Span attributes

Marturia indexes a small set of well-known attributes for first-class filtering and aggregation. Anything else you set is preserved verbatim and visible in the span drawer — never silently dropped.

LLM observability (recommended)

The dashboard's "Cost by model", "Top slowest spans", and per-trace token summaries all read these attributes. If you're calling an LLM, set them.

Attribute	Type	Notes
`llm.model`	string	Model identifier exactly as the provider returns it. Example: `gpt-4o-mini`.
`llm.cost_usd`	double	Cost of this single call in USD. Sum across spans gives the dashboard's spend tile.
`llm.input_tokens`	int	Prompt token count. Provider's billing units, not character count.
`llm.output_tokens`	int	Completion token count.
`llm.temperature`	double	Optional — useful when debugging variance between runs.

User and session

These let you filter the span list to "everything for this customer" or "this user's last 10 sessions" — the most common debugging starting point.

Attribute	Type	Notes
`user.id`	string	Stable identifier for the end user. Don't use email or PII.
`session.id`	string	Group spans into a session boundary (chat thread, request batch, etc.).
`tenant.id`	string	If you operate as multi-tenant SaaS, label spans with your customer's tenant.

Errors

Set span.status to ERROR for any failed operation, and use OTel's standard exception event for stack traces. The dashboard's error-rate tile counts spans where status = ERROR.

# Python
from opentelemetry.trace import Status, StatusCode

try:
    result = call_llm()
except Exception as e:
    span.set_status(Status(StatusCode.ERROR, str(e)))
    span.record_exception(e)
    raise

Semantic conventions

Marturia respects the standard OpenTelemetry semantic conventions — HTTP, database, RPC, messaging. If you're using auto-instrumentation, those attributes already flow through unchanged.

Cardinality

Avoid high-cardinality attributes you don't need. Storing every UUID in attribute keys (rather than values) bloats indexes and slows queries. Use a small fixed set of keys; put unique values in the value field.

Custom attributes

Any attribute we don't index is still stored verbatim and visible in the span drawer. There's no schema enforcement on custom attributes — name them whatever makes sense to your engineers. The dashboard's free-text search hits them.

← Back

OTLP setup

Creating receipts