Expose delivery metrics on /metrics (closes #209)

/metrics carried only the inbound HTTP surface, so a destination
failing for an hour, a growing retry backlog and a stuck-open circuit
breaker were all invisible: the receive side stays healthy in each
case because it is.

New internal/metrics registers, on the existing default registry that
the go-http-metrics recorder and the promhttp handler already share:

- webhooker_events_received_total
- webhooker_delivery_attempts_total
- webhooker_deliveries_succeeded_total
- webhooker_deliveries_failed_total
- webhooker_delivery_retries_total
- webhooker_delivery_duration_seconds
- webhooker_deliveries_pending / _retrying
- webhooker_circuit_breakers_open

The route mounting is untouched.

Every delivery metric carries one label, target_type, whose domain is
the four target-type constants; anything outside it collapses to
"unknown" so no series can be minted from a UUID. Target ids, event
ids and entrypoint ids are deliberately not labels.

An attempt is counted, and its duration observed, only where one was
actually dispatched — the target's own result path, which is also
where the DeliveryResult is written. A delivery an open circuit
breaker refuses sends nothing and records no result row; counting it
would climb the attempts counter with no traffic behind it and pull
the duration quantiles down for as long as the breaker stayed open,
moving the metric the wrong way during the outage it exists to reveal.
The log and database targets now time their own work, so their result
rows carry a real duration too.

The outcome counters move after the status row is written rather than
before, so a transition the database rejected is never reported as an
outcome that happened.

The queue-depth gauges are counted out of the per-webhook databases by
a 30s sampler rather than tracked as deltas, which would need seeding
at startup and would drift on any transition that failed to persist.
They publish an "unknown" series from registration: deliveries queued
against a target that has since been deleted resolve to the empty type
and are folded there, because a backlog behind a deleted target is
precisely the one nobody is watching. The open-breaker gauge is
recounted from the target's breaker registry on every state change.

The orphaned-retry terminal path takes the target type as an argument
rather than attaching the loaded target to the delivery. That path
loads the delivery without its target relation on purpose: a populated
Delivery.Target makes GORM's SaveBeforeAssociations upsert the whole
target row on the status UPDATE, writing the plaintext target config —
the credential, for a slack target — into the per-webhook events
database. A test asserts that path leaves the targets table empty.

README.md

@@ -107,14 +107,26 @@ TTY detection, and security headers are always applied.
 | `DATA_DIR`              | Directory for all SQLite databases  | `/var/lib/webhooker` |
 | `DEBUG`                 | Enable debug logging                | `false`  |
 | `MAINTENANCE_MODE`      | Report `maintenanceMode: true` in the healthcheck JSON. It does not change how any request is served — no maintenance page exists | `false`  |
-| `METRICS_USERNAME`      | Basic auth username for `/metrics`  | `""`     |
-| `METRICS_PASSWORD`      | Basic auth password for `/metrics`  | `""`     |
+| `METRICS_USERNAME`      | Basic auth username for `/metrics`. Must be set together with `METRICS_PASSWORD`; one without the other fails startup | `""`     |
+| `METRICS_PASSWORD`      | Basic auth password for `/metrics`. Must be set together with `METRICS_USERNAME`; one without the other fails startup | `""`     |
 | `SENTRY_DSN`            | Sentry error reporting DSN          | `""`     |
 | `RETENTION_SWEEP_INTERVAL` | How often the retention reaper and archive sweeper run (Go duration, must be positive) | `1h` |
 | `SESSION_IDLE_TIMEOUT`  | Idle session timeout (Go duration)  | `24h`    |
 | `RECEIVER_RATE_LIMIT`   | Receiver requests/minute per IP per entrypoint (10x that per IP across the route) | `120` |
 | `TRUSTED_PROXIES`       | CIDRs whose forwarded headers are trusted (unset: all clients behind a proxy share one rate-limit bucket; a correct login password is never throttled either way) | `""` (none) |
 
+#### Metrics credentials
+
+`METRICS_USERNAME` and `METRICS_PASSWORD` are set together or not at
+all. With both set, `/metrics` is served behind basic auth. With
+neither set, the route is not registered and returns 404. With one set
+and the other empty or unset, the process refuses to start and exits
+non-zero with an error naming both variables — mounting the endpoint
+on the username alone would publish it behind a password that is the
+empty string, and quietly withholding it would deny an endpoint that
+was asked for. The `hasMetricsAuth` field in the startup log and the
+existence of the route are the same value, so they cannot disagree.
+
 #### Trusted proxies
 
 `TRUSTED_PROXIES` is a comma-separated list of CIDR blocks (a bare
@@ -1122,6 +1134,52 @@ delivery as `retrying` and schedules a retry timer for after the
 remaining cooldown period. This ensures no deliveries are lost — they're
 just delayed until the target is healthy again.
 
