OpenTelemetry in Production

Configure sampling, batch export tuning, and graceful flush so your traces hold up under production load

Exporting 100% of traces from a high-throughput app generates significant backend ingestion cost, and spans buffered in BatchSpanProcessor are silently dropped when a runner shuts down if they are not flushed before SIGKILL arrives. For basic tracing setup, start with Custom Traces.

Conditional Tracing

Disable tracing entirely with an environment variable when you do not need it. When no TracerProvider is configured, the OpenTelemetry SDK returns a no-op tracer. All start_as_current_span calls become zero-cost no-ops and no export threads are started.

Python

def setup_tracer(service_name: str):
    import os
    from opentelemetry import trace
    from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
    from opentelemetry.sdk.resources import Resource
    from opentelemetry.sdk.trace import TracerProvider
    from opentelemetry.sdk.trace.export import BatchSpanProcessor

    if os.environ.get("ENABLE_TRACING", "true").lower() != "true":
        return trace.get_tracer(service_name), None

    resource = Resource.create({"service.name": service_name})
    provider = TracerProvider(resource=resource)
    provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))
    trace.set_tracer_provider(provider)
    return trace.get_tracer(service_name), provider

Set ENABLE_TRACING=false as a fal secret to turn off tracing without redeploying:

fal secrets set ENABLE_TRACING=false

Sampling

Sampling reduces the fraction of traces exported. Use it when your app handles many requests per second and exporting every trace would exceed your backend’s ingestion limits or add measurable latency.

ParentBased(root=TraceIdRatioBased(rate)) is the recommended sampler for services that receive calls from other services. It honours the sampling decision made upstream (so a trace that was sampled by a caller is always sampled here too), and applies ratio-based sampling only to new root spans.

Python

def setup_tracer(service_name: str):
    import os
    from opentelemetry import trace
    from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
    from opentelemetry.sdk.resources import Resource
    from opentelemetry.sdk.trace import TracerProvider
    from opentelemetry.sdk.trace.export import BatchSpanProcessor
    from opentelemetry.sdk.trace.sampling import ParentBased, TraceIdRatioBased

    if os.environ.get("ENABLE_TRACING", "true").lower() != "true":
        return trace.get_tracer(service_name), None

    # OTEL_SAMPLE_RATE=0.1 exports 10% of root spans; default is 1.0 (all spans).
    sample_rate = float(os.environ.get("OTEL_SAMPLE_RATE", "1.0"))
    sampler = ParentBased(root=TraceIdRatioBased(sample_rate))

    resource = Resource.create({"service.name": service_name})
    provider = TracerProvider(resource=resource, sampler=sampler)
    provider.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))
    trace.set_tracer_provider(provider)
    return trace.get_tracer(service_name), provider

Start with OTEL_SAMPLE_RATE=0.1 under heavy load and increase it if you need more coverage for debugging.

Batch Export Tuning

BatchSpanProcessor buffers spans in memory and exports them in bulk. Tune these parameters if you see dropped spans under burst traffic or export timeouts in your backend logs.

Python

processor = BatchSpanProcessor(
    OTLPSpanExporter(),
    # Drop oldest spans when the in-memory queue is full.
    # Increase this if you see "SpanProcessor queue is full" warnings.
    max_queue_size=2048,
    # Spans per HTTP request to the collector.
    # Larger batches reduce request overhead at the cost of higher memory use.
    max_export_batch_size=512,
    # How often to flush the buffer, in milliseconds.
    schedule_delay_millis=5000,
    # Give up on an export request after this long.
    export_timeout_millis=30_000,
)

Parameter	Default	When to increase
max_queue_size	2048	Seeing dropped spans under burst traffic
max_export_batch_size	512	High span volume with low per-request overhead
schedule_delay_millis	5000	Backend is overwhelmed by frequent export requests
export_timeout_millis	30000	Backend is slow or on a high-latency network

Flushing on Shutdown

fal gives runners a 5-second grace period between SIGTERM and SIGKILL. teardown() runs during that window, after all in-flight requests finish. See App Lifecycle for the full shutdown sequence.

