DataDog/dd-trace-py v3.6.0

3.6.0

on GitHub

Upgrade Notes

• Profiling

  • Upgrades echion, improving the performance of stack sampler by replacing proc filesystem reads with clock_gettime().
  • Upgrades echion, improving the performance of the stack sampler by reusing memory and reducing the frequency of memory allocation.

• Tracing

  • Sampling formula updated to have the same as the one in the agent and other tracers.

New Features

• LLM Observability
  • Introduces support for submitting custom evaluations to LLM Observability via agent proxy mode.

Bug Fixes

• LLM Observability

  • Resolves an issue where annotating session_id on a LLM Observability span via LLMObs.annotate/annotation_context() as a tag only set it as a tag on the span, but did not update the span's actual session ID value.

• Profiling

  • Reduces the overhead of running the new stack-v2 sampling mechanism via adaptive sampling.
  • Upgrades echion to pick up fixes which enables the stack sampler to avoid extra calls to process_vm_readv() syscalls for Python frame objects for Python versions 3.11+.
  • Reduces overhead of the v2 stack sampler by removing unnecessary and expensive utf-8 validation
  • Reduces the overhead of code provenance reporting when stack v2 is enabled, the default stack sampler.
  • For CPython 3.10+, code provenance correctly classifies frozen modules in Python stdlib when stack v2 is enabled, the default stack sampler.
  • Upgrades echion to pick up a fix which resolves a potential crash. The stack sampler could read off of an empty queue of frames after failing to resolve specific frame information, triggering undefined behavior.

• Tracing

  • Fixes an issue where the sampling decision was hardcoded to USER_KEEP when extracting sampling priority from HTTP headers, if there was no x-datadog-sampling-priority header. The sampling decision is now made using the updated sampling formula.