How to prevent LangGraph GraphInterrupt from flagging tracing spans as errors?

[aliaksandrzakreuski]

Hi

We are using LangGraph's native interrupt()feature for human-in-the-loop workflows (such as user verification and suggestion choices), while monitoring our application using Arize Phoenix via openinference-instrumentation-langchain.

LangGraph uses the GraphInterruptexception as an expected control-flow mechanism to cleanly pause graph execution and save state checkpoints. However, the OpenInference instrumentation captures this escaping exception as a fatal application crash, force-setting the OpenTelemetry span status to ERROR.

Is there an idiomatic or supported way within Phoenix/OpenInference to filter out GraphInterruptfrom being logged as an error?

Thank you

[dosubot[bot]]

Hi @aliaksandrzakreuski! I'm Dosu and I'm helping the Arize team.

The openinference-instrumentation-langchain package has a built-in mechanism for filtering control-flow exceptions via an IGNORED_EXCEPTION_PATTERNS list [1]. When a run completes, the _update_span function checks if the error message matches any pattern in this list — if it does, the span is set to OK instead of ERROR [2]:

IGNORED_EXCEPTION_PATTERNS = [
    r"^Command\(",
    r"^ParentCommand\(",
]

Currently, this list covers Command and ParentCommand (which were added in v0.1.37 to fix a similar issue), but GraphInterrupt is not yet included [1].

Workarounds for now:

1. Monkey-patch the patterns list at app startup:

   from openinference.instrumentation.langchain._tracer import IGNORED_EXCEPTION_PATTERNS
   IGNORED_EXCEPTION_PATTERNS.append(r"^GraphInterrupt\(")

2. Wrap your interrupt logic so the GraphInterrupt doesn't propagate out of the traced span boundary (e.g., catch it at the graph invocation level before it reaches the instrumentation).

This seems like a valid candidate for an upstream fix — adding GraphInterrupt to the ignored patterns list similar to how Command/ParentCommand were handled. You could open a feature request on the OpenInference repo to have it included natively.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[omni-front]

You can use the error_filter option in the openinference-instrumentation-langchain configuration to exclude GraphInterrupt from being logged as an error:

from openinference.instrumentation.langchain import LangchainInstrumentation

instrumentation = LangchainInstrumentation(
    error_filter=lambda exc_type, exc_value, traceback: exc_type is not GraphInterrupt
)

This configuration allows you to specify exceptions that should not be treated as errors in OpenTelemetry spans.