Call force_flush() in teardown(), not in your endpoint handler. Per-request flushing adds latency on every call; BatchSpanProcessor handles periodic exports on its own schedule.

Python

import os

import fal
from fal.toolkit import Image
from pydantic import BaseModel, Field

def setup_tracer(service_name: str):
    from opentelemetry import trace
    from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
    from opentelemetry.sdk.resources import Resource
    from opentelemetry.sdk.trace import TracerProvider
    from opentelemetry.sdk.trace.export import BatchSpanProcessor
    from opentelemetry.sdk.trace.sampling import ParentBased, TraceIdRatioBased

    if os.environ.get("ENABLE_TRACING", "true").lower() != "true":
        return trace.get_tracer(service_name), None

    sample_rate = float(os.environ.get("OTEL_SAMPLE_RATE", "1.0"))
    sampler = ParentBased(root=TraceIdRatioBased(sample_rate))

    resource = Resource.create({"service.name": service_name})
    provider = TracerProvider(resource=resource, sampler=sampler)
    provider.add_span_processor(
        BatchSpanProcessor(
            OTLPSpanExporter(),
            max_queue_size=2048,
            max_export_batch_size=512,
            schedule_delay_millis=5000,
            export_timeout_millis=30_000,
        )
    )
    trace.set_tracer_provider(provider)
    return trace.get_tracer(service_name), provider


class Input(BaseModel):
    prompt: str = Field(description="The prompt to generate an image from")
    num_inference_steps: int = Field(default=20)


class Output(BaseModel):
    image: Image
    trace_id: str


class ProductionApp(fal.App):
    machine_type = "GPU-A100"
    requirements = [
        "hf-transfer==0.1.9",
        "diffusers[torch]==0.32.2",
        "torch==2.10.0",
        "transformers[sentencepiece]==4.51.0",
        "accelerate==1.6.0",
        "opentelemetry-sdk==1.41.0",
        "opentelemetry-exporter-otlp-proto-http==1.41.0",
    ]

    def setup(self):
        # Enable HF Transfer for faster downloads
        os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"

        import torch
        from diffusers import StableDiffusionXLPipeline

        self.tracer, self.tracer_provider = setup_tracer("text-to-image")

        self.pipe = StableDiffusionXLPipeline.from_pretrained(
            "stabilityai/stable-diffusion-xl-base-1.0",
            torch_dtype=torch.float16,
            variant="fp16",
            use_safetensors=True,
        ).to("cuda")
        self.pipe("warmup")

    @fal.endpoint("/")
    def run(self, input: Input) -> Output:
        with self.tracer.start_as_current_span("text-to-image") as root:
            root.set_attribute("model.name", "stable-diffusion-xl-base-1.0")

            with self.tracer.start_as_current_span("inference"):
                result = self.pipe(input.prompt, num_inference_steps=input.num_inference_steps)

            with self.tracer.start_as_current_span("upload"):
                image = Image.from_pil(result.images[0])

            trace_id = format(root.get_span_context().trace_id, "032x")

        return Output(image=image, trace_id=trace_id)

    def teardown(self):
        # Flush buffered spans before SIGKILL (5s grace period).
        # Use 4000ms to leave headroom for SIGKILL at 5000ms.
        if self.tracer_provider:
            self.tracer_provider.force_flush(timeout_millis=4000)

The 5-second grace period is a shared budget across handle_exit(), remaining in-flight requests, teardown(), and force_flush(). If your endpoint handles long-running requests, use handle_exit() to signal them to stop early so teardown() has enough time to flush. See App Lifecycle for the full shutdown sequence.

What’s Next

Custom Traces

Add spans to trace inference stages end to end

Custom Metrics

Counters, histograms, and gauges for your inference workload

Cross-Service Tracing

Connect traces across multiple fal apps

App Lifecycle

Full shutdown sequence and grace period details