+### Metrics
+
+`/metrics` serves one Prometheus registry behind basic auth (see
+[Infrastructure Endpoints](#infrastructure-endpoints)). Alongside the
+inbound HTTP metrics recorded by the middleware, it exposes the
+delivery pipeline — the part of the service that can be failing while
+the receive side looks perfectly healthy, because it is: events are
+arriving and being stored, they are just not getting anywhere.
+
+| Metric | Type | Meaning |
+| ------ | ---- | ------- |
+| `webhooker_events_received_total` | counter | Events received and durably stored. Compare against the delivery counters on one dashboard |
+| `webhooker_delivery_attempts_total` | counter | Delivery attempts actually dispatched to a target. A delivery an open circuit breaker refused is not one: it is counted as a retry instead |
+| `webhooker_deliveries_succeeded_total` | counter | Deliveries that reached `delivered` |
+| `webhooker_deliveries_failed_total` | counter | Deliveries that failed terminally and will not be retried |
+| `webhooker_delivery_retries_total` | counter | Deliveries put back into `retrying` |
+| `webhooker_delivery_duration_seconds` | histogram | Wall time of a single dispatched delivery attempt, the same duration the attempt's `DeliveryResult` records |
+| `webhooker_deliveries_pending` | gauge | Deliveries currently in `pending` |
+| `webhooker_deliveries_retrying` | gauge | Deliveries currently in `retrying` |
+| `webhooker_circuit_breakers_open` | gauge | Circuit breakers currently open |
+
+Every delivery metric carries exactly one label, `target_type`, and
+cardinality is the whole reason for that restriction. A target type is
+one of four compile-time constants, so the label domain is bounded by
+construction; a value outside that set collapses to `unknown` rather
+than minting a series of its own. Target ids, event ids and entrypoint
+ids are deliberately not labels: they are UUIDs minted per operator
+action or per inbound request, a series is never reclaimed once it
+exists, and labelling by any of them would make `/metrics` a memory
+leak that grows with traffic.
+
+The two queue-depth gauges are counted out of the databases by a
+sampler that runs every 30 seconds for as long as the delivery engine
+does, rather than tracked as deltas alongside the status transitions: a
+delta would have to be seeded at startup from rows a previous process
+wrote, and would drift permanently on any transition that failed to
+persist.
+
+Those two gauges also publish an `unknown` series, from startup rather
+than on first occurrence. Deliveries queued against a target that has
+since been deleted are counted there: that backlog is the one nobody is
+watching, so it is the one that must not silently vanish from the
+gauge. The outcome counters move only after the status change has been
+written, so a transition the database rejected is never reported as an
+outcome that happened.
+
 ### Rate Limiting
 
 Global blanket rate limiting middleware (e.g., a per-IP throttle shared
@@ -1682,7 +1740,7 @@ abuse limit later; they are tracked as future work.
 
 | Method | Path       | Description |
 | ------ | ---------- | ----------- |
-| `GET`  | `/metrics` | Prometheus metrics, behind basic auth. The route is registered only when `METRICS_USERNAME` is set; otherwise it does not exist and returns 404 |
+| `GET`  | `/metrics` | Prometheus metrics, behind basic auth. The route is registered only when `METRICS_USERNAME` and `METRICS_PASSWORD` are both set; with neither set it does not exist and returns 404, and with only one set the process refuses to start |
 
