Context

When using LangSmith for tracing and monitoring, you may notice that some traces show an "incomplete" status. This can prevent accurate calculation of full trace metrics, including latency for the entire graph, even when individual step metrics are available.

Answer

An "incomplete" status in LangSmith means that the "end" event for a trace was never successfully received by the LangSmith server. This occurs because LangSmith ingests traces in a background thread to avoid adding overhead to your application.

Common causes for incomplete traces include:

• Process termination before all traces are uploaded

• Runs not sending completion signals

• End times not being recorded for all steps

• Network issues preventing data from reaching LangSmith

1. Update to the latest version of the LangSmith SDK

2. Before your program exits, call the following function to ensure all traces are submitted:

   wait_for_all_tracers()

3. Ensure all runs finish cleanly and send their completion signals

1. For self-hosted deployments: Configure background job settings to prevent status update failures:

   BG_JOB_ISOLATED_LOOPS=true
   BG_JOB_TIMEOUT_SECS=1800
   BG_JOB_SHUTDOWN_GRACE_PERIOD_SECS=300
   

Important notes:

• There is no explicit timeout set by LangSmith for traces

• LangSmith does not automatically retry incomplete traces

• You may still see latency metrics for individual completed steps, but the full trace metrics won't be reliable until marked complete

1. Runs that get stuck in PENDING state due to infrastructure issues may eventually appear as incomplete traces when status updates fail to persist

Additional information available at: https://docs.smith.langchain.com/observability/how_to_guides/trace_with_langchain#ensure-all-traces-are-submitted-before-exiting