kilden

The Python server-side SDK for Kilden, a customer data platform: analytics, campaigns and session replay on one event pipeline. Zero runtime dependencies, fully typed, fork-safe. Python 3.9+.

pip install kilden

from kilden import Client

kilden = Client("sk_your_secret_key")  # the SECRET key — see below
kilden.track("user_42", "order_completed", {"revenue": 99.9, "currency": "CLP"})
kilden.close()  # on shutdown; see "Batching and shutdown"

Events are queued in memory and delivered in batches by a background thread — track() costs nanoseconds and never blocks a request.

Use the secret write key, never the public one. Events signed with the secret key arrive as source=server, verified=true — facts your funnels and revenue reports can trust. The constructor rejects public (wk_…) keys. Keep the secret key out of browsers, mobile apps and frontend bundles; the public key belongs to kilden-sdk-js.

Identity verification

Anyone can open a browser console and send events as ceo@yourcorp.com. Kilden's answer is signed identity tokens: your backend — the only thing that knows who is logged in — signs a short-lived token the browser SDK attaches to its events. Without it, identified browser events stay verified=false and sensitive consumers ignore them.

This SDK makes the signature three lines:

from kilden import IdentitySigner

signer = IdentitySigner(os.environ["KILDEN_IDENTITY_SECRET"], kid="k1")

@app.route("/kilden/identity", methods=["POST"])
@login_required
def kilden_identity():
    token = signer.sign(current_user.id, traits={"plan": current_user.plan})
    return {"distinct_id": current_user.id, "token": token}

Only ever sign an id your backend authenticated. Signing something like request.json["user_id"] lets anyone impersonate anyone — with a "verified" stamp on top. sign() refuses TTLs over 7 days (default 1h); traits become signed traits that override unsigned ones during enrichment. The IdentitySigner is a separate class precisely so a page-rendering controller can sign tokens without touching the event queue.

Feature flags

if kilden.is_enabled("new_checkout", "user_42",
                     person_properties={"plan": "pro"},
                     default=False):
    ...

variant = kilden.get_feature_flag("experiment_button", "user_42")
# False | True | "variant_key"

Flags are evaluated remotely against /decide with a 30-second in-memory cache per distinct_id. person_properties overrides stored person traits for that evaluation only (and bypasses the cache). If Kilden cannot answer within the client timeout you get default back — flag checks never raise and never block beyond one request.

Batching and shutdown

track(), identify() and alias() enqueue; a daemon thread flushes every 10 seconds or 20 events (configurable). Two consequences you should know:

• Call close() when your process exits. It flushes with a 10-second deadline and stops the worker. Without it you rely on the atexit hook, which does not run when the process is SIGKILLed — events queued in the last seconds would be lost.
• The queue is bounded (max_queue_size, default 10 000). When full, new events are dropped and counted in client.dropped_count — the SDK never blocks your request thread and never grows without bound.

Delivery retries 429/5xx/network errors three times with exponential backoff and jitter, honoring Retry-After. Other 4xx responses are dropped immediately — retrying a bad request is spam.

Fork safety (gunicorn, celery, uwsgi)

Preforking servers import your app once and then fork. Without care, each worker inherits the parent's event queue (double-sends) and a dead worker thread (nothing sends at all). This SDK checks the process id on every enqueue: in a forked child it discards the inherited queue — those events belong to the parent — and starts a fresh worker thread. This is tested in CI against a real preforked gunicorn, not simulated. You do not need post-fork hooks; it just works.

Configuration

Client(
    "sk_…",
    host="https://ingest.kilden.io",
    flush_at=20,          # events that trigger a flush
    flush_interval=10,    # seconds between flushes
    max_queue_size=10000,
    timeout=3,            # per HTTP request, seconds
    transport=None,       # bring your own; must expose send(url, body, headers)
    debug=False,
    enabled=True,         # False = full no-op for tests/CI
)

track/identify accept timestamp= (ISO 8601 or datetime; for backfills) and uuid= (idempotency across your own retries). Invalid input after construction is dropped and logged, never raised — telemetry must not take down a request. The constructor is the one place that fails fast.

Spec

This SDK implements the Kilden Server SDK Specification (spec 0.1) and runs its frozen test vectors in CI, byte-for-byte for identity tokens. Behavior changes land in the spec first; divergence from it is a bug worth reporting.

Community

Questions and design conversations: Discussions. Product docs: docs.kilden.io.

License

MIT