feast-dev · ntkathole · Oct 15, 2025 · Oct 6, 2025 · Oct 10, 2025 · Oct 10, 2025
diff --git a/.github/workflows/pr_local_integration_tests.yml b/.github/workflows/pr_local_integration_tests.yml
@@ -42,6 +42,13 @@ jobs:
           enable-cache: true
       - name: Install dependencies
         run: make install-python-dependencies-ci
+      - name: Cache Hadoop tarball
+        uses: actions/cache@v4
+        with:
+          path: ~/hadoop-3.4.2.tar.gz
+          key: hadoop-3.4.2
+      - name: Install Hadoop dependencies
+        run: make install-hadoop-dependencies-ci
       - name: Test local integration tests
         if: ${{ always() }}  # this will guarantee that step won't be canceled and resources won't leak
         run: make test-python-integration-local

diff --git a/Makefile b/Makefile
@@ -87,7 +87,19 @@ install-python-dependencies-ci: ## Install Python CI dependencies in system envi
 	uv pip sync --system sdk/python/requirements/py$(PYTHON_VERSION)-ci-requirements.txt
 	uv pip install --system --no-deps -e .
 
-# Used by multicloud/Dockerfile.dev
+# Used in github actions/ci
+install-hadoop-dependencies-ci: ## Install Hadoop dependencies
+	@if [ ! -f $$HOME/hadoop-3.4.2.tar.gz ]; then \
+		echo "Downloading Hadoop tarball..."; \
+		wget -q https://dlcdn.apache.org/hadoop/common/hadoop-3.4.2/hadoop-3.4.2.tar.gz -O $$HOME/hadoop-3.4.2.tar.gz; \
+	else \
+		echo "Using cached Hadoop tarball"; \
+	fi
+	@if [ ! -d $$HOME/hadoop ]; then \
+		echo "Extracting Hadoop tarball..."; \
+		tar -xzf $$HOME/hadoop-3.4.2.tar.gz -C $$HOME; \
+		mv $$HOME/hadoop-3.4.2 $$HOME/hadoop; \
+	fi
 install-python-ci-dependencies: ## Install Python CI dependencies in system environment using piptools
 	python -m piptools sync sdk/python/requirements/py$(PYTHON_VERSION)-ci-requirements.txt
 	pip install --no-deps -e .
@@ -146,6 +158,9 @@ test-python-integration: ## Run Python integration tests (CI)
 test-python-integration-local: ## Run Python integration tests (local dev mode)
 	FEAST_IS_LOCAL_TEST=True \
 	FEAST_LOCAL_ONLINE_CONTAINER=True \
+	HADOOP_HOME=$$HOME/hadoop \
+	CLASSPATH="$$( $$HADOOP_HOME/bin/hadoop classpath --glob ):$$CLASSPATH" \
+	HADOOP_USER_NAME=root \
 	python -m pytest --tb=short -v -n 8 --color=yes --integration --durations=10 --timeout=1200 --timeout_method=thread --dist loadgroup \
 		-k "not test_lambda_materialization and not test_snowflake_materialization" \
 		-m "not rbac_remote_integration_test" \

@@ -26,6 +26,10 @@ Please see [Registry](../../getting-started/components/registry.md) for a concep
 [snowflake.md](snowflake.md)
 {% endcontent-ref %}
 
