Adds implicit dependency resolution on Python Workflows steps

Author: Caio-Nogueira

src/pyodide/internal/metadata.ts

@@ -58,6 +58,8 @@ const FORCE_NEW_VENDOR_PATH: boolean =
 export const IS_DEDICATED_SNAPSHOT_ENABLED: boolean =
   !!COMPATIBILITY_FLAGS.python_dedicated_snapshot;
 const EXTERNAL_SDK = !!COMPATIBILITY_FLAGS.enable_python_external_sdk;
+export const WORKFLOWS_IMPLICIT_DEPS =
+  !!COMPATIBILITY_FLAGS.python_workflows_implicit_dependencies;
 
 export const LEGACY_GLOBAL_HANDLERS = !NO_GLOBAL_HANDLERS;
 export const LEGACY_VENDOR_PATH = !FORCE_NEW_VENDOR_PATH;

src/pyodide/internal/workers-api/src/workers/_workers.py

@@ -1114,28 +1114,53 @@ def __init__(self, js_step):
         self._js_step = js_step
         self._memoized_dependencies = {}
         self._in_flight = {}
+        self.step_closures = {}
 
     def do(self, name, depends=None, concurrent=False, config=None):
         def decorator(func):
             async def wrapper():
+                # if implicit params are enabled, each param that is not context should be treated as a dependency and resolved
+                # In other words, we introspect the declaration and call a function (need to make sure it's a step) with the same name as the corresponding param
+                # This new code path should discard depends, as we encourage users to implicitly declare their invariant steps in the signature
+                # if the compat flag is disabled, then we just maintain the same legacy behavior
+                results_future_list = depends
+
+                if _pyodide_entrypoint_helper.workflowsImplicitDeps:
+                    sig = inspect.signature(func)
+                    results_future_list = []
+                    for p in sig.parameters.values():
+                        if p.name in self.step_closures:
+                            results_future_list.append(self.step_closures[p.name])
+                        elif p.name == "ctx":
+                            results_future_list.append(p)
+
                 if concurrent:
                     results = await gather(
-                        *[self._resolve_dependency(dep) for dep in depends or []]
+                        *[
+                            self._resolve_dependency(dep)
+                            for dep in results_future_list or []
+                        ]
                     )
                 else:
                     results = [
-                        await self._resolve_dependency(dep) for dep in depends or []
+                        await self._resolve_dependency(dep)
+                        for dep in results_future_list or []
                     ]
-                python_results = [python_from_rpc(result) for result in results]
+                python_results = [
+                    result
+                    if (hasattr(result, "name") and result.name == "ctx")
+                    else python_from_rpc(result)
+                    for result in results
+                ]
                 return await _do_call(self, name, config, func, *python_results)
 
             wrapper._step_name = name
+            self.step_closures[name] = wrapper
             return wrapper
 
         return decorator
 
     def sleep(self, *args, **kwargs):
-        # all types should be primitives - no need for explicit translation
         return self._js_step.sleep(*args, **kwargs)
 
     def sleep_until(self, name, timestamp):
@@ -1154,7 +1179,9 @@ def wait_for_event(self, name, event_type, /, timeout="24 hours"):
         )
 
     async def _resolve_dependency(self, dep):
-        if dep._step_name in self._memoized_dependencies:
+        if hasattr(dep, "name") and dep.name == "ctx":
+            return dep
+        elif dep._step_name in self._memoized_dependencies:
             return self._memoized_dependencies[dep._step_name]
         elif dep._step_name in self._in_flight:
             return await self._in_flight[dep._step_name]
@@ -1163,8 +1190,13 @@ async def _resolve_dependency(self, dep):
 
 
 async def _do_call(entrypoint, name, config, callback, *results):
-    async def _callback():
-        result = callback(*results)
+    async def _callback(ctx=None):
+        # deconstruct the actual ctx object
+        resolved_results = tuple(
+            python_from_rpc(ctx) if hasattr(r, "name") and r.name == "ctx" else r
+            for r in results
+        )
+        result = callback(*resolved_results)
 
         if inspect.iscoroutine(result):
             result = await result

src/pyodide/python-entrypoint-helper.ts

@@ -2,31 +2,35 @@
 // This file is a BUILTIN module that provides the actual implementation for the
 // python-entrypoint.js USER module.
 
