obsforge 0.1.2

Zero-infra drop-in structured logging & observability for Python (Grafana/Loki, Elastic, any stack)

obsforge

Zero-infra, drop-in structured logging & observability for Python.

Add it to your code and your logs come out as structured JSON on stdout — ready for your existing shipper (promtail, filebeat, fluent-bit, otel-collector) to deliver to Grafana/Loki, Elastic, or any stack. obsforge does not run, require, or manage any extra infrastructure. It writes to stdout; your platform already knows how to collect that.

pip install obsforge

import logging, obsforge

obsforge.install_logging_bridge()          # one line, no bootstrap, no code rewrite

logging.getLogger("checkout.auth").warning("login failed", extra={"reason": "bad_password"})

That's the whole integration. Set the service name once via the OpenTelemetry-standard env var and every event is attributable:

export OTEL_SERVICE_NAME=checkout         # otherwise logs show "unknown-service"

---

What problem does it solve?

Most apps log unstructured strings, then teams bolt on regex parsing, inconsistent fields, and accidental high-cardinality labels that blow up Loki. Tracing and logs live in separate worlds. And "observability SDKs" often drag in collectors, agents, or external services you have to operate.

obsforge takes the opposite stance:

• Semantic events, not strings. Every log is a canonical event with a stable schema (event, severity, service, correlation, trace, ...), not a free-form message.
• Cardinality-safe by construction. Output is a three-section document where low-cardinality keys become labels and high-cardinality identifiers (trace_id, user_id, tenant_id, ...) are never labels.
• Correlation built in. W3C trace context + correlation ids flow across HTTP, Celery, Kafka, RabbitMQ, asyncio and background workers — so logs and traces share ids.
• Zero extra infrastructure. stdout-first, JSON-first. No collector, agent, or service is required by the library.
• Drop-in. Works with your existing logging calls; no rewrite.

What it is not

It is not an agent, a daemon, or a hosted service. It doesn't ship logs itself — your platform's collector does. It has no external service dependency in the default path.

---

Features

• 🪵 stdlib logging bridge — route existing logger.* calls through obsforge with one line.
• 🧱 Canonical event model — typed (pydantic), with HTTP / DB / exception / cache / security / business / dependency context.
• 🏷️ Loki label governance — LokiLabelPolicy + StructuredMetadataPolicy; no cardinality footguns.
• 🔗 Distributed correlation — W3C traceparent / baggage, propagated across services, queues, and tasks.
• 🛡️ Security by default — deep PII scrubbing (email, JWT/bearer, card, SSN, IP) across every string field; identity validation on ingress.
• 🧯 Fail-open pipeline — a telemetry fault never breaks your request.
• 🧩 Framework adapters — FastAPI/Starlette, Django/DRF middleware; HTTP-client and DB instrumentation.
• 🔭 Optional OpenTelemetry — opt-in log export + trace bridge (off by default; no collector needed otherwise).
• ✅ Typed & tested — mypy --strict clean, layered test suite, benchmarked.

---

Installation

pip install obsforge                 # core (orjson, pydantic — no other runtime deps)
pip install "obsforge[fastapi]"      # FastAPI / Starlette middleware
pip install "obsforge[django]"       # Django middleware
pip install "obsforge[drf]"          # Django REST Framework exception handler
pip install "obsforge[celery]"       # Celery task correlation
pip install "obsforge[db]"           # SQLAlchemy / psycopg / asyncpg / aiomysql
pip install "obsforge[http]"         # httpx / requests / aiohttp client instrumentation
pip install "obsforge[otel]"         # optional OpenTelemetry export

Python 3.11+ (CI runs 3.11, 3.12, 3.13).

---

Quickstart

1) Drop-in for existing logging code (recommended start)

The minimal integration is a single call — no bootstrap(), no objects to thread through your code. The bridge initializes a default SDK on first use:

import logging, obsforge

obsforge.install_logging_bridge()        # set OTEL_SERVICE_NAME in the environment

log = logging.getLogger("checkout.orders")
log.info("order placed", extra={"order_id": "o_123", "amount_cents": 4200})

try:
    charge(order)
except PaymentError:
    log.exception("charge failed", extra={"order_id": "o_123"})   # traceback captured

When you need explicit configuration (environment, Loki preset, redaction), bootstrap once and the bridge reuses that SDK automatically:

sdk = obsforge.bootstrap(obsforge.ObsforgeSettings(
    service_name="checkout",
    environment="production",
    loki={"preset": "prod"},     # dev | staging | prod label policy
))
obsforge.install_logging_bridge()        # reuses the client registered by bootstrap()
# (or pass it explicitly: obsforge.install_logging_bridge(sdk.event_client))

logger.exception(...) captures a structured exception context. Logs emitted inside a request automatically inherit its correlation id.

2) Explicit semantic events

sdk.logger.log_sync("auth.login.failed", severity=obsforge.Severity.WARNING, reason="bad_password")
await sdk.logger.log("billing.invoice.paid", amount_cents=4200)

@obsforge.instrument(sdk.event_client, "checkout.order.place")
def place_order(...): ...

