Creating Interactive Workflows¶
Flows can pause or suspend execution and automatically resume when they receive type-checked input in Prefect's UI. Flows can also send and receive type-checked input at any time while running, without pausing or suspending. This guide will show you how to use these features to build interactive workflows.
A note on async Python syntax
Most of the example code in this section uses async Python functions and await
. However, as with other Prefect features, you can call these functions with or without await
.
Pausing or suspending a flow until it receives input¶
You can pause or suspend a flow until it receives input from a user in Prefect's UI. This is useful when you need to ask for additional information or feedback before resuming a flow. Such workflows are often called human-in-the-loop (HITL) systems.
What is human-in-the-loop interactivity used for?
Approval workflows that pause to ask a human to confirm whether a workflow should continue are very common in the business world. Certain types of machine learning training and artificial intelligence workflows benefit from incorporating HITL design.
Waiting for input¶
To receive input while paused or suspended use the wait_for_input
parameter in the pause_flow_run
or suspend_flow_run
functions. This parameter accepts one of the following:
- A built-in type like
int
orstr
, or a built-in collection likeList[int]
- A
pydantic.BaseModel
subclass - A subclass of
prefect.input.RunInput
When to use a RunModel
or BaseModel
instead of a built-in type
There are a few reasons to use a RunModel
or BaseModel
. The first is that when you let Prefect automatically create one of these classes for your input type, the field that users will see in Prefect's UI when they click "Resume" on a flow run is named value
and has no help text to suggest what the field is. If you create a RunInput
or BaseModel
, you can change details like the field name, help text, and default value, and users will see those reflected in the "Resume" form.
The simplest way to pause or suspend and wait for input is to pass a built-in type:
from prefect import flow, pause_flow_run, get_run_logger
@flow
def greet_user():
logger = get_run_logger()
user = pause_flow_run(wait_for_input=str)
logger.info(f"Hello, {user}!")
In this example, the flow run will pause until a user clicks the Resume button in the Prefect UI, enters a name, and submits the form.
What types can you pass for wait_for_input
?
When you pass a built-in type such as int
as an argument for the wait_for_input
parameter to pause_flow_run
or suspend_flow_run
, Prefect automatically creates a Pydantic model containing one field annotated with the type you specified. This means you can use any type annotation that Pydantic accepts for model fields with these functions.
Instead of a built-in type, you can pass in a pydantic.BaseModel
class. This is useful if you already have a BaseModel
you want to use:
from prefect import flow, pause_flow_run, get_run_logger
from pydantic import BaseModel
class User(BaseModel):
name: str
age: int
@flow
async def greet_user():
logger = get_run_logger()
user = await pause_flow_run(wait_for_input=User)
logger.info(f"Hello, {user.name}!")
BaseModel
classes are upgraded to RunInput
classes automatically
When you pass a pydantic.BaseModel
class as the wait_for_input
argument to pause_flow_run
or suspend_flow_run
, Prefect automatically creates a RunInput
class with the same behavior as your BaseModel
and uses that instead.
RunInput
classes contain extra logic that allows flows to send and receive them at runtime. You shouldn't notice any difference!
Finally, for advanced use cases like overriding how Prefect stores flow run inputs, you can create a RunInput
class:
from prefect import get_run_logger
from prefect.input import RunInput
class UserInput(RunInput):
name: str
age: int
# Imagine overridden methods here!
def override_something(self, *args, **kwargs):
super().override_something(*args, **kwargs)
@flow
async def greet_user():
logger = get_run_logger()
user = await pause_flow_run(wait_for_input=UserInput)
logger.info(f"Hello, {user.name}!")
Providing initial data¶
You can set default values for fields in your model by using the with_initial_data
method. This is useful when you want to provide default values for the fields in your own RunInput
class.
Expanding on the example above, you could make the name
field default to "anonymous":
from prefect import get_run_logger
from prefect.input import RunInput
class UserInput(RunInput):
name: str
age: int
@flow
async def greet_user():
logger = get_run_logger()
user_input = await pause_flow_run(
wait_for_input=UserInput.with_initial_data(name="anonymous")
)
if user_input.name == "anonymous":
logger.info("Hello, stranger!")
else:
logger.info(f"Hello, {user_input.name}!")
When a user sees the form for this input, the name field will contain "anonymous" as the default.