 #### API (Planned)
 
@@ -1747,6 +1805,7 @@ webhooker/
 │   │   ├── target_log.go      # Log target (stdout)
 │   │   ├── target_config_view.go  # Masked target config for templates
 │   │   ├── archive_sweeper.go # Periodic pruning of idle archives
+│   │   ├── queue_depth.go     # Periodic sampler behind the queue-depth gauges
 │   │   ├── url_mask.go        # Strips credentials from *url.Error
 │   │   └── ssrf.go            # SSRF prevention (IP validation, safe HTTP transport)
 │   ├── handlers/
@@ -1764,6 +1823,8 @@ webhooker/
 │   │   └── lifecycle.go       # Shared stop-hook waiter, bounded by the stop context
 │   ├── logger/
 │   │   └── logger.go          # slog setup with TTY detection
+│   ├── metrics/
+│   │   └── metrics.go         # Delivery Prometheus collectors, labelled by target type
 │   ├── middleware/
 │   │   ├── middleware.go      # Logging, CORS, Auth, Metrics, MetricsAuth, SecurityHeaders, MaxBodySize
 │   │   ├── csrf.go            # CSRF protection middleware (gorilla/csrf)
@@ -1837,7 +1898,8 @@ Applied to all routes in this order:
    Permissions-Policy)
 3. **Logging** — Structured request logging (method, URL, status,
    latency, remote IP, user agent, request ID)
-4. **Metrics** — Prometheus HTTP metrics (if `METRICS_USERNAME` is set)
+4. **Metrics** — Prometheus HTTP metrics (if `METRICS_USERNAME` and
+   `METRICS_PASSWORD` are both set)
 5. **CORS** — Cross-origin resource sharing headers
 6. **Timeout** — 60-second request timeout
 7. **Recoverer** — Panic recovery: one `ERROR` record through
@@ -1869,8 +1931,9 @@ being read and without reaching CSRF, the route group's remaining
 middleware, or the handler. It is not rejected before *any* other
 middleware, though: the global entries listed above all run first, so
 such a request is still logged and given the security headers — and
-counted in the metrics, on a deployment where `METRICS_USERNAME` is
-set and the Metrics middleware is therefore registered at all. The
+counted in the metrics, on a deployment where the `/metrics`
+credentials are set and the Metrics middleware is therefore registered
+at all. The
 rejection itself is logged at `WARN` with the method, path and
 declared length. A chunked request, or
 one that lies about its length, is hard-capped by

go.mod

@@ -13,6 +13,7 @@ require (
 	github.com/gorilla/sessions v1.4.0
 	github.com/joho/godotenv v1.5.1
 	github.com/prometheus/client_golang v1.18.0
+	github.com/prometheus/client_model v0.5.0
 	github.com/slok/go-http-metrics v0.11.0
 	github.com/stretchr/testify v1.8.4
 	go.uber.org/fx v1.20.1
@@ -38,7 +39,6 @@ require (
 	github.com/mattn/go-sqlite3 v1.14.17 // indirect
 	github.com/matttproud/golang_protobuf_extensions/v2 v2.0.0 // indirect
 	github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2 // indirect
-	github.com/prometheus/client_model v0.5.0 // indirect
 	github.com/prometheus/common v0.45.0 // indirect
 	github.com/prometheus/procfs v0.12.0 // indirect
 	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect

internal/config/config.go

@@ -71,6 +71,16 @@ var ErrInvalidPort = errors.New("invalid port")
 // nor a bare IP address.
 var ErrInvalidCIDR = errors.New("invalid CIDR")
 
+// ErrIncompleteMetricsAuth is returned when exactly one of
+// METRICS_USERNAME and METRICS_PASSWORD carries a value. Neither
+// fallback is acceptable: serving /metrics on the username alone
+// publishes an endpoint whose password is the empty string, and
+// silently leaving it unmounted withholds an endpoint the operator
+// asked for. Half-set is a configuration error, so startup fails.
+var ErrIncompleteMetricsAuth = errors.New(
+	"incomplete metrics credentials",
+)
+
 //nolint:revive // ConfigParams is a standard fx naming convention.
 type ConfigParams struct {
 	fx.In
@@ -128,6 +138,21 @@ func (c *Config) IsProd() bool {
 	return c.Environment == EnvironmentProd
 }
 
+// MetricsAuthEnabled reports whether /metrics is served behind basic
+// auth. It is the only answer to that question in the codebase: the
+// route mount, the Prometheus recording middleware and the startup
+// log's hasMetricsAuth field all read this one method, so the log
+// cannot report auth as off while the route is mounted.
+//
+// It requires both credentials rather than the username alone.
+// loadFromEnv already rejects a half-set pair, but a Config built in
+// code bypasses that, and the failure mode this guards is an endpoint
+// mounted with a credential map whose only password is the empty
+// string.
+func (c *Config) MetricsAuthEnabled() bool {
+	return c.MetricsUsername != "" && c.MetricsPassword != ""
+}
+
 // envString returns the value of the named environment variable,
 // or an empty string if not set.
 func envString(key string) string {
@@ -329,6 +354,30 @@ func envPrefixList(key string) ([]netip.Prefix, error) {
 	return prefixes, nil
 }
 
