Add support for serializing modules involved in LambdaOp execution by value (#1741)

* Allow users to specify module serialization hints

This commit adds an optional parameter to Workflow.save so that
users can indicate that certain modules should be serialized by value.
This is necessary if a LambdaOp in a workflow depends on a module
whose source file will not be available in the deployment environment.

Related to #1737.

* Adds automatic inference of LambdaOp module dependencies

This commit adds code to automatically infer LambdaOp module dependencies
in several common cases:

1. in which a function is passed to LambdaOp by name,
2. in which the argument to LambdaOp is a lambda expression that refers
   to a function by a fully-qualified name, and
3. in which the argument to LambdaOp is a lambda expression that refers
   to a function via another variable

The current implementation does not inspect the bodies of any function
passed to LambdaOp, and many corner cases are (necessarily) omitted.  However,
this support should be complete enough to be useful for many users.

Automatic inference is optional (via a parameter to Workflow.save) but it
could be the default in the future.

Related to issue #1737.

* Added tests related to issue #1737

* Fix linter errors

* Workflow.save: reset cloudpickle behavior changes on return

* aligned formatting with black's expectations

---------

Co-authored-by: Karl Higley <[email protected]>
Co-authored-by: rnyak <[email protected]>

Author: 3 people

nvtabular/workflow/workflow.py

@@ -13,10 +13,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
+
+import inspect
 import json
 import logging
 import sys
 import time
+import types
 import warnings
 from typing import TYPE_CHECKING, Optional
 
@@ -31,9 +34,10 @@
 
 from merlin.dag import Graph
 from merlin.dag.executors import DaskExecutor, LocalExecutor
+from merlin.dag.node import iter_nodes
 from merlin.io import Dataset
 from merlin.schema import Schema
-from nvtabular.ops import StatOperator
+from nvtabular.ops import LambdaOp, StatOperator
 from nvtabular.workflow.node import WorkflowNode
 
 LOG = logging.getLogger("nvtabular")
@@ -255,13 +259,68 @@ def _transform_df(self, df):
 
         return LocalExecutor().transform(df, self.output_node, self.output_dtypes)
 
-    def save(self, path):
+    @classmethod
+    def _getmodules(cls, fs):
+        """
+        Returns an imprecise but useful approximation of the list of modules
+        necessary to execute a given list of functions.  This approximation is
+        sound (all modules listed are required by the supplied functions) but not
+        necessarily complete (not all modules required will necessarily be returned).
+
+        For function literals (lambda expressions), this returns
+            1. the names of every module referenced in the lambda expression, e.g.,
+               `m` for `lambda x: m.f(x)` and
+            2. the names of the declaring module for every function referenced in
+               the lambda expression, e.g. `m` for `import m.f; lambda x: f(x)`
+
+        For declared functions, this returns the names of their declaring modules.
+
+        The return value will exclude all built-in modules and (on Python 3.10 or later)
+        all standard library modules.
+        """
+        result = set()
+
+        exclusions = set(sys.builtin_module_names)
+        if hasattr(sys, "stdlib_module_names"):
+            # sys.stdlib_module_names is only available in Python 3.10 and beyond
+            exclusions = exclusions | sys.stdlib_module_names
+
+        for f in fs:
+            if f.__name__ == "<lambda>":
+                for closurevars in [
+                    inspect.getclosurevars(f).globals,
+                    inspect.getclosurevars(f).nonlocals,
+                ]:
+                    for name, val in closurevars.items():
+                        print(f"{name} = {val}")
+                        if isinstance(val, types.ModuleType):
+                            result.add(val)
+                        elif isinstance(val, types.FunctionType):
+                            mod = inspect.getmodule(val)
+                            if mod is not None:
+                                result.add(mod)
+            else:
+                mod = inspect.getmodule(f)
+                if mod is not None:
+                    result.add(mod)
+
+        return [mod for mod in result if mod.__name__ not in exclusions]
+
+    def save(self, path, modules_byvalue=None):
         """Save this workflow to disk
 
         Parameters
         ----------
         path: str
             The path to save the workflow to
+        modules_byvalue:
+            A list of modules that should be serialized by value. This
+            should include any modules that will not be available on
+            the host where this workflow is ultimately deserialized.
+
+            In lieu of an explicit list, pass None to serialize all modules
+            by reference or pass "auto" to use a heuristic to infer which
+            modules to serialize by value.
         """
         # avoid a circular import getting the version
         from nvtabular import __version__ as nvt_version
@@ -290,9 +349,51 @@ def save(self, path):
                 o,
             )
 
