dagster-otel

OpenTelemetry tracing for Dagster ops and assets -- with trace/span IDs correlated into your own log lines -- without giving up ownership of your op/asset definitions to a third-party decorator, and without monkeypatching Dagster internals.

Status: early release, self-tested locally against real Dagster runs (multiprocess, k8s_job_executor, retry-from-failure) + a real trace backend.

Table of Contents

• Installation
• Usage
• Configuration
• Compatibility
• Why this exists
• Contributing
• License

Installation

pip install dagster-otel

Usage

from dagster import asset, job, op

from dagster_otel import traced

@op(...)                # Dagster's own @op still owns op-ness; @traced() is a thin
@traced()                # layer underneath. No @resource/required_resource_keys, no
def upstream_op(context) -> int:  # manual "root" step -- the first @traced() step to
    ...                            # run in a given run just becomes the root.

@op(...)
@traced()
def downstream_op(context, x: int) -> int:
    ...

@asset(...)
@traced()  # same decorator, works for assets too
def downstream_asset(context) -> None:
    ...

@op(...)
@traced  # bare works too, like @op/@asset themselves -- same as @traced()
def another_op(context) -> None:
    ...

@job(...)
def my_job():
    downstream_op(upstream_op())

Works across Dagster's multiprocess and k8s_job_executor executors: each step usually runs in its own process, sometimes on its own node, so trace context is propagated via Dagster's own run storage rather than in-process memory. See docs/design.md for how, and what's verified vs. still assumed.

For @dbt_assets, dagster_otel.dbt.traced_dbt() is a drop-in replacement for @traced() that additionally opens a child span per dbt node (model/seed/test), keyed by the real Dagster asset_key/check_name -- no changes needed to the function body:

from dagster_otel.dbt import traced_dbt

@dbt_assets(manifest=...)
@traced_dbt()
def my_dbt_assets(context, dbt: DbtCliResource):
    yield from dbt.cli(["build"], context=context).stream()

Configuration

Standard OTel environment variables -- nothing bespoke:

@traced() reads these itself (idempotently) the first time it runs in a process -- there's nothing else to wire up, no @resource/required_resource_keys needed. Call configure() yourself only if you want configuration to happen eagerly (e.g. at Definitions load time) rather than lazily on first use.

Without OTEL_EXPORTER_OTLP_ENDPOINT/..._TRACES_ENDPOINT set, no real OTLP exporter is created at all -- spans are still created (propagation and log correlation keep working), just never sent anywhere, so trying @traced() with zero setup never makes a surprise network call. See docs/design.md for the one deliberate tradeoff this makes.

Nesting a whole run's trace under an external caller's (a CI/CD pipeline, a scheduler, another OTel-instrumented system) is a run tag, not an env var -- set EXTERNAL_TRACE_CONTEXT_TAG_KEY (exported from dagster_otel) at launch time:

from dagster_otel import EXTERNAL_TRACE_CONTEXT_TAG_KEY

carrier: dict[str, str] = {}
TraceContextTextMapPropagator().inject(carrier)  # from your own active span
my_job.execute_in_process(tags={EXTERNAL_TRACE_CONTEXT_TAG_KEY: json.dumps(carrier)})

Every root step in the run (the ones that would otherwise seed a fresh trace) checks for this tag first. See docs/design.md for the full verification.

Compatibility

Built and verified against Dagster 1.13.22 and opentelemetry-sdk 1.44.0 (pyproject.toml/uv.lock) -- this is the combination every behavior described here has actually been checked against, including the multiprocess/k8s_job_executor/ retry-from-failure verification in docs/design.md. pyproject.toml declares a much wider floor (dagster >= 1.5) since nothing here relies on version-specific Dagster internals beyond what's documented as an accepted-risk private-API dependency there -- but that wide range isn't individually spot-checked the way it is for dagster-prometheus-exporter. If you hit an incompatibility on another version, please open an issue.

Requires Python 3.10+ (matches Dagster's own floor).

Why this exists

See docs/design.md for the full rationale, including a comparison against prior art (formenergy-observability, a monkeypatch-based prototype) and the design decisions (no monkeypatching, decorators stack under Dagster's own @op/@asset rather than replacing it, log correlation via a public logging.Filter on context.log).

Contributing

See CONTRIBUTING.md for how to set up your toolchain, run checks locally, and submit a pull request. Bug reports and feature requests go through GitHub issues; a security vulnerability goes to SECURITY.md instead.

License

MIT -- see LICENSE.