Azure Function Host Crashes When Parsing PDF Documents with Docling Library

We've started using this python library docling to parse documents. Locally, that works perfectly. On Azure though, we experience the issue that parsing PDFs fails and makes the function host crash. This results in a downtime for the whole function, until it has restarted. Also, I can't find any helpful information as to why parsing PDFs makes the runtime crash. Parsing Word and Powerpoint documents works without any issues on Azure.

As part of the investigation, I've added Application Insights OpenTelemetry instrumentation to the app.

Insights gathered while trying to figure out the core issue:

1. Only PDFs are affected. From a docling perspective, this is reasonable because the processing pipeline for PDFs is quite different than the one for Word and PowerPoint.
2. In order to parse PDFs, docling is downloading appropriate AI models on demand. These downloads are successful, as can be observed in Application Insights.
3. It looks like the failure happens in the moment when we attempt the document conversion (having collected the documents and the necessary AI models to do so).
4. Once the function makes the function host crash (silently, i.e. no exception visible), one can observe the startup logs from a new instance of the same function. As part of these startup logs, the following exception shows up (usually 3 times):
   1. Transport endpoint is not connected : '/home/site/wwwroot/host.json'
   2. Because of that exception, the startup of the function immediately after the crash is not successful and it will take one or two tries to successfully spin up the function again.

Here are some of our environment details:

• We're using Azure functions for Python v2
• We're using python 3.11
• The app currently runs on a consumption plan.
• Dependencies

    azure-functions
    requests
    azure-identity
    
    # to fix (another) deployment issue on azure
    cryptography==43.0.3
    
    # Search
    azure-search-documents
    
    # Ai Service
    openai
    pydantic
    
    # OCR
    docling==2.31.0
    
    # graph ql client
    gql[all]
    
    # Monitoring
    azure-core-tracing-opentelemetry
    azure-monitor-opentelemetry