_version
+            ).hex()
+            idx = self._in_memory_store.entity_keys.get(serialized_key, -1)
+            if idx == -1:
+                results.append((None, None))
+            else:
+                feature_vector = self._index.reconstruct(int(idx))
+                feature_dict = {
+                    name: Value(double_val=value)
+                    for name, value in zip(
+                        self._in_memory_store.feature_names, feature_vector
+                    )
+                }
+                results.append((None, feature_dict))
+        return results
+
+    def online_write_batch(
+        self,
+        config: RepoConfig,
+        table: FeatureView,
+        data: List[Tuple[EntityKey, Dict[str, Value], datetime, Optional[datetime]]],
+        progress: Optional[Callable[[int], Any]],
+    ) -> None:
+        if self._index is None:
+            self._logger.warning("Index is not initialized. Skipping write operation.")
+            return
+
+        feature_vectors = []
+        serialized_keys = []
+
+        for entity_key, feature_dict, _, _ in data:
+            serialized_key = serialize_entity_key(
+                entity_key, config.entity_key_serialization_version
+            ).hex()
+            feature_vector = np.array(
+                [
+                    feature_dict[name].double_val
+                    for name in self._in_memory_store.feature_names
+                ],
+                dtype=np.float32,
+            )
+
+            feature_vectors.append(feature_vector)
+            serialized_keys.append(serialized_key)
+
+        feature_vectors_array = np.array(feature_vectors)
+
+        existing_indices = [
+            self._in_memory_store.entity_keys.get(sk, -1) for sk in serialized_keys
+        ]
+        mask = np.array(existing_indices) != -1
+        if np.any(mask):
+            self._index.remove_ids(
+                np.array([idx for idx in existing_indices if idx != -1])
+            )
+
+        new_indices = np.arange(
+            self._index.ntotal, self._index.ntotal + len(feature_vectors_array)
+        )
+        self._index.add(feature_vectors_array)
+
+        for sk, idx in zip(serialized_keys, new_indices):
+            self._in_memory_store.entity_keys[sk] = idx
+
+        if progress:
+            progress(len(data))
+
+    def retrieve_online_documents(
+        self,
+        config: RepoConfig,
+        table: FeatureView,
+        requested_feature: str,
+        embedding: List[float],
+        top_k: int,
+        distance_metric: Optional[str] = None,
+    ) -> List[
+        Tuple[
+            Optional[datetime],
+            Optional[Value],
+            Optional[Value],
+            Optional[Value],
+        ]
+    ]:
+        if self._index is None:
+            self._logger.warning("Index is not initialized. Returning empty result.")
+            return []
+
+        query_vector = np.array(embedding, dtype=np.float32).reshape(1, -1)
+        distances, indices = self._index.search(query_vector, top_k)
+
+        results: List[
+            Tuple[
+                Optional[datetime],
+                Optional[Value],
+                Optional[Value],
+                Optional[Value],
+            ]
+        ] = []
+        for i, idx in enumerate(indices[0]):
+            if idx == -1:
+                continue
+
+            feature_vector = self._index.reconstruct(int(idx))
+
+            timestamp = Timestamp()
+            timestamp.GetCurrentTime()
+
+            feature_value = Value(string_val=",".join(map(str, feature_vector)))
+            vector_value = Value(string_val=",".join(map(str, feature_vector)))
+            distance_value = Value(float_val=distances[0][i])
+
+            results.append(
+                (
+                    timestamp.ToDatetime(),
+                    feature_value,
+                    vector_value,
+                    distance_value,
+                )
+            )
+
+        return results
+
+    async def online_read_async(
+        self,
+        config: RepoConfig,
+        table: FeatureView,
+        entity_keys: List[EntityKey],
+        requested_features: Optional[List[str]] = None,
+    ) -> List[Tuple[Optional[datetime], Optional[Dict[str, Value]]]]:
+        # Implement async read if needed
+        raise NotImplementedError("Async read is not implemented for FaissOnlineStore")
diff --git a/setup.py b/setup.py
@@ -144,6 +144,8 @@
 
 MSSQL_REQUIRED = ["ibis-framework[mssql]>=9.0.0,<10"]
 
+FAISS_REQUIRED = ["faiss-cpu>=1.7.0,<2"]
+
 CI_REQUIRED = (
     [
         "build",
@@ -210,6 +212,7 @@
     + SQLITE_VEC_REQUIRED
     + SINGLESTORE_REQUIRED
     + OPENTELEMETRY
+    + FAISS_REQUIRED
 )
 
 DOCS_REQUIRED = CI_REQUIRED
@@ -279,6 +282,7 @@
         "sqlite_vec": SQLITE_VEC_REQUIRED,
         "singlestore": SINGLESTORE_REQUIRED,
         "opentelemetry": OPENTELEMETRY,
+        "faiss": FAISS_REQUIRED,
     },
     include_package_data=True,
     license="Apache",