Providing a description with runtime data¶
You can provide a dynamic, markdown description that will appear in the Prefect UI when the flow run pauses. This feature enables context-specific prompts, enhancing clarity and user interaction. Building on the example above:
from datetime import datetime
from prefect import flow, pause_flow_run, get_run_logger
from prefect.input import RunInput
class UserInput(RunInput):
name: str
age: int
@flow
async def greet_user():
logger = get_run_logger()
current_date = datetime.now().strftime("%B %d, %Y")
description_md = f"""
**Welcome to the User Greeting Flow!**
Today's Date: {current_date}
Please enter your details below:
- **Name**: What should we call you?
- **Age**: Just a number, nothing more.
"""
user_input = await pause_flow_run(
wait_for_input=UserInput.with_initial_data(
description=description_md, name="anonymous"
)
)
if user_input.name == "anonymous":
logger.info("Hello, stranger!")
else:
logger.info(f"Hello, {user_input.name}!")
When a user sees the form for this input, the given markdown will appear above the input fields.
Handling custom validation¶
Prefect uses the fields and type hints on your RunInput
or BaseModel
class to validate the general structure of input your flow receives, but you might require more complex validation. If you do, you can use Pydantic validators.
Custom validation runs after the flow resumes
Prefect transforms the type annotations in your RunInput
or BaseModel
class to a JSON schema and uses that schema in the UI for client-side validation. However, custom validation requires running Python logic defined in your RunInput
class. Because of this, validation happens after the flow resumes, so you'll want to handle it explicitly in your flow. Continue reading for an example best practice.
The following is an example RunInput
class that uses a custom field validator:
import pydantic
from prefect.input import RunInput
class ShirtOrder(RunInput):
size: Literal["small", "medium", "large", "xlarge"]
color: Literal["red", "green", "black"]
@pydantic.validator("color")
def validate_age(cls, value, values, **kwargs):
if value == "green" and values["size"] == "small":
raise ValueError(
"Green is only in-stock for medium, large, and XL sizes."
)
return value
In the example, we use Pydantic's validator
decorator to define a custom validation method for the color
field. We can use it in a flow like this:
import pydantic
from prefect import flow, pause_flow_run
from prefect.input import RunInput
class ShirtOrder(RunInput):
size: Literal["small", "medium", "large", "xlarge"]
color: Literal["red", "green", "black"]
@pydantic.validator("color")
def validate_age(cls, value, values, **kwargs):
if value == "green" and values["size"] == "small":
raise ValueError(
"Green is only in-stock for medium, large, and XL sizes."
)
return value
@flow
def get_shirt_order():
shirt_order = pause_flow_run(wait_for_input=ShirtOrder)
If a user chooses any size and color combination other than small
and green
, the flow run will resume successfully. However, if the user chooses size small
and color green
, the flow run will resume, and pause_flow_run
will raise a ValidationError
exception. This will cause the flow run to fail and log the error.
However, what if you don't want the flow run to fail? One way to handle this case is to use a while
loop and pause again if the ValidationError
exception is raised:
from typing import Literal
import pydantic
from prefect import flow, get_run_logger, pause_flow_run
from prefect.input import RunInput
class ShirtOrder(RunInput):
size: Literal["small", "medium", "large", "xlarge"]
color: Literal["red", "green", "black"]
@pydantic.validator("color")
def validate_age(cls, value, values, **kwargs):
if value == "green" and values["size"] == "small":
raise ValueError(
"Green is only in-stock for medium, large, and XL sizes."
)
return value
@flow
def get_shirt_order():
logger = get_run_logger()
shirt_order = None
while shirt_order is None:
try:
shirt_order = pause_flow_run(wait_for_input=ShirtOrder)
except pydantic.ValidationError as exc:
logger.error(f"Invalid size and color combination: {exc}")
logger.info(
f"Shirt order: {shirt_order.size}, {shirt_order.color}"
)
This code will cause the flow run to continually pause until the user enters a valid age.
As an additional step, you may want to use an automation or notification to alert the user to the error.
Sending and receiving input at runtime¶
Use the send_input
and receive_input
functions to send input to a flow or receive input from a flow at runtime. You don't need to pause or suspend the flow to send or receive input.
Why would you send or receive input without pausing or suspending?
You might want to send or receive input without pausing or suspending in scenarios where the flow run is designed to handle real-time data. For instance, in a live monitoring system, you might need to update certain parameters based on the incoming data without interrupting the flow. Another use is having a long-running flow that continually responds to runtime input with low latency. For example, if you're building a chatbot, you could have a flow that starts a GPT Assistant and manages a conversation thread.
