search

temporalio / sdk-python

Temporal Python SDK
MIT License
470 stars 74 forks source link

readme

Temporal Python SDK

Python 3.8+ PyPI MIT

Temporal is a distributed, scalable, durable, and highly available orchestration engine used to execute asynchronous, long-running business logic in a scalable and resilient way.

"Temporal Python SDK" is the framework for authoring workflows and activities using the Python programming language.

Also see:

In addition to features common across all Temporal SDKs, the Python SDK also has the following interesting features:

Type Safe

This library uses the latest typing and MyPy support with generics to ensure all calls can be typed. For example, starting a workflow with an int parameter when it accepts a str parameter would cause MyPy to fail.

Different Activity Types

The activity worker has been developed to work with async def, threaded, and multiprocess activities. While async def activities are the easiest and recommended, care has been taken to make heartbeating and cancellation also work across threads/processes.

Custom asyncio Event Loop

The workflow implementation basically turns async def functions into workflows backed by a distributed, fault-tolerant event loop. This means task management, sleep, cancellation, etc have all been developed to seamlessly integrate with asyncio concepts.

See the blog post introducing the Python SDK for an informal introduction to the features and their implementation.

Contents

Quick Start

We will guide you through the Temporal basics to create a "hello, world!" script on your machine. It is not intended as one of the ways to use Temporal, but in reality it is very simplified and decidedly not "the only way" to use Temporal. For more information, check out the docs references in "Next Steps" below the quick start.

Installation

Install the temporalio package from PyPI.

These steps can be followed to use with a virtual environment and pip:

The SDK is now ready for use. To build from source, see "Building" near the end of this documentation.

NOTE: This README is for the current branch and not necessarily what's released on PyPI.

Implementing a Workflow

Create the following in activities.py:

from temporalio import activity

@activity.defn
def say_hello(name: str) -> str:
    return f"Hello, {name}!"

Create the following in workflows.py:

from datetime import timedelta
from temporalio import workflow

# Import our activity, passing it through the sandbox
with workflow.unsafe.imports_passed_through():
    from .activities import say_hello

@workflow.defn
class SayHello:
    @workflow.run
    async def run(self, name: str) -> str:
        return await workflow.execute_activity(
            say_hello, name, schedule_to_close_timeout=timedelta(seconds=5)
        )

Create the following in run_worker.py:

import asyncio
import concurrent.futures
from temporalio.client import Client
from temporalio.worker import Worker

# Import the activity and workflow from our other files
from .activities import say_hello
from .workflows import SayHello

async def main():
    # Create client connected to server at the given address
    client = await Client.connect("localhost:7233")

    # Run the worker
    with concurrent.futures.ThreadPoolExecutor(max_workers=100) as activity_executor:
        worker = Worker(
          client,
          task_queue="my-task-queue",
          workflows=[SayHello],
          activities=[say_hello],
          activity_executor=activity_executor,
        )
        await worker.run()

if __name__ == "__main__":
    asyncio.run(main())

Assuming you have a Temporal server running on localhost, this will run the worker:

python run_worker.py

Running a Workflow

Create the following script at run_workflow.py:

import asyncio
from temporalio.client import Client

# Import the workflow from the previous code
from .workflows import SayHello

async def main():
    # Create client connected to server at the given address
    client = await Client.connect("localhost:7233")

    # Execute a workflow
    result = await client.execute_workflow(SayHello.run, "my name", id="my-workflow-id", task_queue="my-task-queue")

    print(f"Result: {result}")

if __name__ == "__main__":
    asyncio.run(main())

Assuming you have run_worker.py running from before, this will run the workflow:

python run_workflow.py

The output will be:

Result: Hello, my-name!

Next Steps

Temporal can be implemented in your code in many different ways, to suit your application's needs. The links below will give you much more information about how Temporal works with Python:

Usage

From here, you will find reference documentation about specific pieces of the Temporal Python SDK that were built around Temporal concepts. This section is not intended as a how-to guide -- For more how-to oriented information, check out the links in the Next Steps section above.

Client

A client can be created and used to start a workflow like so:

from temporalio.client import Client

async def main():
    # Create client connected to server at the given address and namespace
    client = await Client.connect("localhost:7233", namespace="my-namespace")

    # Start a workflow
    handle = await client.start_workflow(MyWorkflow.run, "some arg", id="my-workflow-id", task_queue="my-task-queue")

    # Wait for result
    result = await handle.result()
    print(f"Result: {result}")

Some things to note about the above code:

Clients also provide a shallow copy of their config for use in making slightly different clients backed by the same connection. For instance, given the client above, this is how to have a client in another namespace:

config = client.config()
config["namespace"] = "my-other-namespace"
other_ns_client = Client(**config)

Data Conversion

Data converters are used to convert raw Temporal payloads to/from actual Python types. A custom data converter of type temporalio.converter.DataConverter can be set via the data_converter client parameter. Data converters are a combination of payload converters, payload codecs, and failure converters. Payload converters convert Python values to/from serialized bytes. Payload codecs convert bytes to bytes (e.g. for compression or encryption). Failure converters convert exceptions to/from serialized failures.

