OpenTelemetry tracing

reflex serve emits OpenTelemetry GenAI-semantic-convention spans for every /act call. Point Phoenix, Datadog, Honeycomb, New Relic, or any OTLP-compatible backend at it.

Install

pip install 'reflex-vla[tracing]'

Without the [tracing] extra, tracing no-ops silently. Server emits nothing and costs nothing.

Quick start

Phoenix (local dev)
Datadog

# Terminal 1 — Phoenix UI on :6006, OTLP gRPC collector on :4317
pip install arize-phoenix
phoenix serve

# Terminal 2 — Reflex pointing at it
reflex serve ./my-export/ --otel-endpoint localhost:4317 --otel-sample 1.0

Hit /act and open http://localhost:6006 — every request shows up as an act span.

# Datadog Agent exposes OTLP on :4317 when receivers.otlp is enabled in datadog.yaml
reflex serve ./my-export/ \
    --otel-endpoint localhost:4317 \
    --otel-sample 0.1

Spans land under service:reflex-vla in APM.

Flags

Flag	Default	Purpose
`--otel-endpoint`	`$OTEL_EXPORTER_OTLP_ENDPOINT` or `localhost:4317`	OTLP gRPC endpoint
`--otel-sample`	`1.0`	Fraction of traces to sample. Parent-based — child spans inherit the root’s decision.

Sampling guidance

Dev / staging: --otel-sample 1.0
Production, low traffic (under 100 RPS): --otel-sample 1.0 is fine — BatchSpanProcessor decouples export from the hot path
Production, high traffic (over 1000 RPS): --otel-sample 0.1 per OTel GenAI SemConv recommendation

You can also set OTEL_EXPORTER_OTLP_ENDPOINT as an env var.

Attribute reference

Every /act call produces one root span named act.

Standard OTel GenAI

Attribute	Type	Value
`gen_ai.operation.name`	string	`"act"`
`gen_ai.request.model`	string	Export directory path (post policy-versioning, this becomes the routing slot)

Robotics extensions (Reflex-specific, proposed upstream Phase 2)

Attribute	Type	Value
`gen_ai.action.embodiment`	string	Embodiment preset name or `"custom"`
`gen_ai.action.chunk_size`	int	Number of actions returned in this chunk
`gen_ai.action.denoise_steps`	int	Diffusion denoise iterations (when adaptive-denoise is active)

Reflex-specific (prefix intentional; not for upstream)

Attribute	Type	Value
`reflex.instruction`	string	First 512 chars of the instruction text (truncated — never raw bytes)
`reflex.state_dim`	int	Length of the proprio-state vector
`reflex.image_bytes`	int	Size of the posted image in bytes (size only — never the image data)
`reflex.rtc.episode_id`	string	Episode identifier when RTC adapter is active
`reflex.rtc.episode_reset`	bool	True on the first span of a new episode
`reflex.inference_mode`	string	`"decomposed"`, `"monolithic"`, `"native"`, etc.
`reflex.action_chunk_len`	int	Same as `gen_ai.action.chunk_size`; kept for backward compat
`error.type`	string	Populated only when the server returned an error

PII and security

Reflex explicitly does not attribute:

Raw image bytes (could contain faces, text, identifiers on screens)
API keys or auth headers
Raw state vectors when they exceed 512 dims (use reflex.state_dim for size only)

Instructions are truncated to 512 chars. If your instructions carry secrets, filter them client-side before calling /act.

Troubleshooting

“Tracing skipped — pip install reflex-vla[tracing] to enable” — the [tracing] extra isn’t installed.

Spans appear in Phoenix but not in Datadog — Datadog Agent probably has receivers.otlp disabled. Enable OTLP gRPC on :4317 in datadog.yaml.

--otel-sample 0.1 but I see every trace — parent-based sampling inherits from incoming parent context. If an upstream client is propagating a sampled trace context (W3C traceparent), Reflex honors it.

Export slowing down /act — BatchSpanProcessor is async; export failures shouldn’t bubble to request latency. If you see correlation, drop sampling or check collector health.