The most important parameter to the send_input
and receive_input
functions is run_type
, which should be one of the following:
- A built-in type such as
int
orstr
- A
pydantic.BaseModel
class - A
prefect.input.RunInput
class
When to use a BaseModel
or RunInput
instead of a built-in type
Most built-in types and collections of built-in types should work with send_input
and receive_input
, but there is a caveat with nested collection types, such as lists of tuples, e.g. List[Tuple[str, float]])
. In this case, validation may happen after your flow receives the data, so calling receive_input
may raise a ValidationError
. You can plan to catch this exception, but also, consider placing the field in an explicit BaseModel
or RunInput
so that your flow only receives exact type matches.
Let's look at some examples! We'll check out receive_input
first, followed by send_input
, and then we'll see the two functions working together.
Receiving input¶
The following flow uses receive_input
to continually receive names and print a personalized greeting for each name it receives:
from prefect import flow
from prefect.input.run_input import receive_input
@flow
async def greeter_flow():
async for name_input in receive_input(str, timeout=None):
# Prints "Hello, andrew!" if another flow sent "andrew"
print(f"Hello, {name_input}!")
When you pass a type such as str
into receive_input
, Prefect creates a RunInput
class to manage your input automatically. When a flow sends input of this type, Prefect uses the RunInput
class to validate the input. If the validation succeeds, your flow receives the input in the type you specified. In this example, if the flow received a valid string as input, the variable name_input
would contain the string value.
If, instead, you pass a BaseModel
, Prefect upgrades your BaseModel
to a RunInput
class, and the variable your flow sees — in this case, name_input
— is a RunInput
instance that behaves like a BaseModel
. Of course, if you pass in a RunInput
class, no upgrade is needed, and you'll get a RunInput
instance.
If you prefer to keep things simple and pass types such as str
into receive_input
, you can do so. If you need access to the generated RunInput
that contains the received value, pass with_metadata=True
to receive_input
:
from prefect import flow
from prefect.input.run_input import receive_input
@flow
async def greeter_flow():
async for name_input in receive_input(
str,
timeout=None,
with_metadata=True
):
# Input will always be in the field "value" on this object.
print(f"Hello, {name_input.value}!")
Why would you need to use with_metadata=True
?
The primary uses of accessing the RunInput
object for a receive input are to respond to the sender with the RunInput.respond()
function or to access the unique key for an input. Later in this guide, we'll discuss how and why you might use these features.
Notice that we are now printing name_input.value
. When Prefect generates a RunInput
for you from a built-in type, the RunInput
class has a single field, value
, that uses a type annotation matching the type you specified. So if you call receive_input
like this: receive_input(str, with_metadata=True)
, that's equivalent to manually creating the following RunInput
class and receive_input
call:
from prefect import flow
from prefect.input.run_input import RunInput
class GreeterInput(RunInput):
value: str
@flow
async def greeter_flow():
async for name_input in receive_input(GreeterInput, timeout=None):
print(f"Hello, {name_input.value}!")
The type used in receive_input
and send_input
must match
For a flow to receive input, the sender must use the same type that the receiver is receiving. This means that if the receiver is receiving GreeterInput
, the sender must send GreeterInput
. If the receiver is receiving GreeterInput
and the sender sends str
input that Prefect automatically upgrades to a RunInput
class, the types won't match, so the receiving flow run won't receive the input. However, the input will be waiting if the flow ever calls receive_input(str)
!
Keeping track of inputs you've already seen¶
By default, each time you call receive_input
, you get an iterator that iterates over all known inputs to a specific flow run, starting with the first received. The iterator will keep track of your current position as you iterate over it, or you can call next()
to explicitly get the next input. If you're using the iterator in a loop, you should probably assign it to a variable:
from prefect import flow, get_client
from prefect.deployments.deployments import run_deployment
from prefect.input.run_input import receive_input, send_input
EXIT_SIGNAL = "__EXIT__"
@flow
async def sender():
greeter_flow_run = await run_deployment(
"greeter/send-receive", timeout=0, as_subflow=False
)
client = get_client()
# Assigning the `receive_input` iterator to a variable
# outside of the the `while True` loop allows us to continue
# iterating over inputs in subsequent passes through the
# while loop without losing our position.
receiver = receive_input(
str,
with_metadata=True,
timeout=None,
poll_interval=0.1
)
while True:
name = input("What is your name? ")
if not name:
continue
if name == "q" or name == "quit":
await send_input(
EXIT_SIGNAL,
flow_run_id=greeter_flow_run.id
)
print("Goodbye!")
break
await send_input(name, flow_run_id=greeter_flow_run.id)
# Saving the iterator outside of the while loop and
# calling next() on each iteration of the loop ensures
# that we're always getting the newest greeting. If we
# had instead called `receive_input` here, we would
# always get the _first_ greeting this flow received,
# print it, and then ask for a new name.
greeting = await receiver.next()
print(greeting)
So, an iterator helps to keep track of the inputs your flow has already received. But what if you want your flow to suspend and then resume later, picking up where it left off? In that case, you will need to save the keys of the inputs you've seen so that the flow can read them back out when it resumes. You might use a Block, such as a JSONBlock
.