The default data converter supports converting multiple types including:

This notably doesn't include any date, time, or datetime objects as they may not work across SDKs.

Users are strongly encouraged to use a single dataclass for parameter and return types so fields with defaults can be easily added without breaking compatibility.

Classes with generics may not have the generics properly resolved. The current implementation does not have generic type resolution. Users should use concrete types.

Custom Type Data Conversion

For converting from JSON, the workflow/activity type hint is taken into account to convert to the proper type. Care has been taken to support all common typings including Optional, Union, all forms of iterables and mappings, NewType, etc in addition to the regular JSON values mentioned before.

Data converters contain a reference to a payload converter class that is used to convert to/from payloads/values. This is a class and not an instance because it is instantiated on every workflow run inside the sandbox. The payload converter is usually a CompositePayloadConverter which contains a multiple EncodingPayloadConverters it uses to try to serialize/deserialize payloads. Upon serialization, each EncodingPayloadConverter is tried until one succeeds. The EncodingPayloadConverter provides an "encoding" string serialized onto the payload so that, upon deserialization, the specific EncodingPayloadConverter for the given "encoding" is used.

The default data converter uses the DefaultPayloadConverter which is simply a CompositePayloadConverter with a known set of default EncodingPayloadConverters. To implement a custom encoding for a custom type, a new EncodingPayloadConverter can be created for the new type. For example, to support IPv4Address types:

class IPv4AddressEncodingPayloadConverter(EncodingPayloadConverter):
    @property
    def encoding(self) -> str:
        return "text/ipv4-address"

    def to_payload(self, value: Any) -> Optional[Payload]:
        if isinstance(value, ipaddress.IPv4Address):
            return Payload(
                metadata={"encoding": self.encoding.encode()},
                data=str(value).encode(),
            )
        else:
            return None

    def from_payload(self, payload: Payload, type_hint: Optional[Type] = None) -> Any:
        assert not type_hint or type_hint is ipaddress.IPv4Address
        return ipaddress.IPv4Address(payload.data.decode())

class IPv4AddressPayloadConverter(CompositePayloadConverter):
    def __init__(self) -> None:
        # Just add ours as first before the defaults
        super().__init__(
            IPv4AddressEncodingPayloadConverter(),
            *DefaultPayloadConverter.default_encoding_payload_converters,
        )

my_data_converter = dataclasses.replace(
    DataConverter.default,
    payload_converter_class=IPv4AddressPayloadConverter,
)

Imports are left off for brevity.

This is good for many custom types. However, sometimes you want to override the behavior of the just the existing JSON encoding payload converter to support a new type. It is already the last encoding data converter in the list, so it's the fall-through behavior for any otherwise unknown type. Customizing the existing JSON converter has the benefit of making the type work in lists, unions, etc.

The JSONPlainPayloadConverter uses the Python json library with an advanced JSON encoder by default and a custom value conversion method to turn json.loaded values to their type hints. The conversion can be customized for serialization with a custom json.JSONEncoder and deserialization with a custom JSONTypeConverter. For example, to support IPv4Address types in existing JSON conversion:

class IPv4AddressJSONEncoder(AdvancedJSONEncoder):
    def default(self, o: Any) -> Any:
        if isinstance(o, ipaddress.IPv4Address):
            return str(o)
        return super().default(o)
class IPv4AddressJSONTypeConverter(JSONTypeConverter):
    def to_typed_value(
        self, hint: Type, value: Any
    ) -> Union[Optional[Any], _JSONTypeConverterUnhandled]:
        if issubclass(hint, ipaddress.IPv4Address):
            return ipaddress.IPv4Address(value)
        return JSONTypeConverter.Unhandled

class IPv4AddressPayloadConverter(CompositePayloadConverter):
    def __init__(self) -> None:
        # Replace default JSON plain with our own that has our encoder and type
        # converter
        json_converter = JSONPlainPayloadConverter(
            encoder=IPv4AddressJSONEncoder,
            custom_type_converters=[IPv4AddressJSONTypeConverter()],
        )
        super().__init__(
            *[
                c if not isinstance(c, JSONPlainPayloadConverter) else json_converter
                for c in DefaultPayloadConverter.default_encoding_payload_converters
            ]
        )

my_data_converter = dataclasses.replace(
    DataConverter.default,
    payload_converter_class=IPv4AddressPayloadConverter,
)

Now IPv4Address can be used in type hints including collections, optionals, etc.