-        # dump out the full workflow (graph/stats/operators etc) using cloudpickle
-        with fs.open(fs.sep.join([path, "workflow.pkl"]), "wb") as o:
-            cloudpickle.dump(self, o)
+        # track existing by-value modules
+        preexisting_modules_byvalue = set(cloudpickle.list_registry_pickle_by_value())
+
+        # direct cloudpickle to serialize selected modules by value
+        if modules_byvalue is None:
+            modules_byvalue = []
+        elif modules_byvalue == "auto":
+            l_nodes = self.graph.get_nodes_by_op_type(
+                list(iter_nodes([self.graph.output_node])), LambdaOp
+            )
+
+            try:
+                modules_byvalue = Workflow._getmodules([ln.op.f for ln in l_nodes])
+            except RuntimeError as ex:
+                warnings.warn(
+                    "Failed to automatically infer modules to serialize by value. "
+                    f'Reason given was "{str(ex)}"'
+                )
+
+        try:
+            for m in modules_byvalue:
+                if isinstance(m, types.ModuleType):
+                    cloudpickle.register_pickle_by_value(m)
+                elif isinstance(m, str) and m in sys.modules:
+                    cloudpickle.register_pickle_by_value(sys.modules[m])
+        except RuntimeError as ex:
+            warnings.warn(
+                f'Failed to register modules to serialize by value.  Reason given was "{str(ex)}"'
+            )
+
+        try:
+            # dump out the full workflow (graph/stats/operators etc) using cloudpickle
+            with fs.open(fs.sep.join([path, "workflow.pkl"]), "wb") as o:
+                cloudpickle.dump(self, o)
+        finally:
+            # return all modules that we set to serialize by value to by-reference
+            # (i.e., retain modules that were set to serialize by value before this invocation)
+
+            for m in modules_byvalue:
+                if isinstance(m, types.ModuleType):
+                    if m.__name__ not in preexisting_modules_byvalue:
+                        cloudpickle.unregister_pickle_by_value(m)
+                elif isinstance(m, str) and m in sys.modules:
+                    if m not in preexisting_modules_byvalue:
+                        cloudpickle.unregister_pickle_by_value(sys.modules[m])
 
     @classmethod
     def load(cls, path, client=None) -> "Workflow":

tests/unit/workflow/test_workflow.py

@@ -18,6 +18,7 @@
 import math
 import os
 import shutil
+import sys
 
 try:
     import cudf
@@ -666,3 +667,71 @@ def test_workflow_saved_schema(tmpdir):
     for node in postorder_iter_nodes(workflow2.output_node):
         assert node.input_schema is not None
         assert node.output_schema is not None
+
+
+def test_workflow_infer_modules_byvalue(tmp_path):
+    module_fn = tmp_path / "not_a_real_module.py"
+    sys.path.append(str(tmp_path))
+
+    with open(module_fn, "w") as module_f:
+        module_f.write("def identity(col):\n    return col")
+
+    import not_a_real_module
+
+    f_0 = not_a_real_module.identity
+    f_1 = lambda x: not_a_real_module.identity(x)  # noqa
+    f_2 = lambda x: f_0(x)  # noqa
+
+    try:
+        for fn, f in {
+            "not_a_real_module.identity": f_0,
+            "lambda x: not_a_real_module.identity(x)": f_1,
+            "lambda x: f_0(x)": f_2,
+        }.items():
+            assert not_a_real_module in Workflow._getmodules(
+                [f]
+            ), f"inferred module dependencies from {fn}"
+
+    finally:
+        sys.path.pop()
+        del sys.modules["not_a_real_module"]
+
+
+def test_workflow_explicit_modules_byvalue(tmp_path):
+    module_fn = tmp_path / "not_a_real_module.py"
+    sys.path.append(str(tmp_path))
+
+    with open(module_fn, "w") as module_f:
+        module_f.write("def identity(col):\n    return col")
+
+    import not_a_real_module
+
+    wf = nvt.Workflow(["col_a"] >> nvt.ops.LambdaOp(not_a_real_module.identity))
+
+    wf.save(str(tmp_path / "identity-workflow"), modules_byvalue=[not_a_real_module])
+
+    del not_a_real_module
+    del sys.modules["not_a_real_module"]
+    os.unlink(str(tmp_path / "not_a_real_module.py"))
+
+    Workflow.load(str(tmp_path / "identity-workflow"))
+
+
+def test_workflow_auto_infer_modules_byvalue(tmp_path):
+    module_fn = tmp_path / "not_a_real_module.py"
+    sys.path.append(str(tmp_path))
+
+    with open(module_fn, "w") as module_f:
+        module_f.write("def identity(col):\n    return col")
+
+    import not_a_real_module
+
+    wf = nvt.Workflow(["col_a"] >> nvt.ops.LambdaOp(not_a_real_module.identity))
+
+    wf.save(str(tmp_path / "identity-workflow"), modules_byvalue="auto")
+
+    del not_a_real_module
+    del sys.modules["not_a_real_module"]
+    os.unlink(str(tmp_path / "not_a_real_module.py"))
+
+    Workflow.load(str(tmp_path / "identity-workflow"))