Langchain Observability

LangChain Observability (Python)

Overview

Engineer sets LANGCHAIN_TRACING_V2=true and LANGCHAIN_API_KEY=... from the 0.2 docs, restarts the service, and sees zero traces in LangSmith — no errors, no warnings. That is P26: in LangChain 1.0 the canonical env vars are LANGSMITH_TRACING and LANGSMITH_API_KEY. The LANGCHAIN_* names are soft-deprecated and fail silently on any chain that goes through 1.0 middleware or create_react_agent. One-line fix:

export LANGSMITH_TRACING=true
export LANGSMITH_API_KEY=lsv2_...
export LANGSMITH_PROJECT=my-service-prod

Next failure mode: a custom BaseCallbackHandler attached via chain.with_config(callbacks=[meter]) fires on the parent but is silent on LangGraph subgraphs and create_react_agent tool calls — token counts under-report by 30-70% vs the provider dashboard. That is P28: LangGraph creates a child runtime per subgraph, and bound callbacks do not propagate. Pass callbacks at invocation time instead:

await chain.ainvoke(inputs, config={"callbacks": [meter], "configurable": {"tenant_id": t}})

This skill walks through canonical LangSmith setup, a metric-callback template with tenant dimensions, invocation-time propagation, RunnableConfig trace tagging, and a decision tree for LangSmith-only vs OTEL-native (defer to langchain-otel-observability / L33 for OTEL-heavy). Pin: langchain-core 1.0.x, langgraph 1.0.x, langsmith current. LangSmith tracing adds <5ms per-span overhead; metric callbacks add <1ms per fire. Pain-catalog anchors: P26, P28, P04 (cache-token aggregation), P25 (retry double-counting).

Prerequisites

• Python 3.10+
• langchain-core >= 1.0, < 2.0, langgraph >= 1.0, < 2.0
• langsmith (bundled with langchain; upgrade to current for 1.0 env-var support)
• A LangSmith API key (lsv2_...) — free tier at https://smith.langchain.com
• Optional metric sinks: prometheus_client, statsd, or datadog Python packages

Instructions

Step 1 — Enable LangSmith with the canonical 1.0 env vars

LANGSMITH_TRACING=true is the switch. LANGSMITH_API_KEY authenticates. LANGSMITH_PROJECT groups traces by environment — use one project per service-env pair (myapp-prod, myapp-staging), not one per service.

# .env (loaded via python-dotenv or secret manager)
LANGSMITH_TRACING=true
LANGSMITH_API_KEY=lsv2_pt_...
LANGSMITH_PROJECT=my-service-prod

# Legacy fallback names (still work, soft-deprecated — do not use in new code):
# LANGCHAIN_TRACING_V2=true
# LANGCHAIN_API_KEY=lsv2_pt_...
# LANGCHAIN_PROJECT=my-service-prod

Verify in a REPL that the client sees the key before relying on it in production:

from langsmith import Client
c = Client()                       # reads LANGSMITH_API_KEY and LANGSMITH_ENDPOINT
print(c.list_projects(limit=1))   # raises LangSmithAuthError if key is wrong

Do NOT set both LANGCHAIN_TRACING_V2 and LANGSMITH_TRACING — mixed settings have caused stale project routing in 1.0.x. See P26.

For selective sampling in high-traffic services, set LANGSMITH_SAMPLING_RATE=0.1 (10% of runs). Full detail in LangSmith Setup.

Step 2 — Write a metric callback for per-request observability

Subclass BaseCallbackHandler. Record token_in, token_out, latency_ms, tool_calls, and error, tagged with a tenant_id dimension for downstream grouping.

import time
from langchain_core.callbacks import BaseCallbackHandler
from langchain_core.outputs import LLMResult

class MetricCallback(BaseCallbackHandler):
    """Per-LLM-call metrics tagged with tenant_id. Overhead <1ms per event."""