+// resolveMetricsAuth reads the /metrics basic-auth credentials and
+// rejects a half-set pair, naming both variables either way. The
+// error carries neither value: the password is a secret.
+func resolveMetricsAuth() (string, string, error) {
+	username := envString("METRICS_USERNAME")
+	password := envString("METRICS_PASSWORD")
+
+	if (username == "") == (password == "") {
+		return username, password, nil
+	}
+
+	set, empty := "METRICS_USERNAME", "METRICS_PASSWORD"
+	if username == "" {
+		set, empty = empty, set
+	}
+
+	return "", "", fmt.Errorf(
+		"%w: %s is set but %s is empty; METRICS_USERNAME and "+
+			"METRICS_PASSWORD must both be set to serve /metrics, "+
+			"or both be empty to leave it unmounted",
+		ErrIncompleteMetricsAuth, set, empty,
+	)
+}
+
 // resolveEnvironment reads WEBHOOKER_ENVIRONMENT, defaulting to
 // dev, and rejects unrecognised values.
 func resolveEnvironment() (string, error) {
@@ -406,13 +455,18 @@ func loadFromEnv() (*Config, error) {
 		return nil, err
 	}
 
+	metricsUsername, metricsPassword, err := resolveMetricsAuth()
+	if err != nil {
+		return nil, err
+	}
+
 	return &Config{
 		DataDir:                envString("DATA_DIR"),
 		Debug:                  debug,
 		MaintenanceMode:        maintenanceMode,
 		Environment:            environment,
-		MetricsUsername:        envString("METRICS_USERNAME"),
-		MetricsPassword:        envString("METRICS_PASSWORD"),
+		MetricsUsername:        metricsUsername,
+		MetricsPassword:        metricsPassword,
 		Port:                   port,
 		SentryDSN:              envString("SENTRY_DSN"),
 		RetentionSweepInterval: retentionSweepInterval,
@@ -512,8 +566,7 @@ func New(lc fx.Lifecycle, params ConfigParams) (*Config, error) {
 		"receiverRateLimit", s.ReceiverRateLimit,
 		"trustedProxies", len(s.TrustedProxies),
 		"hasSentryDSN", s.SentryDSN != "",
-		"hasMetricsAuth",
-		s.MetricsUsername != "" && s.MetricsPassword != "",
+		"hasMetricsAuth", s.MetricsAuthEnabled(),
 	)
 
 	s.warnSharedRateLimitBucket(log)

internal/config/config_test.go

@@ -26,6 +26,12 @@ const (
 	// cidrPrivateV4 is the sample trusted-proxy block the
 	// TRUSTED_PROXIES cases are built from.
 	cidrPrivateV4 = "10.0.0.0/8"
+
+	// metricsAuthValue is the sample METRICS_PASSWORD the metrics
+	// credential cases are built from. It is asserted absent from
+	// the startup error, so it must not be a substring of either
+	// variable name that error prints.
+	metricsAuthValue = "s3cret"
 )
 
