faker-observability-provider

Generate realistic, correlated, seedable observability test data: OpenTelemetry-style trace trees, log lines in 9 formats, W3C trace contexts, Kubernetes metadata, and per-language stacktraces — as a plain Faker provider with zero extra dependencies.

Built for testing log parsers, OTel pipelines, dashboards, alert rules, and SIEM detections inside your test suite, where a CLI log generator can't reach.

Installation

pip install faker-observability-provider

Quick Start

from faker import Faker
from faker_observability import ObservabilityProvider

fake = Faker()
fake.add_provider(ObservabilityProvider)

fake.traceparent()  # '00-7072ef7b6625afeb5efe109b815d7258-e86a5da65f5eef2f-01'
fake.log_line(fmt="logfmt")
# time=2026-07-06T03:00:38Z level=warn service=checkout-service msg="cache miss rate elevated"

spans = fake.trace(root_service="api-gateway", error=True)  # a correlated span tree
otlp = fake.otlp_json(spans)  # OTLP/JSON ResourceSpans envelope

💡 Tip: Run python showcase.py to see every feature end to end.

Correlated traces (the flagship)

trace() generates a span tree the way a real OTel deployment emits one — paired CLIENT/SERVER spans per hop, timing that actually adds up, a realistic service topology, and error propagation from the failing leaf to the root:

api-gateway · POST /api/checkout [SERVER] 368.7ms  ✗ ERROR
  api-gateway · POST [CLIENT] 27.2ms
    auth-service · POST /auth/token [SERVER] 23.4ms
      auth-service · SELECT shop.orders [CLIENT] 3.0ms
      auth-service · GET [CLIENT] 1.1ms
  api-gateway · POST [CLIENT] 168.2ms  ✗ ERROR
    checkout-service · POST /checkout [SERVER] 167.6ms  ✗ ERROR
      checkout-service · /inventory.InventoryService/CheckStock [CLIENT] 34.4ms
        inventory-service · /inventory.InventoryService/CheckStock [SERVER] 30.9ms
      checkout-service · GET [CLIENT] 80.1ms  ✗ ERROR
        order-service · GET /orders/{id} [SERVER] 78.8ms  ✗ ERROR
          order-service · SELECT shop.orders [CLIENT] 3.0ms  ✗ ERROR

Guaranteed invariants (all enforced by tests):

• one trace_id; unique span ids; a single root span
• children are time-contained inside their parent; sequential siblings never overlap
• every HTTP/gRPC hop is a CLIENT span paired with exactly one SERVER child, status codes matching
• every edge follows the built-in 20-service topology catalog (a DAG)
• error traces have exactly one origin span with an exception event — type and stacktrace matching the owning service's language — and ERROR status + 5xx propagated along the root path only
• span names and attributes follow the OTel semantic conventions ({method} {route} server names, bare-method client names, db.system.name, rpc.system, …)

otlp_json(spans) renders the OTLP/JSON encoding exactly: lowerCamelCase keys, integer kind enums, int64 timestamps as decimal strings, typed attribute values.

Seeding for reproducibility

Everything runs off Faker's seeded RNG — no global random, no uuid4():

fake = Faker()
fake.add_provider(ObservabilityProvider)
fake.seed_instance(1234)

fake.trace(start_time=datetime(2026, 1, 15, 12, 0, tzinfo=timezone.utc))
# byte-identical output on every run with the same seed

With the default start_time (anchored near "now"), same-seed runs reproduce identical ids, structure, and durations; pass an explicit start_time for byte-identical fixtures.

Log formats (9)

json            {"timestamp":"2026-07-06T03:00:38.858Z","level":"info","service":"checkout-service","message":"scheduled task finished","hostname":"checkout-service-bqzx2gc22n-sp4pz","pid":20723,"trace_id":"5815…","span_id":"c4cf…"}
access_json     {"host":"34.220.137.252","user-identifier":"-","datetime":"06/Jul/2026:02:55:52 +0000","method":"GET","request":"/api/checkout","protocol":"HTTP/1.1","status":200,"bytes":59441,"referer":"https://shop.example.com/"}
logfmt          time=2026-07-06T03:00:38Z level=warn service=checkout-service msg="cache miss rate elevated" trace_id=5815… span_id=c4cf…
apache_common   189.109.195.200 - - [06/Jul/2026:02:20:35 +0000] "GET /favicon.ico HTTP/1.1" 304 -
apache_combined 99.242.10.190 - frank [06/Jul/2026:02:15:40 +0000] "GET /static/app.css HTTP/1.1" 200 19338 "https://shop.example.com/products" "python-requests/2.32.0"
apache_error    [Mon Jul 06 02:42:38.621209 2026] [core:info] [pid 7509:tid 727028976795] [client 203.7.80.223:45947] AH02032: Hostname provided via SNI and hostname provided via HTTP have no compatible SSL setup
syslog_rfc3164  <134>Jul  6 02:28:57 inventory-service-nsvxp2nxp-5z76p inventory-service[22371]: user session created
syslog_rfc5424  <134>1 2026-07-06T03:00:38.921Z checkout-service-bqzx2gc22n-sp4pz checkout-service 22349 - [trace@32473 trace_id="5815…" span_id="c4cf…"] cache refreshed from upstream
nginx_error     2026/07/06 02:46:15 [info] 32380#0: *1557 no live upstreams while connecting to upstream, client: 34.220.199.183, server: fraud-service, request: "PUT /api/cart/547452 HTTP/1.1", host: "fraud-service"

Access-log fields are internally plausible: 204/304 render - bytes, write methods never hit static assets, /healthz and /metrics noise arrives with kube-probe/Prometheus user agents, and log_line(span=...) embeds the span's trace context (RFC 5424 carries it as spec-clean structured data under the RFC 5612 documentation enterprise number).

Available Methods

Why this instead of flog / telemetrygen?

Spec fidelity

Formats are implemented against: W3C Trace Context (00 version), OpenTelemetry semantic conventions (stable HTTP/DB/exception/resource registries; experimental messaging conventions as of 2026-07), OTLP 1.x JSON encoding, RFC 5424 / RFC 3164 (including the space-padded day), Apache CLF/error-log and nginx error-log formats, and the Kubernetes name-suffix alphabet. See AGENTS.md for the pinned verification sources.

Non-goals & roadmap

Not an OTel SDK or exporter, not a load generator, and it never emits over the network — it returns values. Metrics (Prometheus/OpenMetrics naming + exposition lines + OTel data points), async CONSUMER span continuations, incident_scenario(), and parallel fan-out timing are on the roadmap and will land additively.

Disclaimer

All generated data is synthetic test data for development and testing only. Names, formats, and topologies mimic real systems so records look realistic, but nothing here describes any real infrastructure or real telemetry.

Contributing

Issues and PRs are welcome — see CONTRIBUTING.md for setup, the checks CI runs, and the determinism, trace/log and catalog rules a change has to keep.

License

MIT