scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-17 23:51:30 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 67

fix(otel): size the message ceiling so every vocabulary fits beside it

Browse source 
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
This commit is contained in:
yucheng 2026-09-10 08:00:51 +00:00
parent 492251c7bc
commit fcaf2d7d98
2 changed files with 14 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
litellm/integrations/otel/mappers/utils.py
View file

@ -32,15 +32,18 @@ core telemetry no matter how many vocabularies are configured.
"""
MAX_MESSAGE_ATTRS_PER_SPAN: Final = DEFAULT_SPAN_ATTRIBUTE_LIMIT // 4
MAX_MESSAGE_ATTRS_PER_SPAN: Final = DEFAULT_SPAN_ATTRIBUTE_LIMIT // 8
"""Span-wide ceiling on attributes spent spelling out chat messages per index.
A conversation is the other unbounded family: two attributes per message, for
the prompt and the response alike, on the same span. Past a few dozen turns the
family alone exceeds the span attribute limit and evicts the core telemetry
written before it. The ceiling covers both directions together, since a budget
handed to each direction separately doubles. The complete conversation still
rides the JSON blob attributes; only the per-index convenience keys are capped.
handed to each direction separately doubles. An eighth is the largest share
that still fits beside the tool ceiling and the core of every vocabulary at
once, request parameters, cost breakdown and identity included. The complete
conversation still rides the JSON blob attributes; only the per-index
convenience keys are capped.
"""

15
tests/test_litellm/integrations/otel/test_otel_v2_emitter.py
View file

@ -461,10 +461,11 @@ def _conversation_payload(turns, choices=1, **overrides):
)
def _conversation_span(mapper_names, payload):
def _conversation_span(mapper_names, payload, legacy_compat=False):
"""The exported LLM-call span for ``payload`` with content capture on."""
cfg = OpenTelemetryV2Config(
exporter="in_memory",
legacy_compat=legacy_compat,
mapper_names=list(mapper_names),
capture_message_content="span_only",
)
@ -541,13 +542,13 @@ def test_message_cap_is_shared_across_input_and_output():
) == (MAX_MESSAGE_ATTRS_PER_SPAN // 2)
def test_fully_populated_arize_span_stays_within_the_attribute_limit():
def test_fully_populated_span_with_every_vocabulary_stays_within_the_attribute_limit():
"""Every capped family maxed at once still leaves the whole core intact.
The Arize / Phoenix composition (``genai`` + ``openinference`` + ``legacy``)
with every request parameter, every cost component, a hundred-plus tools, a
two-hundred-turn prompt and twenty choices is the worst case the two
span-wide ceilings have to absorb together.
Every vocabulary in the registry plus ``legacy``, every request parameter,
every cost component, a hundred-plus tools, a two-hundred-turn prompt and
twenty choices is the worst case the two span-wide ceilings have to absorb
together.
"""
payload = _conversation_payload(
200,
@ -579,7 +580,7 @@ def test_fully_populated_arize_span_stays_within_the_attribute_limit():
)
},
)
span = _conversation_span(["genai", "openinference"], payload)
span = _conversation_span(["genai", "openinference", "langfuse", "weave", "langtrace"], payload, legacy_compat=True)
a = span.attributes
assert span.dropped_attributes == 0