 func TestEnvironmentConfig(t *testing.T) {
@@ -726,3 +732,168 @@ func TestSharedRateLimitBucketWarning(t *testing.T) {
 		})
 	}
 }
+
+// metricsEnv describes what one subtest below puts in the
+// environment for a single METRICS_ variable. A variable that is
+// set to the empty string and one that is not set at all are
+// distinct inputs here, because the reported bug arrived through
+// the first of them.
+type metricsEnv struct {
+	set   bool
+	value string
+}
+
+// unset leaves the variable out of the environment entirely.
+func unset() metricsEnv {
+	return metricsEnv{set: false, value: ""}
+}
+
+// setTo sets the variable, including to the empty string.
+func setTo(value string) metricsEnv {
+	return metricsEnv{set: true, value: value}
+}
+
+// metricsAuthCase is one row of the table in TestMetricsAuthConfig,
+// named so the table can live in its own function and keep the test
+// itself short.
+type metricsAuthCase struct {
+	name        string
+	username    metricsEnv
+	password    metricsEnv
+	expectError bool
+	expectAuth  bool
+}
+
+// metricsAuthCases enumerates every combination of the two
+// credentials, counting "set to the empty string" and "not set at
+// all" as separate inputs on each side.
+func metricsAuthCases() []metricsAuthCase {
+	return []metricsAuthCase{
+		{
+			name:     "both unset leaves metrics unmounted",
+			username: unset(),
+			password: unset(),
+		},
+		{
+			name:     "both empty leaves metrics unmounted",
+			username: setTo(""),
+			password: setTo(""),
+		},
+		{
+			name:       "both set enables metrics auth",
+			username:   setTo("metrics"),
+			password:   setTo(metricsAuthValue),
+			expectAuth: true,
+		},
+		{
+			name:        "username with unset password fails",
+			username:    setTo("metrics"),
+			password:    unset(),
+			expectError: true,
+		},
+		{
+			name:        "username with empty password fails",
+			username:    setTo("metrics"),
+			password:    setTo(""),
+			expectError: true,
+		},
+		{
+			name:        "password with unset username fails",
+			username:    unset(),
+			password:    setTo(metricsAuthValue),
+			expectError: true,
+		},
+		{
+			name:        "password with empty username fails",
+			username:    setTo(""),
+			password:    setTo(metricsAuthValue),
+			expectError: true,
+		},
+	}
+}
+
+// TestMetricsAuthConfig covers every combination of METRICS_USERNAME
+// and METRICS_PASSWORD. Either both carry a value, in which case
+// /metrics is served behind basic auth, or neither does, in which
+// case the route is never mounted. One without the other is a
+// startup error rather than a fallback: mounting on the username
+// alone published /metrics behind a credential map that accepted an
+// empty password, which is the defect this test exists to pin. See
+// https://git.eeqj.de/sneak/webhooker/issues/205.
+func TestMetricsAuthConfig(t *testing.T) {
+	for _, tt := range metricsAuthCases() {
+		t.Run(tt.name, func(t *testing.T) {
+			// Cannot use t.Parallel() here because t.Setenv
+			// is incompatible with parallel subtests.
+			if tt.username.set {
+				t.Setenv("METRICS_USERNAME", tt.username.value)
+			} else {
+				require.NoError(
+					t, os.Unsetenv("METRICS_USERNAME"),
+				)
+			}
+
+			if tt.password.set {
+				t.Setenv("METRICS_PASSWORD", tt.password.value)
+			} else {
+				require.NoError(
+					t, os.Unsetenv("METRICS_PASSWORD"),
+				)
+			}
+
+			if tt.expectError {
+				assertMetricsAuthRejected(t)
+
+				return
+			}
+
+			assertMetricsAuthAccepted(t, tt.expectAuth)
+		})
+	}
+}
+
+// assertMetricsAuthRejected requires that fx refused to build the
+// graph, that the failure is ErrIncompleteMetricsAuth, and that the
+// operator is told both variable names — the point of failing here
+// rather than degrading is that the message says what to fix.
+func assertMetricsAuthRejected(t *testing.T) {
+	t.Helper()
+
+	var cfg *config.Config
+
+	app := fx.New(
+		fx.NopLogger,
+		fx.Provide(globals.New, logger.New, config.New),
+		fx.Populate(&cfg),
+	)
+
+	err := app.Err()
+	require.Error(t, err)
+	require.ErrorIs(t, err, config.ErrIncompleteMetricsAuth)
+	assert.Contains(t, err.Error(), "METRICS_USERNAME")
+	assert.Contains(t, err.Error(), "METRICS_PASSWORD")
+	// The password is a secret and must not reach a startup error.
+	assert.NotContains(t, err.Error(), metricsAuthValue)
+}
+
+// assertMetricsAuthAccepted requires that startup succeeded and that
+// MetricsAuthEnabled — the single value the /metrics mount and the
+// startup log both read — reports what the environment asked for.
+func assertMetricsAuthAccepted(t *testing.T, expectAuth bool) {
+	t.Helper()
+
+	var cfg *config.Config
+
+	app := fxtest.New(
+		t,
+		fx.Provide(globals.New, logger.New, config.New),
+		fx.Populate(&cfg),
+	)
+	require.NoError(t, app.Err())
+
+	app.RequireStart()
+
+	defer app.RequireStop()
+
+	assert.Equal(t, expectAuth, cfg.MetricsAuthEnabled())
+}

internal/delivery/engine.go

@@ -15,6 +15,7 @@ import (
 	"sneak.berlin/go/webhooker/internal/database"
 	"sneak.berlin/go/webhooker/internal/lifecycle"
 	"sneak.berlin/go/webhooker/internal/logger"
+	"sneak.berlin/go/webhooker/internal/metrics"
 )
 
 const (
@@ -139,6 +140,12 @@ type Engine struct {
 	retryCh    chan Task
 	workers    int
 
+	// mtr is the delivery metric set. Production wires the
+	// process-wide one; a test can substitute a set registered on
+	// a private registry so its assertions are not disturbed by
+	// deliveries other tests are making at the same time.
+	mtr *metrics.Set
+
 	// targets maps each target type to its implementation.
 	targets map[database.TargetType]Target
 
@@ -164,6 +171,7 @@ func New(
 		deliveryCh: make(chan Task, deliveryChannelSize),
 		retryCh:    make(chan Task, retryChannelSize),
 		workers:    defaultWorkers,
+		mtr:        metrics.Default(),
 	}
 
 	e.initTargets(&http.Client{
@@ -283,6 +291,10 @@ func (e *Engine) start() {
 
 	go e.retrySweep(ctx)
 
+	e.wg.Add(1)
+
+	go e.queueDepthSampler(ctx)
+
 	e.log.Info(
 		"delivery engine started",
 		"workers", e.workers,
@@ -837,8 +849,15 @@ func (e *Engine) failUnretryableRetry(
 		0,
 	)
 
