OTLP export from the gateway
PALM SOUQ can export to any OTLP endpoint over gRPC or HTTP: an OpenTelemetry Collector you run, Grafana, Honeycomb, New Relic, or your own pipeline. Each proxied request becomes a span, and aggregate token and cost figures are exported as metrics on an interval you choose.
If your application already propagates W3C trace context, the gateway continues the trace. The PALM SOUQ span appears as a child of your application span, so a single trace shows your handler, the gateway decision and the provider call end to end.
Semantic conventions
Spans use the OpenTelemetry GenAI semantic conventions where they exist: gen_ai.request.model, gen_ai.response.model, gen_ai.usage.input_tokens, gen_ai.usage.output_tokens and the operation name. These conventions are still evolving in the OpenTelemetry project, and PALM SOUQ tracks the current version so dashboards built on them keep working.
PALM SOUQ-specific attributes use a palmsouq prefix: palmsouq.tenant, palmsouq.tier, palmsouq.feature, palmsouq.cost_usd, palmsouq.cached_tokens, palmsouq.policy (served, fallback or blocked) and palmsouq.breaker.heuristic when a breaker fires. Prompt and completion content is never attached to spans.
When to choose OTel over a vendor exporter
Teams pick the OpenTelemetry export when they want to own their telemetry pipeline, route LLM data to more than one backend, or keep data inside their own network. With a Collector in your VPC you can sample, enrich or redact before anything leaves.
OpenTelemetry export is included in the Growth plan and above. It runs off the request path, so a slow collector never slows a model call; if the endpoint is unreachable, data is buffered for a short window and then dropped with a counter you can alert on.
Rolling it out safely
Send PALM SOUQ telemetry to a Collector you control rather than straight to a vendor endpoint. The Collector lets you sample traces, drop attributes you do not need and fan out to more than one backend, and it gives you one place to change routing later.
Begin with a modest trace sample ratio and full metrics. Metrics carry the per-tenant cost and token totals, which are what most dashboards need, while traces are most useful for debugging individual sessions.
If your organisation already standardises on OpenTelemetry for services, the gateway span will appear in existing service maps as a dependency between your application and the provider. That makes the latency the gateway adds visible to everyone who reads those maps, which is a useful check on the 5 ms budget.
