Orchestrator container fails to start: ImportError 'LogData' from opentelemetry.sdk._logs

[placerda]

Summary

During deployment of GPT-RAG v2.6.4, the orchestrator container app fails to start due to an ImportError in the OpenTelemetry / Azure Monitor SDK chain. Both container revisions are stuck in a non-running state. The root cause is an incompatibility between azure-monitor-opentelemetry and the installed opentelemetry-sdk version — LogData was moved or removed from opentelemetry.sdk._logs in a newer release.

Affected Release and Components

• GPT-RAG release/tag: v2.6.4
• Components involved:
  • gpt-rag-orchestrator — version/tag: v2.6.1
  • gpt-rag-ingestion
  • gpt-rag-ui
  • gpt-rag-mcp

Deployment Context

• Install type: Fresh install
• Region: Not specified
• Network isolation (NETWORK_ISOLATION): Not specified

Steps to Reproduce

1. Deploy GPT-RAG v2.6.4 infrastructure and orchestrator container app.
2. Check the orchestrator container app revisions — both are not in Running state.
3. Open the container app log stream.

Expected Behavior

The orchestrator container starts successfully and both revisions reach Running state.

Actual Behavior / Error

The container crashes on startup with the following import chain failure:

from .agent_strategy_factory import AgentStrategyFactory
  File "/app/src/strategies/agent_strategy_factory.py", line 6, in <module>
    from .mcp_strategy import McpStrategy
  File "/app/src/strategies/mcp_strategy.py", line 37, in <module>
    from telemetry import Telemetry
  File "/app/src/telemetry/__init__.py", line 1, in <module>
    from .telemetry import Telemetry, ExcludeTraceLogsFilter
  File "/app/src/telemetry/telemetry.py", line 6, in <module>
    from azure.monitor.opentelemetry import configure_azure_monitor
  File "/app/.venv/lib/python3.12/site-packages/azure/monitor/opentelemetry/__init__.py", line 7, in <module>
    from azure.monitor.opentelemetry._configure import configure_azure_monitor
  File "/app/.venv/lib/python3.12/site-packages/azure/monitor/opentelemetry/_configure.py", line 27, in <module>
    from azure.monitor.opentelemetry._constants import (
  File "/app/.venv/lib/python3.12/site-packages/azure/monitor/opentelemetry/_constants.py", line 7, in <module>
    from azure.monitor.opentelemetry.exporter._constants import (
  File "/app/.venv/lib/python3.12/site-packages/azure/monitor/opentelemetry/exporter/__init__.py", line 7, in <module>
    from azure.monitor.opentelemetry.exporter.export.logs._exporter import AzureMonitorLogExporter
  File "/app/.venv/lib/python3.12/site-packages/azure/monitor/opentelemetry/exporter/export/logs/_exporter.py", line 14, in <module>
    from opentelemetry.sdk._logs import LogData
ImportError: cannot import name 'LogData' from 'opentelemetry.sdk._logs'

Additional Context

The error indicates a version mismatch between azure-monitor-opentelemetry-exporter and opentelemetry-sdk. The LogData class was likely moved or renamed in a newer opentelemetry-sdk release, breaking the import in azure-monitor-opentelemetry-exporter. Pinning compatible versions of these packages in requirements.txt should resolve the issue.

[placerda]

Root Cause Analysis

This issue is caused by a breaking change in opentelemetry-sdk v1.39.0 (released 2025-12-03) combined with unpinned OpenTelemetry dependencies in the orchestrator's requirements.txt.

What changed in OpenTelemetry SDK v1.39.0

The LogData class was removed from opentelemetry.sdk._logs in opentelemetry-python v1.39.0 as a [BREAKING] change (PR #4676). The replacement is ReadableLogRecord.

Why the container crashes

The import chain that fails is:

1. agent_strategy_factory.py imports McpStrategy
2. McpStrategy imports Telemetry
3. Telemetry imports azure.monitor.opentelemetry
4. azure.monitor.opentelemetry imports azure.monitor.opentelemetry.exporter
5. The exporter's _exporter.py does from opentelemetry.sdk._logs import LogData
6. Since LogData no longer exists in SDK >= 1.39.0, the import fails

The root issue is a version mismatch between azure-monitor-opentelemetry-exporter and opentelemetry-sdk:

• Older exporter versions (prior to 1.0.0b49) reference LogData, which was removed in SDK 1.39.0
• Exporter 1.0.0b49 has been updated to use ReadableLogRecord instead

Why this is non-deterministic across deployments

The orchestrator's requirements.txt does not pin versions for the following packages:

• azure-monitor-opentelemetry
• azure-monitor-opentelemetry-exporter
• opentelemetry-instrumentation-httpx
• opentelemetry-instrumentation-fastapi

Since the Dockerfile runs pip install -r requirements.txt, the resolved versions depend entirely on what is available on PyPI at the time of the Docker image build. This creates a race condition:

• If pip resolves opentelemetry-sdk >= 1.39.0 and an older exporter (< 1.0.0b49), the container crashes
• If pip resolves the newer exporter 1.0.0b49 (which uses ReadableLogRecord), the container starts normally

This explains why some deployments succeed while others fail with the same source code.

Recommended Fix

Pin the OpenTelemetry packages in requirements.txt to known-compatible versions:

azure-monitor-opentelemetry==1.8.7
azure-monitor-opentelemetry-exporter==1.0.0b49
opentelemetry-instrumentation-httpx==0.61b0
opentelemetry-instrumentation-fastapi==0.61b0

These versions have been verified to work correctly with opentelemetry-sdk==1.40.0. The exporter 1.0.0b49 imports ReadableLogRecord instead of the removed LogData class.

Workaround

For anyone experiencing this issue immediately, rebuilding the Docker image (triggering a fresh pip install) may resolve it if PyPI now serves the compatible exporter version. However, this is not guaranteed without pinning versions.