Workers

Workers host workflows and/or activities. Here's how to run a worker:

import asyncio
import logging
from temporalio.client import Client
from temporalio.worker import Worker
# Import your own workflows and activities
from my_workflow_package import MyWorkflow, my_activity

async def run_worker(stop_event: asyncio.Event):
    # Create client connected to server at the given address
    client = await Client.connect("localhost:7233", namespace="my-namespace")

    # Run the worker until the event is set
    worker = Worker(client, task_queue="my-task-queue", workflows=[MyWorkflow], activities=[my_activity])
    async with worker:
        await stop_event.wait()

Some things to note about the above code:

Workflows

Definition

Workflows are defined as classes decorated with @workflow.defn. The method invoked for the workflow is decorated with @workflow.run. Methods for signals, queries, and updates are decorated with @workflow.signal, @workflow.query and @workflow.update respectively. Here's an example of a workflow:

import asyncio
from datetime import timedelta
from temporalio import workflow

# Pass the activities through the sandbox
with workflow.unsafe.imports_passed_through():
    from .my_activities import GreetingInfo, create_greeting_activity

@workflow.defn
class GreetingWorkflow:
    def __init__(self) -> None:
        self._current_greeting = "<unset>"
        self._greeting_info = GreetingInfo()
        self._greeting_info_update = asyncio.Event()
        self._complete = asyncio.Event()

    @workflow.run
    async def run(self, name: str) -> str:
        self._greeting_info.name = name
        while True:
            # Store greeting
            self._current_greeting = await workflow.execute_activity(
                create_greeting_activity,
                self._greeting_info,
                start_to_close_timeout=timedelta(seconds=5),
            )
            workflow.logger.debug("Greeting set to %s", self._current_greeting)

            # Wait for salutation update or complete signal (this can be
            # cancelled)
            await asyncio.wait(
                [
                    asyncio.create_task(self._greeting_info_update.wait()),
                    asyncio.create_task(self._complete.wait()),
                ],
                return_when=asyncio.FIRST_COMPLETED,
            )
            if self._complete.is_set():
                return self._current_greeting
            self._greeting_info_update.clear()

    @workflow.signal
    async def update_salutation(self, salutation: str) -> None:
        self._greeting_info.salutation = salutation
        self._greeting_info_update.set()

    @workflow.signal
    async def complete_with_greeting(self) -> None:
        self._complete.set()

    @workflow.query
    def current_greeting(self) -> str:
        return self._current_greeting

    @workflow.update
    def set_and_get_greeting(self, greeting: str) -> str:
      old = self._current_greeting
      self._current_greeting = greeting
      return old

This assumes there's an activity in my_activities.py like:

from dataclasses import dataclass
from temporalio import workflow

@dataclass
class GreetingInfo:
    salutation: str = "Hello"
    name: str = "<unknown>"

@activity.defn
def create_greeting_activity(info: GreetingInfo) -> str:
    return f"{info.salutation}, {info.name}!"

Some things to note about the above workflow code:

Here are the decorators that can be applied:

Running

To start a locally-defined workflow from a client, you can simply reference its method like so:

from temporalio.client import Client
from my_workflow_package import GreetingWorkflow

async def create_greeting(client: Client) -> str:
    # Start the workflow
    handle = await client.start_workflow(GreetingWorkflow.run, "my name", id="my-workflow-id", task_queue="my-task-queue")
    # Change the salutation
    await handle.signal(GreetingWorkflow.update_salutation, "Aloha")
    # Tell it to complete
    await handle.signal(GreetingWorkflow.complete_with_greeting)
    # Wait and return result
    return await handle.result()

Some things to note about the above code:

Invoking Activities

Invoking Child Workflows

Timers

Conditions

Asyncio and Determinism

Workflows must be deterministic. Workflows are backed by a custom asyncio event loop. This means many of the common asyncio calls work as normal. Some asyncio features are disabled such as:

Also, there are some asyncio utilities that internally use set() which can make them non-deterministic from one worker to the next. Therefore the following asyncio functions have workflow-module alternatives that are deterministic:

Asyncio Cancellation

Cancellation is done using asyncio task cancellation. This means that tasks are requested to be cancelled but can catch the asyncio.CancelledError, thus allowing them to perform some cleanup before allowing the cancellation to proceed (i.e. re-raising the error), or to deny the cancellation entirely. It also means that asyncio.shield() can be used to protect tasks against cancellation.

The following tasks, when cancelled, perform a Temporal cancellation:

When the workflow itself is requested to cancel, Task.cancel is called on the main workflow task. Therefore, asyncio.CancelledError can be caught in order to handle the cancel gracefully.

Workflows follow asyncio cancellation rules exactly which can cause confusion among Python developers. Cancelling a task doesn't always cancel the thing it created. For example, given task = asyncio.create_task(workflow.start_child_workflow(..., calling task.cancel does not cancel the child workflow, it only cancels the starting of it, which has no effect if it has already started. However, cancelling the result of handle = await workflow.start_child_workflow(... or task = asyncio.create_task(workflow.execute_child_workflow(... does cancel the child workflow.

Also, due to Temporal rules, a cancellation request is a state not an event. Therefore, repeated cancellation requests are not delivered, only the first. If the workflow chooses swallow a cancellation, it cannot be requested again.