The following flow receives input for 30 seconds then suspends itself, which exits the flow and tears down infrastructure:
from prefect import flow, get_run_logger, suspend_flow_run
from prefect.blocks.system import JSON
from prefect.context import get_run_context
from prefect.input.run_input import receive_input
EXIT_SIGNAL = "__EXIT__"
@flow
async def greeter():
logger = get_run_logger()
run_context = get_run_context()
assert run_context.flow_run, "Could not see my flow run ID"
block_name = f"{run_context.flow_run.id}-seen-ids"
try:
seen_keys_block = await JSON.load(block_name)
except ValueError:
seen_keys_block = JSON(
value=[],
)
try:
async for name_input in receive_input(
str,
with_metadata=True,
poll_interval=0.1,
timeout=30,
exclude_keys=seen_keys_block.value
):
if name_input.value == EXIT_SIGNAL:
print("Goodbye!")
return
await name_input.respond(f"Hello, {name_input.value}!")
seen_keys_block.value.append(name_input.metadata.key)
await seen_keys_block.save(
name=block_name,
overwrite=True
)
except TimeoutError:
logger.info("Suspending greeter after 30 seconds of idle time")
await suspend_flow_run(timeout=10000)
As this flow processes name input, it adds the key of the flow run input to the seen_keys_block
. When the flow later suspends and then resumes, it reads the keys it has already seen out of the JSON Block and passes them as the exlude_keys
parameter to receive_input
.
Responding to the input's sender¶
When your flow receives input from another flow, Prefect knows the sending flow run ID, so the receiving flow can respond by calling the respond
method on the RunInput
instance the flow received. There are a couple of requirements:
- You will need to pass in a
BaseModel
orRunInput
, or usewith_metadata=True
- The flow you are responding to must receive the same type of input you send in order to see it.
The respond
method is equivalent to calling send_input(..., flow_run_id=sending_flow_run.id)
, but with respond
, your flow doesn't need to know the sending flow run's ID.
Now that we know about respond
, let's make our greeter_flow
respond to name inputs instead of printing them:
from prefect import flow
from prefect.input.run_input import receive_input
@flow
async def greeter():
async for name_input in receive_input(
str,
with_metadata=True,
timeout=None
):
await name_input.respond(f"Hello, {name_input.value}!")
Cool! There's one problem left: this flow runs forever! We need a way to signal that it should exit. Let's keep things simple and teach it to look for a special string:
from prefect import flow
from prefect.input.run_input import receive_input
EXIT_SIGNAL = "__EXIT__"
@flow
async def greeter():
async for name_input in receive_input(
str,
with_metadata=True,
poll_interval=0.1,
timeout=None
):
if name_input.value == EXIT_SIGNAL:
print("Goodbye!")
return
await name_input.respond(f"Hello, {name_input.value}!")
With a greeter
flow in place, we're ready to create the flow that sends greeter
names!
Sending input¶
You can send input to a flow with the send_input
function. This works similarly to receive_input
and, like that function, accepts the same run_input
argument, which can be a built-in type such as str
, or else a BaseModel
or RunInput
subclass.
When can you send input to a flow run?
You can send input to a flow run as soon as you have the flow run's ID. The flow does not have to be receiving input for you to send input. If you send a flow input before it is receiving, it will see your input when it calls receive_input
(as long as the types in the send_input
and receive_input
calls match!)
Next, we'll create a sender
flow that starts a greeter
flow run and then enters a loop, continuously getting input from the terminal and sending it to the greeter flow:
@flow
async def sender():
greeter_flow_run = await run_deployment(
"greeter/send-receive", timeout=0, as_subflow=False
)
receiver = receive_input(str, timeout=None, poll_interval=0.1)
client = get_client()
while True:
flow_run = await client.read_flow_run(greeter_flow_run.id)
if not flow_run.state or not flow_run.state.is_running():
continue
name = input("What is your name? ")
if not name:
continue
if name == "q" or name == "quit":
await send_input(
EXIT_SIGNAL,
flow_run_id=greeter_flow_run.id
)
print("Goodbye!")
break
await send_input(name, flow_run_id=greeter_flow_run.id)
greeting = await receiver.next()
print(greeting)
There's more going on here than in greeter
, so let's take a closer look at the pieces.