+{% content-ref url="hdfs.md" %}
+[hdfs.md](hdfs.md)
+{% endcontent-ref %}
+
 {% content-ref url="remote.md" %}
 [remote.md](remote.md)
 {% endcontent-ref %}
@@ -0,0 +1,42 @@
+# HDFS Registry
+
+## Description
+
+HDFS registry provides support for storing the protobuf representation of your feature store objects (data sources, feature views, feature services, etc.) in Hadoop Distributed File System (HDFS).
+
+While it can be used in production, there are still inherent limitations with a file-based registries, since changing a single field in the registry requires re-writing the whole registry file. With multiple concurrent writers, this presents a risk of data loss, or bottlenecks writes to the registry since all changes have to be serialized (e.g. when running materialization for multiple feature views or time ranges concurrently).
+
+### Pre-requisites
+
+The HDFS registry requires Hadoop 3.3+ to be installed and the `HADOOP_HOME` environment variable set.
+
+### Authentication and User Configuration
+
+The HDFS registry is using `pyarrow.fs.HadoopFileSystem` and **does not** support specifying HDFS users or Kerberos credentials directly in the `feature_store.yaml` configuration. It relies entirely on the Hadoop and system environment configuration available to the process running Feast.
+
+By default, `pyarrow.fs.HadoopFileSystem` inherits authentication from the underlying Hadoop client libraries and environment variables, such as:
+
+- `HADOOP_USER_NAME`
+- `KRB5CCNAME`
+- `hadoop.security.authentication`
+- Any other relevant properties in `core-site.xml` and `hdfs-site.xml`
+
+For more information, refer to:
+- [pyarrow.fs.HadoopFileSystem API Reference](https://arrow.apache.org/docs/python/generated/pyarrow.fs.HadoopFileSystem.html)
+- [Hadoop Security: Simple & Kerberos Authentication](https://hadoop.apache.org/docs/stable/hadoop-project-dist/hadoop-common/SecureMode.html)
+
+## Example
+
+An example of how to configure this would be:
+
+{% code title="feature_store.yaml" %}
+```yaml
+project: feast_hdfs
+registry:
+  path: hdfs://[YOUR NAMENODE HOST]:[YOUR NAMENODE PORT]/[PATH TO REGISTRY]/registry.pb
+  cache_ttl_seconds: 60
+online_store: null
+offline_store: null
+```
+{% endcode %}
+
@@ -0,0 +1,121 @@
+import json
+import uuid
+from pathlib import Path, PurePosixPath
+from typing import Optional
+from urllib.parse import urlparse
+
+from pyarrow import fs
+
+from feast.infra.registry.registry_store import RegistryStore
+from feast.protos.feast.core.Registry_pb2 import Registry as RegistryProto
+from feast.repo_config import RegistryConfig
+from feast.utils import _utc_now
+
+
+class HDFSRegistryStore(RegistryStore):
+    """HDFS implementation of RegistryStore.
+    registryConfig.path should be a hdfs path like hdfs://namenode:8020/path/to/registry.db
+    """
+
+    def __init__(self, registry_config: RegistryConfig, repo_path: Path):
+        try:
+            from pyarrow.fs import HadoopFileSystem
+        except ImportError as e:
+            from feast.errors import FeastExtrasDependencyImportError
+
+            raise FeastExtrasDependencyImportError(
+                "pyarrow.fs.HadoopFileSystem", str(e)
+            )
+        uri = registry_config.path
+        self._uri = urlparse(uri)
+        if self._uri.scheme != "hdfs":
+            raise ValueError(
+                f"Unsupported scheme {self._uri.scheme} in HDFS path {uri}"
+            )
+        self._hdfs = HadoopFileSystem(self._uri.hostname, self._uri.port or 8020)
+        self._path = PurePosixPath(self._uri.path)
+
+    def get_registry_proto(self):
+        registry_proto = RegistryProto()
+        if _check_hdfs_path_exists(self._hdfs, str(self._path)):
+            with self._hdfs.open_input_file(str(self._path)) as f:
+                registry_proto.ParseFromString(f.read())
+            return registry_proto
+        raise FileNotFoundError(
+            f'Registry not found at path "{self._uri.geturl()}". Have you run "feast apply"?'
+        )
+
+    def update_registry_proto(self, registry_proto: RegistryProto):
+        self._write_registry(registry_proto)
+
+    def teardown(self):
+        if _check_hdfs_path_exists(self._hdfs, str(self._path)):
+            self._hdfs.delete_file(str(self._path))
+        else:
+            # Nothing to do
+            pass
+
+    def _write_registry(self, registry_proto: RegistryProto):
+        """Write registry protobuf to HDFS."""
+        registry_proto.version_id = str(uuid.uuid4())
+        registry_proto.last_updated.FromDatetime(_utc_now())
+
+        dir_path = self._path.parent
+        if not _check_hdfs_path_exists(self._hdfs, str(dir_path)):
+            self._hdfs.create_dir(str(dir_path), recursive=True)
+
+        with self._hdfs.open_output_stream(str(self._path)) as f:
+            f.write(registry_proto.SerializeToString())
+
+    def set_project_metadata(self, project: str, key: str, value: str):
+        """Set a custom project metadata key-value pair in the registry (HDFS backend)."""
+        registry_proto = self.get_registry_proto()
+        found = False
+
+        for pm in registry_proto.project_metadata:
+            if pm.project == project:
+                # Load JSON metadata from project_uuid
+                try:
+                    meta = json.loads(pm.project_uuid) if pm.project_uuid else {}
+                except Exception:
+                    meta = {}
+
+                if not isinstance(meta, dict):
+                    meta = {}
+
+                meta[key] = value
+                pm.project_uuid = json.dumps(meta)
+                found = True
+                break
+
+        if not found:
+            # Create new ProjectMetadata entry
+            from feast.project_metadata import ProjectMetadata
+
+            pm = ProjectMetadata(project_name=project)
+            pm.project_uuid = json.dumps({key: value})
+            registry_proto.project_metadata.append(pm.to_proto())
+
+        # Write back
+        self.update_registry_proto(registry_proto)
+
+    def get_project_metadata(self, project: str, key: str) -> Optional[str]:
+        """Get custom project metadata key from registry (HDFS backend)."""
+        registry_proto = self.get_registry_proto()
+
+        for pm in registry_proto.project_metadata:
+            if pm.project == project:
+                try:
+                    meta = json.loads(pm.project_uuid) if pm.project_uuid else {}
+                except Exception:
+                    meta = {}
+
+                if not isinstance(meta, dict):
+                    return None
+                return meta.get(key, None)
+        return None
+
+
+def _check_hdfs_path_exists(hdfs, path: str) -> bool:
+    info = hdfs.get_file_info([path])[0]
+    return info.type != fs.FileType.NotFound
@@ -62,12 +62,14 @@
     "S3RegistryStore": "feast.infra.registry.s3.S3RegistryStore",
     "FileRegistryStore": "feast.infra.registry.file.FileRegistryStore",
     "AzureRegistryStore": "feast.infra.registry.contrib.azure.azure_registry_store.AzBlobRegistryStore",
+    "HDFSRegistryStore": "feast.infra.registry.contrib.hdfs.hdfs_registry_store.HDFSRegistryStore",
 }
 