Workflow Utilities

While running in a workflow, in addition to features documented elsewhere, the following items are available from the temporalio.workflow package:

Exceptions

This default can be changed by providing a list of exception types to workflow_failure_exception_types when creating a Worker or failure_exception_types on the @workflow.defn decorator. If a workflow-thrown exception is an instance of any type in either list, it will fail the workflow (or update) instead of the workflow task. This means a value of [Exception] will cause every exception to fail the workflow instead of the workflow task. Also, as a special case, if temporalio.workflow.NondeterminismError (or any superclass of it) is set, non-deterministic exceptions will fail the workflow. WARNING: These settings are experimental.

Signal and update handlers

Signal and update handlers are defined using decorated methods as shown in the example above. Client code sends signals and updates using workflow_handle.signal, workflow_handle.execute_update, or workflow_handle.start_update. When the workflow receives one of these requests, it starts an asyncio.Task executing the corresponding handler method with the argument(s) from the request.

The handler methods may be async def and can do all the async operations described above (e.g. invoking activities and child workflows, and waiting on timers and conditions). Notice that this means that handler tasks will be executing concurrently with respect to each other and the main workflow task. Use asyncio.Lock and asyncio.Semaphore if necessary.

Your main workflow task may finish as a result of successful completion, cancellation, continue-as-new, or failure. You should ensure that all in-progress signal and update handler tasks have finished before this happens; if you do not, you will see a warning (the warning can be disabled via the workflow.signal/workflow.update decorators). One way to ensure that handler tasks have finished is to wait on the workflow.all_handlers_finished condition:

await workflow.wait_condition(workflow.all_handlers_finished)

External Workflows

Testing

Workflow testing can be done in an integration-test fashion against a real server, however it is hard to simulate timeouts and other long time-based code. Using the time-skipping workflow test environment can help there.

The time-skipping temporalio.testing.WorkflowEnvironment can be created via the static async start_time_skipping(). This internally downloads the Temporal time-skipping test server to a temporary directory if it doesn't already exist, then starts the test server which has special APIs for skipping time.

NOTE: The time-skipping test environment does not work on ARM. The SDK will try to download the x64 binary on macOS for use with the Intel emulator, but for Linux or Windows ARM there is no proper time-skipping test server at this time.

Automatic Time Skipping

Anytime a workflow result is waited on, the time-skipping server automatically advances to the next event it can. To manually advance time before waiting on the result of a workflow, the WorkflowEnvironment.sleep method can be used.

Here's a simple example of a workflow that sleeps for 24 hours:

import asyncio
from temporalio import workflow

@workflow.defn
class WaitADayWorkflow:
    @workflow.run
    async def run(self) -> str:
        await asyncio.sleep(24 * 60 * 60)
        return "all done"

An integration test of this workflow would be way too slow. However the time-skipping server automatically skips to the next event when we wait on the result. Here's a test for that workflow:

from temporalio.testing import WorkflowEnvironment
from temporalio.worker import Worker

async def test_wait_a_day_workflow():
    async with await WorkflowEnvironment.start_time_skipping() as env:
        async with Worker(env.client, task_queue="tq1", workflows=[WaitADayWorkflow]):
            assert "all done" == await env.client.execute_workflow(WaitADayWorkflow.run, id="wf1", task_queue="tq1")

That test will run almost instantly. This is because by calling execute_workflow on our client, we have asked the environment to automatically skip time as much as it can (basically until the end of the workflow or until an activity is run).

To disable automatic time-skipping while waiting for a workflow result, run code inside a with env.auto_time_skipping_disabled(): block.

Manual Time Skipping

Until a workflow is waited on, all time skipping in the time-skipping environment is done manually via WorkflowEnvironment.sleep.

Here's workflow that waits for a signal or times out:

import asyncio
from temporalio import workflow

@workflow.defn
class SignalWorkflow:
    def __init__(self) -> None:
        self.signal_received = False

    @workflow.run
    async def run(self) -> str:
        # Wait for signal or timeout in 45 seconds
        try:
            await workflow.wait_condition(lambda: self.signal_received, timeout=45)
            return "got signal"
        except asyncio.TimeoutError:
            return "got timeout"

    @workflow.signal
    def some_signal(self) -> None:
        self.signal_received = True

To test a normal signal, you might:

from temporalio.testing import WorkflowEnvironment
from temporalio.worker import Worker

async def test_signal_workflow():
    async with await WorkflowEnvironment.start_time_skipping() as env:
        async with Worker(env.client, task_queue="tq1", workflows=[SignalWorkflow]):
            # Start workflow, send signal, check result
            handle = await env.client.start_workflow(SignalWorkflow.run, id="wf1", task_queue="tq1")
            await handle.signal(SignalWorkflow.some_signal)
            assert "got signal" == await handle.result()

But how would you test the timeout part? Like so:

from temporalio.testing import WorkflowEnvironment
from temporalio.worker import Worker

async def test_signal_workflow_timeout():
    async with await WorkflowEnvironment.start_time_skipping() as env:
        async with Worker(env.client, task_queue="tq1", workflows=[SignalWorkflow]):
            # Start workflow, advance time past timeout, check result
            handle = await env.client.start_workflow(SignalWorkflow.run, id="wf1", task_queue="tq1")
            await env.sleep(50)
            assert "got timeout" == await handle.result()

Also, the current time of the workflow environment can be obtained via the async WorkflowEnvironment.get_current_time method.