+	// The type is passed rather than assigned onto d: the delivery
+	// is loaded here without its target relation, and populating
+	// d.Target would make GORM's SaveBeforeAssociations upsert the
+	// whole target row — plaintext config, which for a slack target
+	// is the credential — into the per-webhook event database. See
+	// https://git.eeqj.de/sneak/webhooker/issues/206.
 	e.updateDeliveryStatus(
-		webhookDB, d, database.DeliveryStatusFailed,
+		webhookDB, d, target.Type,
+		database.DeliveryStatusFailed,
 	)
 }
 
@@ -859,7 +878,8 @@ func (e *Engine) processDelivery(
 		)
 
 		e.updateDeliveryStatus(
-			webhookDB, d, database.DeliveryStatusFailed,
+			webhookDB, d, d.Target.Type,
+			database.DeliveryStatusFailed,
 		)
 
 		return
@@ -868,6 +888,24 @@ func (e *Engine) processDelivery(
 	target.Deliver(ctx, webhookDB, d, task, e)
 }
 
+// observeAttempt counts one delivery attempt that was actually
+// dispatched to a target, and records how long it took.
+//
+// It is called from the dispatch paths rather than from around
+// Target.Deliver, because Deliver is also entered for deliveries
+// that never reach the wire: a delivery an open circuit breaker
+// refuses sends nothing, records no DeliveryResult, and is
+// rescheduled. Counting those would climb the attempts counter with
+// no traffic behind it and fill the duration histogram with
+// microsecond samples, which would make the delivery-duration
+// quantiles improve during exactly the outage they exist to reveal.
+func (e *Engine) observeAttempt(
+	t database.TargetType, dur time.Duration,
+) {
+	e.mtr.DeliveryAttempted(t)
+	e.mtr.ObserveDeliveryDuration(t, dur)
+}
+
 // recordResult persists a DeliveryResult row describing a
 // single attempt. It is a cross-target helper the targets
 // call.
@@ -901,10 +939,22 @@ func (e *Engine) recordResult(
 }
 
 // updateDeliveryStatus persists a new status for a delivery.
-// It is a cross-target helper the targets call.
+// It is a cross-target helper the targets call, and therefore the
+// single point where a delivery's outcome — delivered, terminally
+// failed, or put back into retry — is counted.
+//
+// The target type is a parameter rather than read off d.Target
+// because one caller — failUnretryableRetry — deliberately holds a
+// delivery loaded without its target relation, and must keep it that
+// way: a populated d.Target makes GORM upsert the target row, config
+// and all, into the per-webhook database.
+//
+// The counter moves only after the row is written, so a transition
+// the database rejected is not claimed as an outcome that happened.
 func (e *Engine) updateDeliveryStatus(
 	webhookDB *gorm.DB,
 	d *database.Delivery,
+	targetType database.TargetType,
 	status database.DeliveryStatus,
 ) {
 	err := webhookDB.Model(d).
@@ -916,7 +966,11 @@ func (e *Engine) updateDeliveryStatus(
 			"status", status,
 			"error", err,
 		)
+
+		return
 	}
+
+	e.mtr.DeliveryStatusChanged(targetType, status)
 }
 