 REGISTRY_STORE_CLASS_FOR_SCHEME = {
     "gs": "GCSRegistryStore",
     "s3": "S3RegistryStore",
     "file": "FileRegistryStore",
+    "hdfs": "HDFSRegistryStore",
     "": "FileRegistryStore",
 }
 
@@ -143,7 +145,7 @@ def get_registry_store_class_from_scheme(registry_path: str):
     if uri.scheme not in REGISTRY_STORE_CLASS_FOR_SCHEME:
         raise Exception(
             f"Registry path {registry_path} has unsupported scheme {uri.scheme}. "
-            f"Supported schemes are file, s3 and gs."
+            f"Supported schemes are file, s3, gs and hdfs."
         )
     else:
         registry_store_type = REGISTRY_STORE_CLASS_FOR_SCHEME[uri.scheme]

@@ -22,8 +22,12 @@
 
 import grpc_testing
 import pandas as pd
+import pyarrow.fs as fs
 import pytest
 from pytest_lazyfixture import lazy_fixture
+from testcontainers.core.container import DockerContainer
+from testcontainers.core.network import Network
+from testcontainers.core.waiting_utils import wait_for_logs
 from testcontainers.mysql import MySqlContainer
 from testcontainers.postgres import PostgresContainer
 
@@ -280,6 +284,60 @@ def sqlite_registry():
     yield SqlRegistry(registry_config, "project", None)
 
 
+@pytest.fixture(scope="function")
+def hdfs_registry():
+    HADOOP_NAMENODE_IMAGE = "bde2020/hadoop-namenode:2.0.0-hadoop3.2.1-java8"
+    HADOOP_DATANODE_IMAGE = "bde2020/hadoop-datanode:2.0.0-hadoop3.2.1-java8"
+    HDFS_CLUSTER_NAME = "feast-hdfs-cluster"
+    HADOOP_NAMENODE_WAIT_LOG = "namenode.NameNode: NameNode RPC up"
+    HADOOP_DATANODE_WAIT_LOG = "datanode.DataNode: .*successfully registered with NN"
+    with Network() as network:
+        namenode = None
+        datanode = None
+
+        try:
+            namenode = (
+                DockerContainer(HADOOP_NAMENODE_IMAGE)
+                .with_network(network)
+                .with_env("CLUSTER_NAME", HDFS_CLUSTER_NAME)
+                .with_exposed_ports(8020)
+                .with_network_aliases("namenode")
+                .with_kwargs(hostname="namenode")
+                .start()
+            )
+            wait_for_logs(namenode, HADOOP_NAMENODE_WAIT_LOG, timeout=120)
+            namenode_ip = namenode.get_container_host_ip()
+            namenode_port = int(namenode.get_exposed_port(8020))
+
+            datanode = (
+                DockerContainer(HADOOP_DATANODE_IMAGE)
+                .with_network(network)
+                .with_exposed_ports(9867)
+                .with_env("CLUSTER_NAME", HDFS_CLUSTER_NAME)
+                .with_env("CORE_CONF_fs_defaultFS", "hdfs://namenode:8020")
+                .with_network_aliases("datanode")
+                .with_kwargs(hostname="datanode")
+                .start()
+            )
+
+            wait_for_logs(datanode, HADOOP_DATANODE_WAIT_LOG, timeout=120)
+
+            hdfs = fs.HadoopFileSystem(host=namenode_ip, port=namenode_port)
+            hdfs.create_dir("/feast")
+            registry_path = f"hdfs://{namenode_ip}:{namenode_port}/feast/registry.db"
+            with hdfs.open_output_stream(registry_path) as f:
+                f.write(b"")
+
+            registry_config = RegistryConfig(path=registry_path, cache_ttl_seconds=600)
+            reg = Registry("project", registry_config, None)
+            yield reg
+        finally:
+            if datanode:
+                datanode.stop()
+            if namenode:
+                namenode.stop()
+
+
 class GrpcMockChannel:
     def __init__(self, service, servicer):
         self.service = service
@@ -350,6 +408,10 @@ def mock_remote_registry():
             lazy_fixture("mock_remote_registry"),
             marks=pytest.mark.rbac_remote_integration_test,
         ),
+        pytest.param(
+            lazy_fixture("hdfs_registry"),
+            marks=pytest.mark.xdist_group(name="hdfs_registry"),
+        ),
     ]
 
 sql_fixtures = [