Mocking Activities

Activities are just functions decorated with @activity.defn. Simply write different ones and pass those to the worker to have different activities called during the test.

Workflow Sandbox

By default workflows are run in a sandbox to help avoid non-deterministic code. If a call that is known to be non-deterministic is performed, an exception will be thrown in the workflow which will "fail the task" which means the workflow will not progress until fixed.

The sandbox is not foolproof and non-determinism can still occur. It is simply a best-effort way to catch bad code early. Users are encouraged to define their workflows in files with no other side effects.

The sandbox offers a mechanism to pass through modules from outside the sandbox. By default this already includes all standard library modules and Temporal modules. For performance and behavior reasons, users are encouraged to pass through all third party modules whose calls will be deterministic. This includes modules containing the activities to be referenced in workflows. See "Passthrough Modules" below on how to do this.

If you are getting an error like:

temporalio.worker.workflow_sandbox.restrictions.RestrictedWorkflowAccessError: Cannot access http.client.IncompleteRead.\_mroentries_\ from inside a workflow. If this is code from a module not used in a workflow or known to only be used deterministically from a workflow, mark the import as pass through.

Then you are either using an invalid construct from the workflow, this is a known limitation of the sandbox, or most commonly this is from a module that is safe to pass through (see "Passthrough Modules" section below).

How the Sandbox Works

The sandbox is made up of two components that work closely together:

Global state isolation is performed by using exec. Upon workflow start, the file that the workflow is defined in is imported into a new sandbox created for that workflow run. In order to keep the sandbox performant a known set of "passthrough modules" are passed through from outside of the sandbox when they are imported. These are expected to be side-effect free on import and have their non-deterministic aspects restricted. By default the entire Python standard library, temporalio, and a couple of other modules are passed through from outside of the sandbox. To update this list, see "Customizing the Sandbox".

Restrictions preventing known non-deterministic library calls are achieved using proxy objects on modules wrapped around the custom importer set in the sandbox. Many restrictions apply at workflow import time and workflow run time, while some restrictions only apply at workflow run time. A default set of restrictions is included that prevents most dangerous standard library calls. However it is known in Python that some otherwise-non-deterministic invocations, like reading a file from disk via open or using os.environ, are done as part of importing modules. To customize what is and isn't restricted, see "Customizing the Sandbox".

Avoiding the Sandbox

There are three increasingly-scoped ways to avoid the sandbox. Users are discouraged from avoiding the sandbox if possible.

To remove restrictions around a particular block of code, use with temporalio.workflow.unsafe.sandbox_unrestricted():. The workflow will still be running in the sandbox, but no restrictions for invalid library calls will be applied.

To run an entire workflow outside of a sandbox, set sandboxed=False on the @workflow.defn decorator when defining it. This will run the entire workflow outside of the workflow which means it can share global state and other bad things.

To disable the sandbox entirely for a worker, set the Worker init's workflow_runner keyword argument to temporalio.worker.UnsandboxedWorkflowRunner(). This value is defaulted to temporalio.worker.workflow_sandbox.SandboxedWorkflowRunner() so by changing it to the unsandboxed runner, the sandbox will not be used at all.

Customizing the Sandbox

⚠️ WARNING: APIs in the temporalio.worker.workflow_sandbox module are not yet considered stable and may change in future releases.

When creating the Worker, the workflow_runner is defaulted to temporalio.worker.workflow_sandbox.SandboxedWorkflowRunner(). The SandboxedWorkflowRunner's init accepts a restrictions keyword argument that is defaulted to SandboxRestrictions.default. The SandboxRestrictions dataclass is immutable and contains three fields that can be customized, but only two have notable value. See below.

Passthrough Modules

By default the sandbox completely reloads non-standard-library and non-Temporal modules for every workflow run. To make the sandbox quicker and use less memory when importing known-side-effect-free third party modules, they can be marked as passthrough modules.

For performance and behavior reasons, users are encouraged to pass through all third party modules whose calls will be deterministic.

One way to pass through a module is at import time in the workflow file using the imports_passed_through context manager like so:

# my_workflow_file.py

from temporalio import workflow

with workflow.unsafe.imports_passed_through():
    import pydantic

@workflow.defn
class MyWorkflow:
    ...

Alternatively, this can be done at worker creation time by customizing the runner's restrictions. For example:

my_worker = Worker(
  ...,
  workflow_runner=SandboxedWorkflowRunner(
    restrictions=SandboxRestrictions.default.with_passthrough_modules("pydantic")
  )
)

In both of these cases, now the pydantic module will be passed through from outside of the sandbox instead of being reloaded for every workflow run.