First, we use run_deployment
to start a greeter
flow run. This means we must have a worker or flow.serve()
running in separate process. That process will begin running greeter
while sender
continues to execute. Calling run_deployment(..., timeout=0)
ensures that sender
won't wait for the greeter
flow run to complete, because it's running a loop and will only exit when we send EXIT_SIGNAL
.
Next, we capture the iterator returned by receive_input
as receiver
. This flow works by entering a loop, and on each iteration of the loop, the flow asks for terminal input, sends that to the greeter
flow, and then runs receiver.next()
to wait until it receives the response from greeter
.
Next, we let the terminal user who ran this flow exit by entering the string q
or quit
. When that happens, we send the greeter
flow an exit signal so it will shut down too.
Finally, we send the new name to greeter
. We know that greeter
is going to send back a greeting as a string, so we immediately wait for new string input. When we receive the greeting, we print it and continue the loop that gets terminal input.
Seeing a complete example¶
Finally, let's see a complete example of using send_input
and receive_input
. Here is what the greeter
and sender
flows look like together:
import asyncio
import sys
from prefect import flow, get_client
from prefect.blocks.system import JSON
from prefect.context import get_run_context
from prefect.deployments.deployments import run_deployment
from prefect.input.run_input import receive_input, send_input
EXIT_SIGNAL = "__EXIT__"
@flow
async def greeter():
run_context = get_run_context()
assert run_context.flow_run, "Could not see my flow run ID"
block_name = f"{run_context.flow_run.id}-seen-ids"
try:
seen_keys_block = await JSON.load(block_name)
except ValueError:
seen_keys_block = JSON(
value=[],
)
async for name_input in receive_input(
str,
with_metadata=True,
poll_interval=0.1,
timeout=None
):
if name_input.value == EXIT_SIGNAL:
print("Goodbye!")
return
await name_input.respond(f"Hello, {name_input.value}!")
seen_keys_block.value.append(name_input.metadata.key)
await seen_keys_block.save(
name=block_name,
overwrite=True
)
@flow
async def sender():
greeter_flow_run = await run_deployment(
"greeter/send-receive", timeout=0, as_subflow=False
)
receiver = receive_input(str, timeout=None, poll_interval=0.1)
client = get_client()
while True:
flow_run = await client.read_flow_run(greeter_flow_run.id)
if not flow_run.state or not flow_run.state.is_running():
continue
name = input("What is your name? ")
if not name:
continue
if name == "q" or name == "quit":
await send_input(
EXIT_SIGNAL,
flow_run_id=greeter_flow_run.id
)
print("Goodbye!")
break
await send_input(name, flow_run_id=greeter_flow_run.id)
greeting = await receiver.next()
print(greeting)
if __name__ == "__main__":
if sys.argv[1] == "greeter":
asyncio.run(greeter.serve(name="send-receive"))
elif sys.argv[1] == "sender":
asyncio.run(sender())
To run the example, you'll need a Python environment with Prefect installed, pointed at either an open-source Prefect server instance or Prefect Cloud.
With your environment set up, start a flow runner in one terminal with the following command:
python my_file_name greeter
For example, with Prefect Cloud, you should see output like this:
╭──────────────────────────────────────────────────────────────────────────────────────────────────╮
│ Your flow 'greeter' is being served and polling for scheduled runs! │
│ │
│ To trigger a run for this flow, use the following command: │
│ │
│ $ prefect deployment run 'greeter/send-receive' │
│ │
│ You can also run your flow via the Prefect UI: │
│ https://app.prefect.cloud/account/...(a URL for your account) │
│ │
╰──────────────────────────────────────────────────────────────────────────────────────────────────╯
Then start the greeter process in another terminal:
python my_file_name sender
You should see output like this:
11:38:41.800 | INFO | prefect.engine - Created flow run 'gregarious-owl' for flow 'sender'
11:38:41.802 | INFO | Flow run 'gregarious-owl' - View at https://app.prefect.cloud/account/...
What is your name?
Type a name and press the enter key to see a greeting, and you'll see sending and receiving in action:
What is your name? andrew
Hello, andrew!