 func truncate(s string, maxLen int) string {

internal/delivery/engine_integration_test.go

@@ -886,6 +886,82 @@ func TestSweepSingleRetry_TypeNoLongerRetries(
 	)
 }
 
+// TestFailUnretryableRetry_WritesNoTargetRow proves the
+// orphaned-retry terminal path leaves no target row — and so no
+// plaintext target config — in the per-webhook event database.
+//
+// That path loads the delivery without its Target relation on
+// purpose. Populating d.Target makes GORM's SaveBeforeAssociations
+// upsert the whole target row on the status UPDATE, which for a slack
+// target writes the incoming-webhook credential into events-*.db.
+// See https://git.eeqj.de/sneak/webhooker/issues/206.
+func TestFailUnretryableRetry_WritesNoTargetRow(
+	t *testing.T,
+) {
+	t.Parallel()
+
+	s := newISetup(t)
+
+	iCreateWebhook(
+		t, s.MainDB, s.WebhookID, "no-target-row",
+	)
+
+	targetID := uuid.New().String()
+
+	// A Slack incoming-webhook URL: the target config IS the
+	// credential, which is what makes a leaked target row a
+	// disclosure rather than a curiosity.
+	hookURL := "https://hooks.slack.com/services/T00/B00/x"
+
+	iCreateTarget(t, s.MainDB, targetID,
+		s.WebhookID, "credential-bearing",
+		database.TargetTypeLog, iHTTPConfig(hookURL), 5,
+	)
+
+	event := iSeedEvent(
+		t, s.WebhookDB, s.WebhookID, `{"orphaned":"retry"}`,
+	)
+
+	d := iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusRetrying,
+	)
+
+	iSeedFailedResult(t, s.WebhookDB, d.ID)
+
+	s.Engine.ExportSweepWebhookRetries(
+		context.Background(), s.WebhookID,
+	)
+
+	iAssertStatus(t, s.WebhookDB, d.ID,
+		database.DeliveryStatusFailed,
+	)
+
+	// The table exists in the per-webhook database because GORM
+	// migrates the Delivery relation's model alongside it. It must
+	// stay empty.
+	var targetRows int64
+
+	require.NoError(t, s.WebhookDB.
+		Table("targets").
+		Count(&targetRows).Error)
+
+	assert.Zero(t, targetRows,
+		"orphaned-retry terminal failure wrote a target row "+
+			"into the per-webhook event database",
+	)
+
+	var configs []string
+
+	require.NoError(t, s.WebhookDB.
+		Table("targets").
+		Pluck("config", &configs).Error)
+
+	assert.NotContains(
+		t, strings.Join(configs, " "), hookURL,
+	)
+}
+
 func TestRecoverSingleRetry_UnknownTargetType(
 	t *testing.T,
 ) {

internal/delivery/export_test.go

@@ -10,6 +10,7 @@ import (
 	"go.uber.org/fx"
 	"gorm.io/gorm"
 	"sneak.berlin/go/webhooker/internal/database"
+	"sneak.berlin/go/webhooker/internal/metrics"
 )
 
 // ErrExportArchiveWriterEvicted exposes the sentinel returned by
@@ -253,6 +254,7 @@ func NewTestEngine(
 		deliveryCh: make(chan Task, deliveryChannelSize),
 		retryCh:    make(chan Task, retryChannelSize),
 		workers:    workers,
+		mtr:        metrics.Default(),
 	}
 	e.initTargets(client)
 
@@ -267,6 +269,7 @@ func NewTestEngineSmallRetry(
 	e := &Engine{
 		log:     log,
 		retryCh: make(chan Task, 1),
+		mtr:     metrics.Default(),
 	}
 	e.initTargets(nil)
 
@@ -289,12 +292,25 @@ func NewTestEngineWithDB(
 		deliveryCh: make(chan Task, deliveryChannelSize),
 		retryCh:    make(chan Task, retryChannelSize),
 		workers:    workers,
+		mtr:        metrics.Default(),
 	}
 	e.initTargets(client)
 
 	return e
 }
 
+// ExportSetMetrics substitutes the engine's metric set, so a test can
+// assert on collectors registered on a private registry instead of
+// the process-wide ones every other test is also moving.
+func (e *Engine) ExportSetMetrics(mtr *metrics.Set) {
+	e.mtr = mtr
+}
+
+// ExportSampleQueueDepths runs one queue depth sample synchronously.
+func (e *Engine) ExportSampleQueueDepths(ctx context.Context) {
+	e.sampleQueueDepths(ctx)
+}
+
 // NewTestCircuitBreaker creates a CircuitBreaker with
 // custom settings for testing.
 func NewTestCircuitBreaker(