Invalid Module Members

SandboxRestrictions.invalid_module_members contains a root matcher that applies to all module members. This already has a default set which includes things like datetime.date.today() which should never be called from a workflow. To remove this restriction:

my_restrictions = dataclasses.replace(
    SandboxRestrictions.default,
    invalid_module_members=SandboxRestrictions.invalid_module_members_default.with_child_unrestricted(
      "datetime", "date", "today",
    ),
)
my_worker = Worker(..., workflow_runner=SandboxedWorkflowRunner(restrictions=my_restrictions))

Restrictions can also be added by |'ing together matchers, for example to restrict the datetime.date class from being used altogether:

my_restrictions = dataclasses.replace(
    SandboxRestrictions.default,
    invalid_module_members=SandboxRestrictions.invalid_module_members_default | SandboxMatcher(
      children={"datetime": SandboxMatcher(use={"date"})},
    ),
)
my_worker = Worker(..., workflow_runner=SandboxedWorkflowRunner(restrictions=my_restrictions))

See the API for more details on exact fields and their meaning.

Known Sandbox Issues

Below are known sandbox issues. As the sandbox is developed and matures, some may be resolved.

Global Import/Builtins

Currently the sandbox references/alters the global sys.modules and builtins fields while running workflow code. In order to prevent affecting other sandboxed code, thread locals are leveraged to only intercept these values during the workflow thread running. Therefore, technically if top-level import code starts a thread, it may lose sandbox protection.

Sandbox is not Secure

The sandbox is built to catch many non-deterministic and state sharing issues, but it is not secure. Some known bad calls are intercepted, but for performance reasons, every single attribute get/set cannot be checked. Therefore a simple call like setattr(temporalio.common, "__my_key", "my value") will leak across sandbox runs.

The sandbox is only a helper, it does not provide full protection.

Sandbox Performance

The sandbox does not add significant CPU or memory overhead for workflows that are in files which only import standard library modules. This is because they are passed through from outside of the sandbox. However, every non-standard-library import that is performed at the top of the same file the workflow is in will add CPU overhead (the module is re-imported every workflow run) and memory overhead (each module independently cached as part of the workflow run for isolation reasons). This becomes more apparent for large numbers of workflow runs.

To mitigate this, users should:

Extending Restricted Classes

Extending a restricted class causes Python to instantiate the restricted metaclass which is unsupported. Therefore if you attempt to use a class in the sandbox that extends a restricted class, it will fail. For example, if you have a class MyZipFile(zipfile.ZipFile) and try to use that class inside a workflow, it will fail.

Classes used inside the workflow should not extend restricted classes. For situations where third-party modules need to at import time, they should be marked as pass through modules.

Certain Standard Library Calls on Restricted Objects

If an object is restricted, internal C Python validation may fail in some cases. For example, running dict.items(os.__dict__) will fail with:

descriptor 'items' for 'dict' objects doesn't apply to a '_RestrictedProxy' object

This is a low-level check that cannot be subverted. The solution is to not use restricted objects inside the sandbox. For situations where third-party modules need to at import time, they should be marked as pass through modules.

is_subclass of ABC-based Restricted Classes

Due to https://bugs.python.org/issue44847, classes that are wrapped and then checked to see if they are subclasses of another via is_subclass may fail (see also this wrapt issue).

Compiled Pydantic Sometimes Using Wrong Types

If the Pydantic dependency is in compiled form (the default) and you are using a Pydantic model inside a workflow sandbox that uses a datetime type, it will grab the wrong validator and use date instead. This is because our patched form of issubclass is bypassed by compiled Pydantic.

To work around, either don't use datetime-based Pydantic model fields in workflows, or mark datetime library as passthrough (means you lose protection against calling the non-deterministic now()), or use non-compiled Pydantic dependency.

Activities

Definition

Activities are decorated with @activity.defn like so:

from temporalio import activity

@activity.defn
def say_hello_activity(name: str) -> str:
    return f"Hello, {name}!"

Some things to note about activity definitions:

Types of Activities

There are 3 types of activity callables accepted and described below: synchronous multithreaded, synchronous multiprocess/other, and asynchronous. Only positional parameters are allowed in activity callables.

Synchronous Activities

Synchronous activities, i.e. functions that do not have async def, can be used with workers, but the activity_executor worker parameter must be set with a concurrent.futures.Executor instance to use for executing the activities.

All long running, non-local activities should heartbeat so they can be cancelled. Cancellation in threaded activities throws but multiprocess/other activities does not. The sections below on each synchronous type explain further. There are also calls on the context that can check for cancellation. For more information, see "Activity Context" and "Heartbeating and Cancellation" sections later.

Note, all calls from an activity to functions in the temporalio.activity package are powered by contextvars. Therefore, new threads starting inside of activities must copy_context() and then .run() manually to ensure temporalio.activity calls like heartbeat still function in the new threads.

If any activity ever throws a concurrent.futures.BrokenExecutor, the failure is consisted unrecoverable and the worker will fail and shutdown.

