Skip to content


Don't Panic! If you experience an error with Prefect, there are many paths to understanding and resolving it. The first troubleshooting step is confirming that you are running the latest version of Prefect. If you are not, be sure to upgrade to the latest version, since the issue may have already been fixed. Beyond that, there are several categories of errors:

  • The issue may be in your flow code, in which case you should carefully read the logs.
  • The issue could be with how you are authenticated, and whether or not you are connected to Cloud.
  • The issue might have to do with how your code is executed.


Prefect is constantly evolving, adding new features and fixing bugs. Chances are that a patch has already been identified and released. Search existing issues for similar reports and check out the Release Notes. Upgrade to the newest version with the following command:

pip install --upgrade prefect

Different components may use different versions of Prefect:

  • Cloud will generally always be the newest version. Cloud is continuously deployed by the Prefect team. When using a self-hosted server, you can control this version.
  • Workers and agents typically don't change versions frequently, and are usually whatever the latest version was at the time of creation. Workers and agents provision infrastructure for flow runs, so upgrading them may help with infrastructure problems.
  • Flows could use a different version than the worker or agent that created them, especially when running in different environments. Suppose your worker and flow both use the latest official Docker image, but your worker was created a month ago. Your worker will often be on an older version than your flow.

Integration Versions

Keep in mind that integrations are versioned and released independently of the core Prefect library. They should be upgraded simultaneously with the core library, using the same method.


In many cases, there will be an informative stack trace in Prefect's logs. Read it carefully, locate the source of the error, and try to identify the cause.

There are two types of logs:

  • Flow and task logs are always scoped to a flow. They are sent to Prefect and are viewable in the UI.
  • Worker and agent logs are not scoped to a flow and may have more information on what happened before the flow started. These logs are generally only available where the worker or agent is running.

If your flow and task logs are empty, there may have been an infrastructure issue that prevented your flow from starting. Check your worker logs for more details.

If there is no clear indication of what went wrong, try updating the logging level from the default INFO level to the DEBUG level. Settings such as the logging level are propagated from the worker environment to the flow run environment and can be set via environment variables or the prefect config set CLI:

# Using the CLI

# Using environment variables

The DEBUG logging level produces a high volume of logs so consider setting it back to INFO once any issues are resolved.


When using Prefect Cloud, there are the additional concerns of authentication and authorization. The Prefect API authenticates users and service accounts - collectively known as actors - with API keys. Missing, incorrect, or expired API keys will result in a 401 response with detail Invalid authentication credentials. Use the following command to check your authentication, replacing $PREFECT_API_KEY with your API key:

curl -s -H "Authorization: Bearer $PREFECT_API_KEY" ""

Users vs Service Accounts

Service accounts - sometimes referred to as bots - represent non-human actors that interact with Prefect such as workers and CI/CD systems. Each human that interacts with Prefect should be represented as a user. User API keys start with pnu_ and service account API keys start with pnb_.

Supposing the response succeeds, let's check our authorization. Actors can be members of workspaces. An actor attempting an action in a workspace they are not a member of will result in a 404 response. Use the following command to check your actor's workspace memberships:

curl -s -H "Authorization: Bearer $PREFECT_API_KEY" ""

Formatting JSON

Python comes with a helpful tool for formatting JSON. Append the following to the end of the command above to make the output more readable: | python -m json.tool

Make sure your actor is a member of the workspace you are working in. Within a workspace, an actor has a role which grants them certain permissions. Insufficient permissions will result in an error. For example, starting an agent or worker with the Viewer role, will result in errors.


Prefect flows can be executed locally by the user, or remotely by a worker or agent. Local execution generally means that you - the user - run your flow directly with a command like python Remote execution generally means that a worker runs your flow via a deployment, optionally on different infrastructure.

With remote execution, the creation of your flow run happens separately from its execution. Flow runs are assigned to a work pool and a work queue. For flow runs to execute, a worker must be subscribed to the work pool and work queue, otherwise the flow runs will go from Scheduled to Late. Ensure that your work pool and work queue have a subscribed worker.

Local and remote execution can also differ in their treatment of relative imports. If switching from local to remote execution results in local import errors, try replicating the behavior by executing the flow locally with the -m flag (i.e. python -m flow instead of python Read more about -m here.

API tests return an unexpected 307 Redirected

Summary: Requests require a trailing / in the request URL.

If you write a test that does not include a trailing / when making a request to a specific endpoint:

async def test_example(client):
    response = await"/my_route")
    assert response.status_code == 201

You'll see a failure like:

E       assert 307 == 201
E        +  where 307 = <Response [307 Temporary Redirect]>.status_code

To resolve this, include the trailing /:

async def test_example(client):
    response = await"/my_route/")
    assert response.status_code == 201

Note: requests to nested URLs may exhibit the opposite behavior and require no trailing slash:

async def test_nested_example(client):
    response = await"/my_route/filter/")
    assert response.status_code == 307

    response = await"/my_route/filter")
    assert response.status_code == 200

Reference: "HTTPX disabled redirect following by default" in 0.22.0.

pytest.PytestUnraisableExceptionWarning or ResourceWarning

As you're working with one of the FlowRunner implementations, you may get an error like this one:

E               pytest.PytestUnraisableExceptionWarning: Exception ignored in: <ssl.SSLSocket fd=-1, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0>
E               Traceback (most recent call last):
E                 File ".../pytest_asyncio/", line 306, in setup
E                   res = await func(**_add_kwargs(func, kwargs, event_loop, request))
E               ResourceWarning: unclosed <ssl.SSLSocket fd=10, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('', 60605), raddr=('', 6443)>

.../_pytest/ PytestUnraisableExceptionWarning

This error is saying that your test suite (or the prefect library code) opened a connection to something (like a Docker daemon or a Kubernetes cluster) and didn't close it.

It may help to re-run the specific test with PYTHONTRACEMALLOC=25 pytest ... so that Python can display more of the stack trace where the connection was opened.