-import {
-  beforeRequest,
-  loadPyodide,
-  clearSignals,
-} from 'pyodide-internal:python';
 import { enterJaegerSpan } from 'pyodide-internal:jaeger';
-import { patchLoadPackage } from 'pyodide-internal:setupPackages';
+import { default as Limiter } from 'pyodide-internal:limiter';
 import {
   IS_WORKERD,
+  LEGACY_GLOBAL_HANDLERS,
+  LEGACY_INCLUDE_SDK,
   LOCKFILE,
-  TRANSITIVE_REQUIREMENTS,
   MAIN_MODULE_NAME,
-  WORKERD_INDEX_URL,
   SHOULD_SNAPSHOT_TO_DISK,
+  TRANSITIVE_REQUIREMENTS,
+  WORKERD_INDEX_URL,
   WORKFLOWS_ENABLED,
-  LEGACY_GLOBAL_HANDLERS,
-  LEGACY_INCLUDE_SDK,
+  WORKFLOWS_IMPLICIT_DEPS,
 } from 'pyodide-internal:metadata';
-import { default as Limiter } from 'pyodide-internal:limiter';
 import {
-  PythonWorkersInternalError,
+  beforeRequest,
+  clearSignals,
+  loadPyodide,
+} from 'pyodide-internal:python';
+import { patchLoadPackage } from 'pyodide-internal:setupPackages';
+import {
+  LOADED_SNAPSHOT_TYPE,
+  maybeCollectDedicatedSnapshot,
+} from 'pyodide-internal:snapshot';
+import {
   PythonUserError,
+  PythonWorkersInternalError,
   reportError,
 } from 'pyodide-internal:util';
-import { LOADED_SNAPSHOT_TYPE } from 'pyodide-internal:snapshot';
 
 type PyFuture<T> = Promise<T> & { copy(): PyFuture<T>; destroy(): void };
 
@@ -69,8 +73,8 @@ export type PyodideEntrypointHelper = {
   cloudflareSocketsModule: any;
   workerEntrypoint: any;
   patchWaitUntil: typeof patchWaitUntil;
+  workflowsImplicitDeps: boolean;
 };
-import { maybeCollectDedicatedSnapshot } from 'pyodide-internal:snapshot';
 
 // Function to import JavaScript modules from Python
 let _pyodide_entrypoint_helper: PyodideEntrypointHelper | null = null;
@@ -96,6 +100,7 @@ export function setDoAnImport(
     cloudflareSocketsModule,
     workerEntrypoint,
     patchWaitUntil,
+    workflowsImplicitDeps: WORKFLOWS_IMPLICIT_DEPS,
   };
 }
 

src/pyodide/types/runtime-generated/metadata.d.ts

@@ -6,6 +6,7 @@ declare namespace MetadataReader {
     python_dedicated_snapshot?: boolean;
     enable_python_external_sdk?: boolean;
     python_check_rng_state?: boolean;
+    python_workflows_implicit_dependencies?: boolean;
   }
 
   const isWorkerd: () => boolean;

src/workerd/io/compatibility-date.capnp

@@ -1256,4 +1256,11 @@ struct CompatibilityFlags @0x8f8c1b68151b6cef {
     $experimental;
   # Enables precise timers with 3ms granularity. This provides more accurate timing for performance
   # measurements and time-sensitive operations.
+
+  pythonWorkflowsImplicitDeps @149 :Bool
+    $compatEnableFlag("python_workflows_implicit_dependencies")
+    $compatDisableFlag("no_python_workflows_implicit_dependencies")
+    $impliedByAfterDate(name = "pythonWorkers", date = "2026-01-15");
+  # replaces depends param on steps to an implicit approach with step callables passed as params
+  # these steps are called internally and act as dependencies
 }

src/workerd/server/tests/python/workflow-entrypoint/worker.js

@@ -2,9 +2,15 @@ import { RpcTarget } from 'cloudflare:workers';
 import * as assert from 'node:assert';
 
 class Context extends RpcTarget {
+  constructor(shouldSendCtx) {
+    super();
+    this.shouldSendCtx = shouldSendCtx;
+  }
+
   async do(name, fn) {
     try {
-      const result = await fn();
+      const ctx = { attempt: '1', metadata: 'expected_return_metadata' };
+      const result = this.shouldSendCtx ? await fn(ctx) : await fn();
       return result;
     } catch (e) {
       console.log(`Error received: ${e.name} Message: ${e.message}`);
@@ -16,10 +22,35 @@ class Context extends RpcTarget {
 