Synchronous Multithreaded Activities

If activity_executor is set to an instance of concurrent.futures.ThreadPoolExecutor then the synchronous activities are considered multithreaded activities. If max_workers is not set to at least the worker's max_concurrent_activities setting a warning will be issued. Besides activity_executor, no other worker parameters are required for synchronous multithreaded activities.

By default, cancellation of a synchronous multithreaded activity is done via a temporalio.exceptions.CancelledError thrown into the activity thread. Activities that do not wish to have cancellation thrown can set no_thread_cancel_exception=True in the @activity.defn decorator.

Code that wishes to be temporarily shielded from the cancellation exception can run inside with activity.shield_thread_cancel_exception():. But once the last nested form of that block is finished, even if there is a return statement within, it will throw the cancellation if there was one. A try + except temporalio.exceptions.CancelledError would have to surround the with to handle the cancellation explicitly.

Synchronous Multiprocess/Other Activities

If activity_executor is set to an instance of concurrent.futures.Executor that is not concurrent.futures.ThreadPoolExecutor, then the synchronous activities are considered multiprocess/other activities. Users should prefer threaded activities over multiprocess ones since, among other reasons, threaded activities can raise on cancellation.

These require special primitives for heartbeating and cancellation. The shared_state_manager worker parameter must be set to an instance of temporalio.worker.SharedStateManager. The most common implementation can be created by passing a multiprocessing.managers.SyncManager (i.e. result of multiprocessing.managers.Manager()) to temporalio.worker.SharedStateManager.create_from_multiprocessing().

Also, all of these activity functions must be "picklable".

Asynchronous Activities

Asynchronous activities are functions defined with async def. Asynchronous activities are often much more performant than synchronous ones. When using asynchronous activities no special worker parameters are needed.

⚠️ WARNING: Do not block the thread in async def Python functions. This can stop the processing of the rest of the Temporal.

Cancellation for asynchronous activities is done via asyncio.Task.cancel. This means that asyncio.CancelledError will be raised (and can be caught, but it is not recommended). A non-local activity must heartbeat to receive cancellation and there are other ways to be notified about cancellation (see "Activity Context" and "Heartbeating and Cancellation" later).

Activity Context

During activity execution, an implicit activity context is set as a context variable. The context variable itself is not visible, but calls in the temporalio.activity package make use of it. Specifically:

With the exception of in_activity(), if any of the functions are called outside of an activity context, an error occurs. Synchronous activities cannot call any of the async functions.

Heartbeating and Cancellation

In order for a non-local activity to be notified of cancellation requests, it must be given a heartbeat_timeout at invocation time and invoke temporalio.activity.heartbeat() inside the activity. It is strongly recommended that all but the fastest executing activities call this function regularly. "Types of Activities" has specifics on cancellation for synchronous and asynchronous activities.

In addition to obtaining cancellation information, heartbeats also support detail data that is persisted on the server for retrieval during activity retry. If an activity calls temporalio.activity.heartbeat(123, 456) and then fails and is retried, temporalio.activity.info().heartbeat_details will return an iterable containing 123 and 456 on the next run.

Heartbeating has no effect on local activities.

Worker Shutdown

An activity can react to a worker shutdown. Using is_worker_shutdown or one of the wait_for_worker_shutdown functions an activity can react to a shutdown.

When the graceful_shutdown_timeout worker parameter is given a datetime.timedelta, on shutdown the worker will notify activities of the graceful shutdown. Once that timeout has passed (or if wasn't set), the worker will perform cancellation of all outstanding activities.

The shutdown() invocation will wait on all activities to complete, so if a long-running activity does not at least respect cancellation, the shutdown may never complete.

Testing

Unit testing an activity or any code that could run in an activity is done via the temporalio.testing.ActivityEnvironment class. Simply instantiate this and any callable + params passed to run will be invoked inside the activity context. The following are attributes/methods on the environment that can be used to affect calls activity code might make to functions on the temporalio.activity package.

Workflow Replay

Given a workflow's history, it can be replayed locally to check for things like non-determinism errors. For example, assuming history_str is populated with a JSON string history either exported from the web UI or from tctl, the following function will replay it:

from temporalio.client import WorkflowHistory
from temporalio.worker import Replayer

async def run_replayer(history_str: str):
  replayer = Replayer(workflows=[SayHello])
  await replayer.replay_workflow(WorkflowHistory.from_json(history_str))

This will throw an error if any non-determinism is detected.

Replaying from workflow history is a powerful concept that many use to test that workflow alterations won't cause non-determinisms with past-complete workflows. The following code will make sure that all workflow histories for a certain workflow type (i.e. workflow class) are safe with the current code.

from temporalio.client import Client, WorkflowHistory
from temporalio.worker import Replayer

async def check_past_histories(my_client: Client):
  replayer = Replayer(workflows=[SayHello])
  await replayer.replay_workflows(
    await my_client.list_workflows("WorkflowType = 'SayHello'").map_histories(),
  )

OpenTelemetry Support

OpenTelemetry support requires the optional opentelemetry dependencies which are part of the opentelemetry extra. When using pip, running

pip install temporalio[opentelemetry]

will install needed dependencies. Then the temporalio.contrib.opentelemetry.TracingInterceptor can be created and set as an interceptor on the interceptors argument of Client.connect. When set, spans will be created for all client calls and for all activity and workflow invocations on the worker, spans will be created and properly serialized through the server to give one proper trace for a workflow execution.