3) FastAPI

import obsforge
from fastapi import FastAPI
from obsforge.integrations.fastapi.middleware import FastAPIObservabilityMiddleware

sdk = obsforge.bootstrap(obsforge.ObsforgeSettings(service_name="api"))
app = FastAPI()
app.add_middleware(FastAPIObservabilityMiddleware, api_engine=sdk.api_engine)

See examples/ for FastAPI, Django, worker, and plain-logging apps.

4) OpenTelemetry (optional)

pip install "obsforge[otel]"

from opentelemetry.sdk._logs import LoggerProvider
from opentelemetry.sdk._logs.export import BatchLogRecordProcessor
from opentelemetry.exporter.otlp.proto.http._log_exporter import OTLPLogExporter

provider = LoggerProvider()
provider.add_log_record_processor(BatchLogRecordProcessor(OTLPLogExporter()))

sdk = obsforge.bootstrap(
    obsforge.ObsforgeSettings(service_name="api", otel={"logs_enabled": True}),
    otel_logger_provider=provider,   # inject the configured provider to actually export
)
# Events are also emitted as OTel LogRecords with trace_id in the record context.

OTel export is off by default — stdout remains the default path. Opting in (logs_enabled=True / traces_enabled=True) requires injecting the matching provider (otel_logger_provider= / otel_tracer_provider=); enabling it without one raises ConfigurationError rather than silently dropping every record.

---

What the output looks like

Each event is one JSON line on stdout, split into three sections:

{
  "labels": { "service": "checkout", "environment": "production", "level": "warning", "event_kind": "api", "outcome": "failed" },
  "structured_metadata": { "trace_id": "0af7...", "correlation_id": "...", "user_id": "u_1", "fingerprint": "..." },
  "body": { "event_name": "auth.login.failed", "message": "login failed", "severity": "warning", "...": "..." }
}

• labels → low-cardinality, safe to index in Loki.
• structured_metadata → high-cardinality ids; queryable, never labels.
• body → the full canonical event (your log line).

How your stack consumes it (no library infra)

Backend	How
Grafana / Loki	promtail or Alloy maps labels→labels, structured_metadata→structured metadata, body→log line (config)
Elastic	filebeat / fluent-bit ingest the JSON line; query by body.* / structured_metadata.*
Anything	it's JSON on stdout — if your platform collects stdout, it just works

Grafana / Loki (promtail snippet)

Map the three sections explicitly — never auto-flatten the whole document into labels:

pipeline_stages:
  - json:
      expressions: { labels: labels, structured_metadata: structured_metadata, body: body }
  - labels:
      service:
      environment:
      level:
      event_kind:
      outcome:
  - structured_metadata:
      trace_id:
      correlation_id:
      user_id:
  - output:
      source: body

Full promtail/Alloy + OTLP reference: docs/operational/loki_otel_setup.md.

---

Governance & safety (defaults you get for free)

• No cardinality footguns: trace_id, user_id, tenant_id, request_id, correlation_id, session_id, event_id can never be emitted as labels.
• PII scrubbing: emails, bearer/JWT tokens, card numbers, SSNs and IPs are redacted from every string field — message, payload previews, headers, exception text, DB query text.
• Fail-open: any error inside the pipeline is logged-and-dropped, never raised into your hot path.
• Trusted propagation: inbound identifiers are length- and control-char-validated; baggage is capped.

---

Configuration

obsforge.ObsforgeSettings(
    service_name="checkout",
    environment="production",
    min_severity=obsforge.Severity.INFO,      # sampler drops below-threshold events
    loki={"preset": "prod"},                  # label policy preset
    security={"scrub_pii": True, "trust_inbound_identity": True},
    otel={"logs_enabled": False},             # opt-in; requires the otel extra + a provider
)

Per-subsystem settings exist for HTTP, DB, exceptions, and distributed correlation — see obsforge.config.settings.

---

Performance

~28k–37k events/sec/core single-threaded; ~35 µs per logger.info(...) through the bridge — about 8–9× plain stdlib formatting, and that delta buys the typed event, PII scrubbing, correlation, and Loki-governed JSON. Throughput holds steady under concurrent load (Python 3.13, no-op sink). Run it yourself:

python benchmarks/run.py            # overhead + stdlib baseline + concurrency sweep

Details and methodology: docs/operational/performance.md.

---

Documentation

• Adoption guide — full integration walkthrough
• Architecture — layers and canonical event shape
• Loki / OTel setup — promtail/Alloy + OTLP config
• Performance
• Support matrix
• Publishing
• Changelog

---

Compatibility

Python	3.11, 3.12, 3.13 (CI matrix)
Frameworks	FastAPI/Starlette, Django/DRF (via extras)
DB drivers	SQLAlchemy, psycopg, asyncpg, aiomysql (via db extra)
HTTP clients	httpx, requests, aiohttp (via http extra)
Backends	Grafana/Loki, Elastic, any stdout collector

Postgres, Kafka, RabbitMQ, Celery brokers and an OTLP collector are never run or required by obsforge — they're only relevant to instrumentation tests you opt into.

---

Development

python3 -m venv .venv && source .venv/bin/activate   # Python 3.11+
pip install -e ".[dev,fastapi,django,db,http,otel]"

ruff check src tests          # lint
mypy src                      # strict type-check
pytest -m "not requires_postgres and not requires_mysql and not requires_kafka and not requires_rabbitmq and not requires_broker"

---

License

MIT — see LICENSE.