    def __init__(self, tenant_id: str, sink) -> None:
        self.tenant_id = tenant_id
        self.sink = sink
        self._starts: dict[str, float] = {}

    def on_llm_start(self, serialized, prompts, *, run_id, **kwargs) -> None:
        self._starts[str(run_id)] = time.perf_counter()

    def on_llm_end(self, response: LLMResult, *, run_id, **kwargs) -> None:
        t0 = self._starts.pop(str(run_id), time.perf_counter())
        elapsed_ms = (time.perf_counter() - t0) * 1000   # wall-clock latency
        tags = {"tenant_id": self.tenant_id}
        for gen in response.generations:
            for g in gen:
                meta = getattr(g.message, "usage_metadata", None) or {}
                self.sink.incr("llm.token_in",   meta.get("input_tokens", 0),  tags)
                self.sink.incr("llm.token_out",  meta.get("output_tokens", 0), tags)
                # P04 — aggregate Anthropic cache reads across calls
                cache = meta.get("input_token_details", {}).get("cache_read", 0)
                self.sink.incr("llm.cache_read", cache, tags)
        self.sink.hist("llm.latency_ms", elapsed_ms, tags)

    def on_llm_error(self, error, *, run_id, **kwargs) -> None:
        self._starts.pop(str(run_id), None)
        self.sink.incr("llm.error", 1, {"tenant_id": self.tenant_id,
                                         "error_type": type(error).__name__})

    def on_tool_end(self, output, *, run_id, **kwargs) -> None:
        self.sink.incr("llm.tool_calls", 1, {"tenant_id": self.tenant_id})

A thin sink protocol (incr, hist) swaps between Prometheus, StatsD, or Datadog. Alternative sinks (LangSmith-only, OTEL) do not need this callback at all — see Step 5. Full sink adapters and P25 retry dedupe in Custom Metrics Callback.

Step 3 — Pass callbacks via config["callbacks"] at invocation (P28)

This is the single most common observability bug in LangGraph 1.0 services. Binding callbacks at definition time does not propagate into subgraphs or create_react_agent tool nodes — those create child runtimes with their own callback scope.

# WRONG — fires on parent runnable only; silent on subgraphs (P28)
agent_bound = agent.with_config(callbacks=[MetricCallback(tenant_id, sink)])
result = await agent_bound.ainvoke(inputs)

# RIGHT — propagates to every runnable, subgraph, and tool call
meter = MetricCallback(tenant_id, sink)
result = await agent.ainvoke(
    inputs,
    config={
        "callbacks": [meter],
        "configurable": {"thread_id": session_id, "tenant_id": tenant_id},
        "tags": ["prod", f"tenant:{tenant_id}"],
        "metadata": {"request_id": req_id, "tier": "enterprise"},
    },
)

Construct the callback inside the request handler so it captures a fresh tenant_id per request — and in that pattern, invocation-time config is the only way callbacks reach subgraphs. See Trace Metadata and Tagging for the full RunnableConfig shape.

Step 4 — Tag and annotate traces via RunnableConfig

LangSmith indexes two per-request fields: tags (flat list, filterable) and metadata (key-value, searchable). Fix conventions early — LangSmith has no rename tool.

config = {
    "callbacks": [meter],
    "tags": [
        "env:prod",                # environment
        f"tenant:{tenant_id}",     # tenant
        f"tier:{tenant_tier}",     # plan tier
        f"feature:{feature_flag}", # A/B experiment arm
    ],
    "metadata": {
        "request_id": req_id,
        "user_id": user_id,
        "session_id": session_id,
        "app_version": os.environ["APP_VERSION"],
    },
    "run_name": "agent_main",      # LangSmith UI label; overrides chain class name
}

Hierarchical tag conventions (env:prod, tenant:acme, tier:enterprise) make LangSmith filters work. Free-form tags ("important", "check-me") do not. See Trace Metadata and Tagging.

Step 5 — Pick a sink and the stack shape

The callback handler is the integration point. Options, in decreasing order of fit:

• LangSmith only — zero additional overhead; tracing already covers latency and token accounting. Fine for solo dev, small teams, and LLM-native ops.
• Prometheus (pull) — best fit for Kubernetes + existing Prom stack. Export via prometheus_client HTTP endpoint. Watch tenant label cardinality.
• StatsD / Datadog (push) — UDP fire-and-forget; sub-1ms overhead. Safe on high-throughput async services. Use datadog.dogstatsd for tag support.
• OTEL native — multi-service distributed tracing. Defer to langchain-otel-observability (L33); do not reimplement here.