Protobuf 3.x vs 4.x

Python currently has two somewhat-incompatible protobuf library versions - the 3.x series and the 4.x series. Python currently recommends 4.x and that is the primary supported version. Some libraries like Pulumi require 4.x. Other libraries such as ONNX and Streamlit, for one reason or another, have/will not leave 3.x.

To support these, Temporal Python SDK allows any protobuf library >= 3.19. However, the C extension in older Python versions can cause issues with the sandbox due to global state sharing. Temporal strongly recommends using the latest protobuf 4.x library unless you absolutely cannot at which point some proto libraries may have to be marked as Passthrough Modules.

Known Compatibility Issues

Below are known compatibility issues with the Python SDK.

gevent Patching

When using gevent.monkey.patch_all(), asyncio event loops can get messed up, especially those using custom event loops like Temporal. See this gevent issue. This is a known incompatibility and users are encouraged to not use gevent in asyncio applications (including Temporal). But if you must, there is a sample showing how it is possible.

Development

The Python SDK is built to work with Python 3.8 and newer. It is built using SDK Core which is written in Rust.

Building

Prepare

To build the SDK from source for use as a dependency, the following prerequisites are required:

macOS note: If errors are encountered, it may be better to install Python and Rust as recommended from their websites instead of via brew.

With the prerequisites installed, first clone the SDK repository recursively:

git clone --recursive https://github.com/temporalio/sdk-python.git
cd sdk-python

Use poetry to install the dependencies with --no-root to not install this package (because we still need to build it):

poetry install --no-root --all-extras

Build

Now perform the release build:

This will take a while because Rust will compile the core project in release mode (see Local SDK development environment for the quicker approach to local development).

poetry build

The compiled wheel doesn't have the exact right tags yet for use, so run this script to fix it:

poe fix-wheel

The whl wheel file in dist/ is now ready to use.

Use

The wheel can now be installed into any virtual environment.

For example, create a virtual environment somewhere and then run the following inside the virtual environment:

pip install wheel
pip install /path/to/cloned/sdk-python/dist/*.whl

Create this Python file at example.py:

import asyncio
from temporalio import workflow, activity
from temporalio.client import Client
from temporalio.worker import Worker

@workflow.defn
class SayHello:
    @workflow.run
    async def run(self, name: str) -> str:
        return f"Hello, {name}!"

async def main():
    client = await Client.connect("localhost:7233")
    async with Worker(client, task_queue="my-task-queue", workflows=[SayHello]):
        result = await client.execute_workflow(SayHello.run, "Temporal",
            id="my-workflow-id", task_queue="my-task-queue")
        print(f"Result: {result}")

if __name__ == "__main__":
    asyncio.run(main())

Assuming there is a local Temporal server running, execute the file with python (or python3 if necessary):

python example.py

It should output:

Result: Hello, Temporal!

Local SDK development environment

For local development, it is often quicker to use debug builds and a local virtual environment.

While not required, it often helps IDEs if we put the virtual environment .venv directory in the project itself. This can be configured system-wide via:

poetry config virtualenvs.in-project true

Now perform the same steps as the "Prepare" section above by installing the prerequisites, cloning the project, installing dependencies, and generating the protobuf code:

git clone --recursive https://github.com/temporalio/sdk-python.git
cd sdk-python
poetry install --no-root --all-extras

Now compile the Rust extension in develop mode which is quicker than release mode:

poe build-develop

That step can be repeated for any Rust changes made.

The environment is now ready to develop in.

Testing

To execute tests:

poe test

This runs against Temporalite. To run against the time-skipping test server, pass --workflow-environment time-skipping. To run against the default namespace of an already-running server, pass the host:port to --workflow-environment. Can also use regular pytest arguments. For example, here's how to run a single test with debug logs on the console:

poe test -s --log-cli-level=DEBUG -k test_sync_activity_thread_cancel_caught

Proto Generation and Testing

To allow for backwards compatibility, protobuf code is generated on the 3.x series of the protobuf library. To generate protobuf code, you must be on Python <= 3.10, and then run poetry add "protobuf<4". Then the protobuf files can be generated via poe gen-protos. Tests can be run for protobuf version 3 by setting the TEMPORAL_TEST_PROTO3 env var to 1 prior to running tests.

Do not commit poetry.lock or pyproject.toml changes. To go back from this downgrade, restore both of those files and run poetry install --no-root --all-extras. Make sure you poe format the results.

For a less system-intrusive approach, you can (note this approach may have a bug):

docker build -f scripts/_proto/Dockerfile .
docker run --rm -v "${PWD}/temporalio/api:/api_new" -v "${PWD}/temporalio/bridge/proto:/bridge_new" <just built image sha>
poe format

Style