KrakenD API Gateway

KrakenD is an open-source API gateway — the front door that routes, aggregates and shapes requests to your backends. A gateway’s failure modes are its own: it can serve errors (5xx), serve slowly (latency), stop serving entirely (silence), or serve fine while a backend behind it fails or times out. This system type has a starter check for each.

Key: krakend · applied automatically to services emitting krakend.-prefixed metrics.

Getting the telemetry

No Collector receiver is needed: KrakenD instruments itself via krakend-otel and exports OTLP traces and metrics directly. Point its exporter at a Collector that attaches your Sluicio ingest key:

krakend.json (extra_config, minimal)

"telemetry/opentelemetry": {
  "service_name": "krakend-gateway",
  "exporters": {
    "otlp": [
      { "name": "collector", "host": "http://otel-collector", "port": 4318, "use_http": true }
    ]
  }
}

The full walkthrough — the Collector in between (Community Edition can’t attach auth headers itself), the plaintext-TLS pitfall, what the spans look like — is in Sending KrakenD telemetry to Sluicio.

What Sluicio detects

Metrics whose names start with krakend., emitted by krakend-otel alongside the gateway’s traces and the http.client.* backend metrics.

Starter checks

Check	Condition	Severity	Why it matters
Gateway 5xx responses	Error traces where the span attribute http.response.status_code ≥ 500	Warning	Users are getting errors. See the note below — this check needs a cell setting to work.
High p95 latency	p95 request latency > 2,000 ms	Warning	The gateway is slow — its own overhead or a slow backend; either way, callers feel it.
Gateway silent	No traces received for 15 minutes	Warning	A gateway that stops reporting is down, cut off from the Collector, or has lost its telemetry config.
Backend unreachable	http.client.request.failed.count increasing (delta)	Warning	The gateway can’t reach a backend — connection refused, DNS, TLS. Callers see errors or degraded aggregates.
Backend timeouts	http.client.request.timedout.count increasing (delta)	Warning	A backend answers, but slower than the configured timeout — the classic slow-dependency failure.

The 5xx check needs “Treat HTTP 5xx as errors”

KrakenD records a failed request only as a span attribute and leaves the span status at Ok/Unset, so an unconfigured cell counts those traces as healthy. Enable Settings → System → General → “Treat HTTP 5xx as errors” (admin-only, off by default) so 5xx spans are stored as errors — the check works from that point on. The KrakenD guide explains the mechanics.

Tuning notes

• p95 threshold. 2,000 ms is a conservative default for an aggregating gateway. An edge gateway fronting fast APIs should run far tighter — set it from your real latency profile with headroom above normal p95.
• Silence window. 15 minutes suits steady traffic. A gateway that’s quiet at night will false-positive — widen the window, or scope the check to business hours.
• Backend checks are gateway-wide deltas. Any failing backend trips them. Once you know your traffic, consider per-route alert rules on failed traces (attribute conditions on http.route) to pinpoint which backend — the KrakenD guide shows how.