ntkathole · Apr 30, 2026 · Mar 3, 2026 · Mar 4, 2026 · Mar 9, 2026 · Mar 17, 2026
diff --git a/.secrets.baseline b/.secrets.baseline
diff --git a/pixi.lock b/pixi.lock
diff --git a/pyproject.toml b/pyproject.toml
@@ -153,7 +153,7 @@ test = [
 ]
 
 ci = [
-    "feast[test, aws, azure, cassandra, clickhouse, couchbase, delta, docling, duckdb, elasticsearch, faiss, gcp, ge, go, grpcio, hazelcast, hbase, ibis, image, k8s, mcp, milvus, mssql, mysql, openlineage, opentelemetry, oracle, spark, trino, postgres, pytorch, qdrant, rag, ray, redis, singlestore, snowflake, sqlite_vec]",
+    "feast[test, aws, azure, cassandra, clickhouse, couchbase, delta, docling, duckdb, elasticsearch, faiss, gcp, ge, go, grpcio, hazelcast, hbase, ibis, image, k8s, mcp, milvus, mongodb, mssql, mysql, openlineage, opentelemetry, oracle, spark, trino, postgres, pytorch, qdrant, rag, ray, redis, singlestore, snowflake, sqlite_vec]",
     "build",
     "virtualenv==20.23.0",
     "dbt-artifacts-parser",

@@ -0,0 +1,72 @@
+# MongoDB Offline Store
+
+This offline store lets you train models and run batch scoring directly from it.
+All feature views share a single collection (`feature_history`). Reads use
+MongoDB aggregation pipelines with a compound index, so per-entity cost is
+O(log n_observations) regardless of collection size, and K feature views with the same
+entity key collapse into one round-trip instead of K (1 if your data shares a unique id.)
+
+## Schema
+
+All feature views share one collection (default: `feature_history`), discriminated by the `feature_view` field.
+
+```javascript
+// Collection: feature_history
+{
+  "entity_id":       Binary("..."),           // Serialized entity key (bytes)
+  "feature_view":    "driver_stats",          // Discriminator
+  "features": {                               // Nested subdocument
+    "trips_today":   5,
+    "rating":        4.8
+  },
+  "event_timestamp": ISODate("2024-01-15T10:00:00Z"),
+  "created_at":      ISODate("2024-01-15T10:00:01Z")
+}
+```
+## Index
+
+The store creates one compound index lazily on first use. This index supports every query issued..
+
+```javascript
+db.feature_history.createIndex({
+  "entity_id":       1,
+  "feature_view":    1,
+  "event_timestamp": -1,
+  "created_at":      -1
+})
+
+```
+## Configuration
+
+```yaml
+offline_store:
+  type: feast.infra.offline_stores.contrib.mongodb_offline_store.mongodb.MongoDBOfflineStore
+  connection_string: mongodb://localhost:27017
+  database: feast
+  collection: feature_history  # optional, default: feature_history
+```
+
+## Key Features
+
+**Query-collapse** — Feature views that share the same join key set are grouped into a single MongoDB aggregation round-trip instead of one per feature view. Reduces round-trips from K to the number of unique join key signatures, often one.
+
+**Scoring path** — When `entity_df` contains unique entity IDs, a `$match + $sort + $group` pipeline performs server-side deduplication returning at most one document per `(entity_id, feature_view)`. The compound index makes per-entity cost O(log n_obs).
+
+**Training path** — When `entity_df` contains repeated entity IDs at different timestamps, the `$group` stage is omitted and `pandas.merge_asof` performs per-row point-in-time joins optimized in C.
+
+**`strict_pit`** — `get_historical_features` accepts a `strict_pit` keyword argument (default `True`). With `strict_pit=True` (default, safe for training), documents whose timestamp is strictly after the entity request timestamp are returned as `NULL`. Set `strict_pit=False` for real-time inference where you always want the most recent observation.
+
+
+## Writing Data
+
+Use `offline_write_batch` (called automatically by `feast materialize`) to write feature observations:
+
+```python
+store.write_to_offline_store(feature_view_name, df)
+```
+
+Documents are appended; `pull_latest` and the scoring path select the highest `created_at` at read time.
+
+## Memory Behaviour
+
+The store filters by entity key in `$match` rather than loading the entire collection. Memory usage is bounded by the number of unique entity IDs × documents per entity, not the total collection size.
@@ -0,0 +1,8 @@
+import feast.version
+
+try:
+    from pymongo.driver_info import DriverInfo
+
+    DRIVER_METADATA = DriverInfo(name="Feast", version=feast.version.get_version())
+except ImportError:
+    DRIVER_METADATA = None  # type: ignore[assignment]