Context

You may encounter errors when LangSmith attempts to upload trace data, particularly with larger payloads. These errors typically manifest as "Failed to send compressed multipart ingest" warnings with various HTTP status codes (403, 413, 422, 500) and can prevent traces from appearing in your LangSmith dashboard.

Answer

The multipart ingest errors can occur for several reasons depending on your setup and the specific error code. Here are the most common causes and solutions:

For 403 Forbidden Errors

This typically indicates an authentication issue:

1. Verify your API key is valid and hasn't expired

2. Regenerate your API key in the LangSmith dashboard

3. Update your application configuration with the new API key

4. Ensure the API key has the correct workspace permissions

For 413 Payload Too Large Errors

This occurs when your trace data exceeds size limits:

1. For self-hosted deployments, increase the frontend.maxBodySize parameter in your Helm values (default is 25MB)

2. Check if you have nginx ingress controllers with restrictive body size limits

3. Consider reducing the amount of data being traced or using attachments instead of including large files in state

For 422 Unprocessable Entity Errors

These errors often indicate data corruption or configuration issues:

1. Reduce batch size by setting max_batch_size_bytes on the Client() constructor. A starting value of 5MB helps isolate whether payload size is the cause:

   from langsmith import Client
   client = Client(max_batch_size_bytes=5_242_880)

2. For self-hosted deployments, ensure blob storage is properly configured if enabled

3. Check that platform-backend pods have write access to temporary directories

For Self-Hosted Deployments

Additional considerations for self-hosted LangSmith instances:

1. Ensure blob storage (S3/Azure) is properly configured with correct credentials

2. Verify platform-backend pods have sufficient resources and aren't memory constrained

3. Check that the platform-backend service can write to /tmp directories

4. Consider increasing the number of platform-backend replicas for high-throughput workloads

For Large State Management

If you're working with large data in LangGraph state:

1. Use LangSmith attachments for large files instead of including them in state

2. Implement selective field masking using Client(hide_inputs=..., hide_outputs=...)

3. Remove large fields from state after they're no longer needed in your workflow

General Troubleshooting

1. Ensure you're using the latest LangSmith SDK version

2. If uploads terminate prematurely (e.g. in serverless or short-lived environments), explicitly flush pending traces before your process exits:

   • Python: client.flush()

   • JS/TS: await client.awaitPendingTraceBatches()

   • See: https://docs.langchain.com/langsmith/serverless-environments

3. Check your network connectivity and any proxy configurations that might interfere with uploads

If these steps don't resolve the issue, check your LangSmith backend logs for more specific error details, particularly in the platform-backend pods for self-hosted deployments.