 export default {
   async test(ctrl, env, ctx) {
-    // JS types
-    const stubStep = new Context();
+    let stubStep = new Context(true);
+
+    // Tests forward compat - i.e.: python workflows should be compatible with steps that pass a ctx argument
+    // this param is optional and searched by name. Meaning it's not positional
+    let resp = await env.PythonWorkflow.run(
+      {
+        foo: 'bar',
+      },
+      stubStep
+    );
+    assert.deepStrictEqual(resp, 'foobar');
+
+    // Tests backwards compat - i.e.: new logic shouldn't break workflows until a release is done with
+    // ctx being passed as a step argument - this is not controlled via compat flag
+    // Moreover, this workflow also tests that dependencies are no longer resolved through the old code path
+    stubStep = new Context(false);
+    resp = await env.PythonWorkflowBackwardsCompat.run(
+      {
+        foo: 'bar',
+      },
+      stubStep
+    );
+    assert.deepStrictEqual(resp, 'foo');
 
-    const resp = await env.PythonWorkflow.run(
+    // Tests backwards compat for dependency resolution - previously deps were not following a name based
+    // approach. Instead, they were passed in the same order as they were declared in the depends param
+    // This test also doesn't pass any ctx to steps
+    stubStep = new Context(false);
+    resp = await env.PythonWorkflowDepends.run(
       {
         foo: 'bar',
       },

src/workerd/server/tests/python/workflow-entrypoint/workflow-entrypoint.wd-test

@@ -3,21 +3,37 @@ using Workerd = import "/workerd/workerd.capnp";
 const config :Workerd.Config = (
   services = [
     (name = "py", worker = .pyWorker),
+    (name = "pyOld", worker = .pyWorkerOld),
     (name = "js", worker = .jsWorker),
   ],
 );
 
 const pyWorker :Workerd.Worker = (
-  compatibilityDate = "2025-03-04",
+  compatibilityDate = "2025-12-04",
 
-  compatibilityFlags = [%PYTHON_FEATURE_FLAGS, "python_workflows"],
+  compatibilityFlags = [%PYTHON_FEATURE_FLAGS, "python_workflows", "python_workflows_implicit_dependencies"],
 
   modules = [
     (name = "workflow.py", pythonModule = embed "workflow.py"),
   ],
 
   bindings = [
     (name = "PythonWorkflow", service = (name = "py", entrypoint = "WorkflowEntrypointExample")),
+    (name = "PythonWorkflowBackwardsCompat", service = (name = "py", entrypoint = "WorkflowEntrypointBackwardsCompat")),
+  ],
+);
+
+const pyWorkerOld :Workerd.Worker = (
+  compatibilityDate = "2025-12-04",
+
+  compatibilityFlags = [%PYTHON_FEATURE_FLAGS, "python_workflows", "no_python_workflows_implicit_dependencies"],
+
+  modules = [
+    (name = "workflow-old.py", pythonModule = embed "workflow-old.py"),
+  ],
+
+  bindings = [
+    (name = "PythonWorkflowDepends", service = (name = "pyOld", entrypoint = "PythonWorkflowDepends")),
   ],
 );
 
@@ -32,5 +48,7 @@ const jsWorker :Workerd.Worker = (
 
   bindings = [
     (name = "PythonWorkflow", service = (name = "py", entrypoint = "WorkflowEntrypointExample")),
+    (name = "PythonWorkflowBackwardsCompat", service = (name = "py", entrypoint = "WorkflowEntrypointBackwardsCompat")),
+    (name = "PythonWorkflowDepends", service = (name = "pyOld", entrypoint = "PythonWorkflowDepends")),
   ],
 );

src/workerd/server/tests/python/workflow-entrypoint/workflow-old.py

@@ -0,0 +1,33 @@
+# Copyright (c) 2025 Cloudflare, Inc.
+# Licensed under the Apache 2.0 license found in the LICENSE file or at:
+#     https://opensource.org/licenses/Apache-2.0
+
+from workers import WorkflowEntrypoint
+
+
+class PythonWorkflowDepends(WorkflowEntrypoint):
+    # The purpose of this workflow is testing that depends are no longer resolved
+    #
+    async def run(self, event, step):
+        @step.do("step_1")
+        async def step_1():
+            # tests backwards compat with workflows that don't have ctx in the step callback
+            print("Executing step 1")
+            return "foo"
+
+        @step.do("step_2")
+        async def step_2():
+            print("Executing step 2")
+            return "bar"
+
+        @step.do("step_3", depends=[step_1, step_2], concurrent=True)
+        async def step_3(result1=(), result2=()):
+            assert result1 == "foo"
+            assert result2 == "bar"
+            return result1 + result2
+
+        return await step_3()
+
+
+async def test(ctrl, env, ctx):
+    pass