Decision tree:

Existing OTEL stack (Collector, Tempo, Jaeger)?
├── YES → OTEL-native (L33). LangSmith optional for prompt inspection.
└── NO  → LLM-specific features (prompt inspection, evals, queues) enough?
         ├── YES → LangSmith only. Add MetricCallback only for tenant cost.
         └── NO  → Hybrid: LangSmith for prompts + Prometheus/Datadog for SLOs.
                   See references/hybrid-langsmith-otel.md for split-point rules.

Mixing paths without a plan creates double-emission and conflicting trace IDs. See Custom Metrics Callback for Prometheus / StatsD / Datadog sink implementations, plus dedupe for P25 retry double-counts; see Hybrid LangSmith + OTEL for the split-point contract.

Step 6 — Feed runs back into evals

Real traffic is the best eval set. Route a sampled subset of production runs into a LangSmith annotation queue for human review; the queue feeds Dataset objects replayable against candidate models.

from langsmith import Client
Client().create_annotation_queue(
    name="prod-regressions",
    description="1% sample, weekly review",
)
# Add metadata={"eval_candidate": "true"} on 1% of runs — LangSmith UI has
# a rule to route into the queue by metadata filter.

Keep annotation queues under 500 runs/week (reviewers saturate past that). See LangSmith Setup for the queue and dataset flow.

Output

• LangSmith tracing on via LANGSMITH_TRACING / LANGSMITH_API_KEY / LANGSMITH_PROJECT with a langsmith.Client() smoke-check
• MetricCallback(BaseCallbackHandler) emitting token_in, token_out, cache_read, latency_ms, tool_calls, error tagged with tenant_id
• All chain invocations pass config={"callbacks": [...], ...} at invoke time so metrics propagate to subgraphs and agent tools
• RunnableConfig carries hierarchical tags (env:*, tenant:*, tier:*) and structured metadata (request_id, user_id, session_id)
• One metric sink wired (Prometheus, StatsD, Datadog, or LangSmith-only)
• Explicit choice recorded for LangSmith / OTEL / hybrid / custom

Error Handling

Examples

Multi-tenant SaaS: per-tenant cost dashboard

A production SaaS has 200 tenants on a shared LangGraph agent. Finance wants weekly cost reports per tenant. The MetricCallback records token_in, token_out, and cache_read tagged with tenant_id; Prometheus scrapes the /metrics endpoint; Grafana aggregates sum by (tenant_id) (rate(llm_token_out_total[1w])) * 0.0000015 for Sonnet output cost. The invocation-time config["callbacks"] propagation is load-bearing here — without it, subgraph tool calls (the bulk of token spend) go uncounted. See Custom Metrics Callback for the full Prometheus integration.

Debugging missing traces in staging

A team deploys a new LangGraph service to staging. No traces show up in LangSmith. Checking: (1) LANGSMITH_TRACING spelled correctly — yes; (2) API key valid — langsmith.Client().list_projects(limit=1) returns ok; (3) project name matches — LANGSMITH_PROJECT=myservice-staging. Traces appear in the default project, not myservice-staging. Root cause: the env var was set in the runtime env-file but the process was started before the env-file was sourced. Client() read LANGSMITH_PROJECT at import time. Fix: restart the process cleanly. See LangSmith Setup for the process-order checklist.

Feeding prod traffic to an eval dataset

A team wants to validate a Claude 4.6 → Claude 4.7 upgrade against recent prod runs. They add metadata={"eval_candidate": "pre-upgrade"} to 1% of runs for one week, create a LangSmith dataset from the tagged runs, then replay against the new model and diff outputs. The sampling rule lives in LangSmith UI, filtered by metadata.eval_candidate. See LangSmith Setup for the annotation-queue and dataset-creation flow.

Resources

• LangSmith Observability concepts
• LangSmith env variable reference
• LangChain callbacks (1.0)
• BaseCallbackHandler API
• RunnableConfig API
• For OTEL-native instrumentation: langchain-otel-observability (L33) in this pack
• Pack pain catalog: docs/pain-catalog.md (entries P04, P25, P26, P28)

Source: jeremylongshore/claude-code-plugins-plus-skills → skills/.curated/langchain-observability/SKILL.md

Also appears in: jeremylongshore/claude-code-plugins-plus-skills/plugins/saas-packs/langchain-py-pack/skills/langchain-observability/SKILL.md