Logging — the print() replacement worth learning

Every backend job you'll ever hold uses logging, not print. Reasons that add up:

• Levels — turn debug on locally, keep it off in prod.
• Configurable per-module — quiet a noisy library without touching call sites.
• Structured — JSON logs feed straight into Grafana / Datadog / Loki.
• Includes tracebacks — log.exception() captures the stack for free.

The five levels (memorize the order)

• DEBUG — verbose inspection while developing.
• INFO — normal events worth remembering (user signed up, payment ok).
• WARNING — retryable or degraded state (gateway timeout, retrying).
• ERROR — the operation failed, but the app still runs.
• CRITICAL — the app can't continue (DB down, kernel OOM).

Set the threshold once via basicConfig(level=...). In prod, most services run at INFO.

Lazy formatting — the one perf rule

log.debug("payload=%s", payload)      # RIGHT — string built only if DEBUG is on
log.debug(f"payload={payload}")       # WRONG — string built every call

Matters when you leave debug calls in hot paths.

Handlers = where logs go

Default: stderr. Real apps chain multiple handlers:

• StreamHandler → stdout (for Docker, Kubernetes to pick up).
• FileHandler → a file (with RotatingFileHandler for size caps).
• Third-party: Sentry, Datadog, CloudWatch — all plug in the same way.

Structured (JSON) logs

Modern log aggregators want machine-readable input. Instead of "charge ok user=alice amount=199" write {"event": "charge_ok", "user": "alice", "amount": 199}. You can grep, filter, aggregate by fields.

Libraries that do this for you: structlog (great), loguru (nice API, some pin against it), or roll your own JsonFormatter like the starter code shows.

Don't log secrets

Password, API keys, JWT tokens, PII — never. Add a redactor filter or an allowlist for the fields you log. This bites everyone once.

Try it

1. Add a second logger "gateway" and log two INFO events under it — notice getLogger("payments") vs getLogger("gateway") in output.
2. Wrap a divide by zero, log at ERROR with log.exception, then read the traceback in the output.