Imported from jeremylongshore/tons-of-skills-marketplace (skills/.curated/langchain-otel-observability/SKILL.md). Install upstream with npx skills add jeremylongshore/tons-of-skills-marketplace --skill langchain-otel-observability. Copyright stays with the author (MIT).

LangChain OTEL Observability (Python)

Overview

An engineer wires OpenTelemetry expecting to see prompts and responses in Honeycomb. The traces land — but only timing, model name, and token counts appear. The prompt body is blank. This is not a bug: it's the OTEL GenAI semantic-conventions privacy-safe default (P27), where OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT is off. The instinct is to flip it on and move on. On a multi-tenant workload that flip is a leak — the next engineer to search traces for Tenant A sees Tenant B's PII in the results, because redaction was supposed to happen upstream and never did.

A second trap lives inside LangGraph. A BaseCallbackHandler attached to the parent runnable never fires on inner agent tool calls, because LangGraph creates a child runtime per subgraph and callbacks do not inherit (P28). Spans inside subgraphs appear orphaned in the waterfall — or they do not appear at all — and SLO dashboards under-count latency on the exact calls that matter most: the nested agent loops.

This skill wires LangChain 1.0 / LangGraph 1.0 into an OTEL-native backend (Jaeger, Honeycomb, Grafana Tempo, Datadog) with a correct content-capture policy, subgraph-aware span propagation, and five LLM-specific SLOs (p95 / p99 latency, error rate, cost-per-request, TTFT) with burn-rate alerts. Pin: langchain-core 1.0.x, langgraph 1.0.x, opentelemetry-instrumentation-langchain >= 0.33, OTEL GenAI semconv as of 2026-04. Pain-catalog anchors: P27, P28 (and cross-references P04, P34, P37).

Prerequisites

• Python 3.10+
• langchain-core >= 1.0, < 2.0, langgraph >= 1.0, < 2.0
• An OTEL-native backend picked: Jaeger (dev), Honeycomb / Tempo / Datadog (prod)
• For multi-tenant: upstream redaction middleware already in place (see langchain-security-basics and langchain-middleware-patterns)
• Access to set env vars at deploy time (OTLP_ENDPOINT, API keys)

Instructions

Step 1 — Install the SDK and instrumentor, configure the exporter

pip install \
  opentelemetry-api \
  opentelemetry-sdk \
  opentelemetry-exporter-otlp-proto-http \
  "opentelemetry-instrumentation-langchain>=0.33"

import os
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.sdk.resources import Resource
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.instrumentation.langchain import LangchainInstrumentor

resource = Resource.create({
    "service.name": "my-langchain-app",
    "service.version": "1.0.0",
    "deployment.environment": os.getenv("ENV", "dev"),
})
provider = TracerProvider(resource=resource)
provider.add_span_processor(BatchSpanProcessor(
    OTLPSpanExporter(
        endpoint=os.environ["OTLP_ENDPOINT"],       # per-backend; see matrix
        headers=_parse_headers(os.getenv("OTLP_HEADERS", "")),
    ),
    max_queue_size=2048,        # spans buffered before drop; raise for high volume
    max_export_batch_size=512,  # batched export keeps per-span overhead under 1ms
))
trace.set_tracer_provider(provider)

LangchainInstrumentor().instrument()   # emits gen_ai.* attrs on every run

BatchSpanProcessor keeps per-span overhead well under 1 ms. Use SimpleSpanProcessor only in local dev — it blocks the call path per span.

Per-backend OTLP_ENDPOINT and header config lives in Backend Setup Matrix — Jaeger, Honeycomb, Grafana Tempo, Datadog.

Step 2 — Verify the GenAI attribute schema

Trigger one call and inspect what landed in the backend. LangChain 1.0 emits these gen_ai.* attributes natively on every chat-model span:

Missing anything? Likely a stale instrumentor version or an outdated provider package. The full emitted-vs-custom matrix plus LangGraph's span taxonomy (LangGraph.invoke → LangGraph.node.* → LangGraph.subgraph.*) is in GenAI Semantic Conventions.

Step 3 — Decide on prompt-content capture (critical — do not skip)

The engineer's instinct is to flip the capture flag to see prompts. Before flipping it, classify the workload into one of these buckets:

# trusted single-tenant ONLY
export OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=true
export TRACELOOP_TRACE_CONTENT=true   # OpenLLMetry alias; set both to be safe

Leave unset (default) anywhere else. To capture bodies in a multi-tenant system, wire redaction middleware upstream of the model call first — see Prompt Content Policy and cross-reference pack siblings langchain-security-basics (PII redaction middleware pattern, P34) and langchain-middleware-patterns (middleware order: redact → cache → model, P24). Failure pattern P27 — prompts missing from traces because capture was never opted in — is the #1 first-day OTEL complaint; make the decision explicit instead of surprise-flipping the flag in prod.

