chore: Remove Bigframes installation in remote training custom job

PiperOrigin-RevId: 568668981

vertexai/preview/_workflow/executor/training.py

@@ -506,8 +506,6 @@ def remote_training(invokable: shared._Invokable, rewrapper: Any):
             VERTEX_AI_DEPENDENCY_PATH,
             "absl-py==1.4.0",
         ]
-    if bf:
-        vertex_requirements.append("bigframes==0.1.1")
 
     requirements = []
 

vertexai/preview/_workflow/serialization_engine/serializers.py

@@ -1124,7 +1124,7 @@ def serialize(
             raise ValueError(f"Invalid gcs path: {gcs_path}")
 
         BigframeSerializer._metadata.dependencies = (
-            supported_frameworks._get_deps_if_bigframe(to_serialize)
+            supported_frameworks._get_bigframe_deps()
         )
 
         # Check if index.name is default and set index.name if not

vertexai/preview/_workflow/shared/supported_frameworks.py

@@ -267,13 +267,7 @@ def _get_deps_if_pandas_dataframe(possible_dataframe: Any) -> List[str]:
     if _is_pandas_dataframe(possible_dataframe):
         dep_version = version.Version(pd.__version__).base_version
         deps.append(f"pandas=={dep_version}")
-        try:
-            import pyarrow as pa
-
-            pyarrow_version = version.Version(pa.__version__).base_version
-            deps.append(f"pyarrow=={pyarrow_version}")
-        except ImportError:
-            deps.append("pyarrow")
+        deps += _get_pyarrow_deps()
     # Note: it's likely that a DataFrame can be changed to other format, and
     # therefore needs to be serialized by CloudPickleSerializer. An example
     # is sklearn's Transformer.fit_transform() method, whose output is always
@@ -282,18 +276,26 @@ def _get_deps_if_pandas_dataframe(possible_dataframe: Any) -> List[str]:
     return deps
 
 
-def _get_deps_if_bigframe(possible_dataframe: Any) -> List[str]:
+def _get_bigframe_deps() -> List[str]:
     deps = []
-    if _is_bigframe(possible_dataframe):
-        dep_version = version.Version(bf.__version__).base_version
-        deps.append(f"bigframes=={dep_version}")
-
-    # Note: it's likely that a DataFrame can be changed to other format, and
-    # therefore needs to be serialized by CloudPickleSerializer. An example
-    # is sklearn's Transformer.fit_transform() method, whose output is always
-    # a ndarray.
-    deps += _get_cloudpickle_deps()
+    # Note: bigframe serialization can only occur locally so bigframes
+    # should not be installed remotely. Pandas and pyarrow are required
+    # to deserialize for sklearn bigframes though.
     deps += _get_pandas_deps()
+    deps += _get_pyarrow_deps()
+    return deps
+
+
+def _get_pyarrow_deps() -> List[str]:
+    deps = []
+    try:
+        global pyarrow
+        import pyarrow
+
+        dep_version = version.Version(pyarrow.__version__).base_version
+        deps.append(f"pyarrow=={dep_version}")
+    except ImportError:
+        deps.append("pyarrow")
     return deps