●Change8

v4.8.0rc2

📦 datadog-sdkView on GitHub →
✨ 7 features🐛 12 fixes⚡ 1 deprecations🔧 8 symbols

Summary

This release introduces new tracing support for Azure CosmosDB and llama-index, adds an AI Guardrail integration for LiteLLM, and fixes several bugs related to profiling, internal thread leaks, and LLM Observability span reporting.

Migration Steps

  1. If using RAGAS integration for LLM Observability, switch to manually submitting evaluation results and refer to the external evaluation documentation.
  2. To enable eager flushing for CI Visibility test event buffering fixes in pytest-xdist worker crashes, set `DD_TRACE_PARTIAL_FLUSH_MIN_SPANS=1`.
  3. To enable automatic log correlation for CI Visibility, set `DD_AGENTLESS_LOG_SUBMISSION_ENABLED=true` for agentless setups, or `DD_LOGS_INJECTION=true` when using the Datadog Agent.

✨ New Features

  • ASM adds a LiteLLM proxy guardrail integration for Datadog AI Guard using `ddtrace.appsec.ai_guard.ai_guard.integrations.litellm.DatadogAIGuardGuardrail`. Requires `litellm>=1.46.1`.
  • Added tracing support for Azure CosmosDB, tracing CRUD operations on databases, containers, and items.
  • CI Visibility adds automatic log correlation and submission so that test logs appear alongside their corresponding test run in Datadog. Enable via `DD_AGENTLESS_LOG_SUBMISSION_ENABLED=true` or `DD_LOGS_INJECTION=true`.
  • Adds APM tracing and LLM Observability support for `llama-index-core>=0.11.0`, tracing LLM calls, query engines, retrievers, embeddings, and agents.
  • Tracing now supports exporting traces in OTLP HTTP/JSON format via libdatadog by setting `OTEL_TRACES_EXPORTER=otlp`.
  • LLM Observability introduces a `decorator` tag to spans traced by a function decorator.
  • LLM Observability experiments now accept a `pydantic_evals` `ReportEvaluator` as a summary evaluator if its `evaluate` return annotation is exactly `ScalarResult`.

🐛 Bug Fixes

  • Profiling: Fixed lock profiling samples not appearing in the Thread Timeline view for events collected on macOS.
  • Profiling: Fixed a rare crash that could occur post-fork in fork-based applications.
  • Profiling: Fixed a bug in Lock Profiling that could cause crashes when accessing attributes of custom Lock subclasses (e.g., in Ray).
  • Internal: Fixed a potential internal thread leak in fork-heavy applications.
  • Internal: Resolved an issue where a `ModuleNotFoundError` could be raised at startup in Python environments without the `_ctypes` extension module.
  • Internal: Fixed a crash that could occur post-fork in fork-heavy applications.
  • LLM Observability: Fixed incorrect span hierarchy when using ddtrace SDK alongside OTel-based instrumentation (e.g., Strands Agents), ensuring OTel gen_ai spans are nested correctly.
  • LLM Observability: Fixed multimodal OpenAI chat completion inputs being rendered as raw iterable objects; content parts are now materialized and formatted as readable text.
  • LLM Observability: Fixed `model_name` and `model_provider` reporting for AWS Bedrock LLM spans to use values matching backend pricing data.
  • LLM Observability: Fixed an issue where deferred tools (`defer_loading=True`) in Anthropic and OpenAI integrations caused span payloads to include full tool descriptions/schemas; definitions are now stripped, preserving only the tool name.
  • CI Visibility: Fixed loss of buffered test events caused by pytest-xdist worker crashes (`os._exit`, SIGKILL, segfault) by enabling eager flushing via `DD_TRACE_PARTIAL_FLUSH_MIN_SPANS=1`.
  • CI Visibility: Fixed git metadata upload falling back to sending full 30-day commit history upon failure of the `/search_commits` endpoint; the upload now aborts on failure.

Affected Symbols

ddtrace.appsec.ai_guard.integrations.litellm.DatadogAIGuardGuardrailllama-index-coreOTEL_TRACES_EXPORTERpydantic_evals.evaluators.ReportEvaluatorScalarResultReportEvaluatorContextReportAnalysisLLMObs

⚡ Deprecations

  • Support for the RAGAS integration in LLM Observability has been removed. Users should manually submit RAGAS evaluation results as an alternative.