Step 4 — Propagate callbacks through subgraphs (P28)

LangGraph creates a child runtime per subgraph. Callbacks bound at the parent definition time do not inherit:

# WRONG — subagent spans orphaned or missing (P28)
agent = create_react_agent(model=llm, tools=tools).with_config(
    callbacks=[my_handler]  # bound at definition time; children do not see it
)
agent.invoke({"messages": [...]})

# RIGHT — pass callbacks at invocation via config; they propagate down
agent.invoke(
    {"messages": [...]},
    config={"callbacks": [my_handler]}  # invocation-time; inherited by children
)

The same rule applies to custom attribute handlers (e.g. the CostAttributeHandler in the semantic-conventions reference that stamps gen_ai.usage.cost_usd on each model span). Attach via config["callbacks"], never via .with_config(). Failure pattern P28 symptom: SLO dashboards show low latency because the slow nested spans are missing entirely, not because the nested calls are fast.

Step 5 — Define LLM SLOs and dashboards

Five SLIs matter from day one. All five derive from gen_ai.* span attributes — no second pipeline required:

Concrete Honeycomb / PromQL / Datadog queries for each SLI, plus multi-window multi-burn-rate alerts (14.4× / 1h fast burn, 6× / 6h slow burn), are in LLM SLO Dashboards.

Step 6 — Tune sampling

Defaults are wrong for two ends of the volume spectrum:

from opentelemetry.sdk.trace.sampling import TraceIdRatioBased

# Low/medium volume — keep everything for debuggability
# (< ~100 req/s) — SDK default 100% is fine

# High volume — head-sample, but carve out errors + slow spans via tail sampling
# at the OTEL Collector (see references/llm-slo-dashboards.md)
provider = TracerProvider(
    resource=resource,
    sampler=TraceIdRatioBased(0.10),  # 10% head sample
)

Watch out: head sampling at 10% means 90% of p99 outliers are discarded before they reach the backend — p99 metrics become noisy and biased toward the median. For tail-latency SLOs, move sampling to a Collector with tailsamplingprocessor so errors and slow spans (latency > 5000ms) are always kept while the rest is probabilistically sampled at 10%. Typical trace overhead with BatchSpanProcessor at the 512-span batch size: under 1 ms per span; recommended sampling rate for high-volume production is 1-10%.

Output

• OTEL exporter wired to a chosen backend (Jaeger / Honeycomb / Tempo / Datadog)
• opentelemetry-instrumentation-langchain emitting gen_ai.* attrs on every LangChain and LangGraph span
• Explicit prompt-content capture decision recorded against a workload bucket, with the multi-tenant guardrail enforced upstream
• Callbacks propagated via config["callbacks"] at invocation time so subgraph spans nest correctly under their parent node
• Five LLM SLOs (p95 / p99 latency, error rate, cost-per-request, TTFT) with dashboards and MWMBR burn-rate alerts
• Sampling strategy matched to workload volume and SLO precision needs

Error Handling

Examples

Running Jaeger locally for dev-loop tracing

Spin up Jaeger in Docker, point the SDK at http://localhost:4318/v1/traces, leave content capture on (it's dev, inputs are synthetic). You get a generic span waterfall — no LLM-specific UX, but good for verifying the instrumentor emits what you expect before paying for a SaaS backend.

See Backend Setup Matrix for the docker run command and SDK config.

Investigating an agent latency incident in Honeycomb

Honeycomb's BubbleUp over gen_ai.request.model, gen_ai.usage.input_tokens, and tool call count is the fastest path from "p95 spiked at 14:00" to "one specific tool took 20 s because the vectorstore was slow." Requires content-capture-off by default so you can turn the team loose on search without PII-leak worries.

See LLM SLO Dashboards for the exact Honeycomb query shape.

Dual-exporting during a LangSmith → Tempo migration

Register two BatchSpanProcessors — one to LangSmith's OTLP endpoint, one to Tempo. Run both for two weeks, compare waterfalls, cut over. LangSmith handles LLM-specific analytics; Tempo handles unified trace search across LLM and non-LLM services in your Grafana stack.

See Backend Setup Matrix dual-export section.

Resources

• OTEL GenAI semantic conventions
• OTEL Python SDK
• OpenLLMetry LangChain instrumentation
• Honeycomb OTLP ingest
• Grafana Tempo
• Datadog OTLP
• Google SRE — Alerting on SLOs
• Pack cross-references: langchain-security-basics (redaction, P34), langchain-middleware-patterns (order: redact → cache → model, P24), langchain-model-inference (cost callback pattern, P04)
• Pack pain catalog: docs/pain-catalog.md — P27 (content-capture default), P28 (subgraph callback propagation), P04 (cache token aggregation), P34 (prompt injection), P37 (secrets in env / prompts)