internal/delivery/metrics_test.go

@@ -0,0 +1,545 @@
+package delivery_test
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/google/uuid"
+	"github.com/prometheus/client_golang/prometheus"
+	dto "github.com/prometheus/client_model/go"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/webhooker/internal/database"
+	"sneak.berlin/go/webhooker/internal/delivery"
+	"sneak.berlin/go/webhooker/internal/metrics"
+)
+
+// Metric names as exposed on /metrics.
+const (
+	mAttempts  = "webhooker_delivery_attempts_total"
+	mSucceeded = "webhooker_deliveries_succeeded_total"
+	mFailed    = "webhooker_deliveries_failed_total"
+	mRetries   = "webhooker_delivery_retries_total"
+	mDuration  = "webhooker_delivery_duration_seconds"
+	mPending   = "webhooker_deliveries_pending"
+	mRetrying  = "webhooker_deliveries_retrying"
+	mBreakers  = "webhooker_circuit_breakers_open"
+)
+
+const (
+	mTypeHTTP    = "http"
+	mTypeLog     = "log"
+	mTypeUnknown = "unknown"
+)
+
+// mIsolate gives the setup's engine a metric set registered on a
+// private registry. The process-wide collectors are moved by every
+// other delivery test running in parallel, so exact assertions are
+// only possible against a registry this test owns.
+func mIsolate(
+	t *testing.T, s iSetup,
+) *prometheus.Registry {
+	t.Helper()
+
+	reg := prometheus.NewRegistry()
+	s.Engine.ExportSetMetrics(metrics.New(reg))
+
+	return reg
+}
+
+// mFind returns the series of the named metric carrying the given
+// target_type label.
+func mFind(
+	t *testing.T,
+	reg *prometheus.Registry,
+	name, targetType string,
+) *dto.Metric {
+	t.Helper()
+
+	families, err := reg.Gather()
+	require.NoError(t, err)
+
+	for _, fam := range families {
+		if fam.GetName() != name {
+			continue
+		}
+
+		for _, m := range fam.GetMetric() {
+			for _, label := range m.GetLabel() {
+				if label.GetName() == "target_type" &&
+					label.GetValue() == targetType {
+					return m
+				}
+			}
+		}
+	}
+
+	t.Fatalf(
+		"metric %s{target_type=%q} not found",
+		name, targetType,
+	)
+
+	return nil
+}
+
+func mCounter(
+	t *testing.T,
+	reg *prometheus.Registry,
+	name, targetType string,
+) float64 {
+	t.Helper()
+
+	return mFind(t, reg, name, targetType).
+		GetCounter().GetValue()
+}
+
+func mGauge(
+	t *testing.T,
+	reg *prometheus.Registry,
+	name, targetType string,
+) float64 {
+	t.Helper()
+
+	return mFind(t, reg, name, targetType).
+		GetGauge().GetValue()
+}
+
+// mHTTPDurations returns how many samples the delivery duration
+// histogram holds for the http target type, which is the type every
+// test here times.
+func mHTTPDurations(
+	t *testing.T, reg *prometheus.Registry,
+) uint64 {
+	t.Helper()
+
+	return mFind(t, reg, mDuration, mTypeHTTP).
+		GetHistogram().GetSampleCount()
+}
+
+// TestDeliveryMetrics_SuccessAndRetryExhaustion drives one delivery
+// that succeeds and one that fails every attempt until its retries
+// are exhausted, and asserts every delivery counter across both.
+func TestDeliveryMetrics_SuccessAndRetryExhaustion(
+	t *testing.T,
+) {
+	t.Parallel()
+
+	s := newISetup(t)
+	reg := mIsolate(t, s)
+
+	mDeliverOK(t, s)
+
+	assert.InDelta(t, 1.0,
+		mCounter(t, reg, mAttempts, mTypeHTTP), 0)
+	assert.InDelta(t, 1.0,
+		mCounter(t, reg, mSucceeded, mTypeHTTP), 0)
+	assert.InDelta(t, 0.0,
+		mCounter(t, reg, mFailed, mTypeHTTP), 0)
+	assert.InDelta(t, 0.0,
+		mCounter(t, reg, mRetries, mTypeHTTP), 0)
+	assert.Equal(t, uint64(1),
+		mHTTPDurations(t, reg))
+
+	mExhaustRetries(t, s)
+
+	// Two further attempts: the first is retried, the second is
+	// the last one allowed and fails the delivery terminally.
+	assert.InDelta(t, 3.0,
+		mCounter(t, reg, mAttempts, mTypeHTTP), 0)
+	assert.InDelta(t, 1.0,
+		mCounter(t, reg, mSucceeded, mTypeHTTP), 0)
+	assert.InDelta(t, 1.0,
+		mCounter(t, reg, mRetries, mTypeHTTP), 0)
+	assert.InDelta(t, 1.0,
+		mCounter(t, reg, mFailed, mTypeHTTP), 0)
+	assert.Equal(t, uint64(3),
+		mHTTPDurations(t, reg))
+
+	// Two consecutive failures are below the trip threshold.
+	assert.InDelta(t, 0.0,
+		mGauge(t, reg, mBreakers, mTypeHTTP), 0)
+
+	// The label is the target type and nothing finer: two http
+	// targets shared one series, and no other type's moved.
+	assert.InDelta(t, 0.0,
+		mCounter(t, reg, mAttempts, mTypeLog), 0)
+	assert.InDelta(t, 0.0,
+		mCounter(t, reg, mFailed, mTypeLog), 0)
+}
+
+// mDeliverOK delivers one event to a target that answers 200.
+func mDeliverOK(t *testing.T, s iSetup) {
+	t.Helper()
+
+	ts := httptest.NewServer(http.HandlerFunc(
+		func(w http.ResponseWriter, _ *http.Request) {
+			w.WriteHeader(http.StatusOK)
+		},
+	))
+	defer ts.Close()
+
+	event := iSeedEvent(
+		t, s.WebhookDB, s.WebhookID, `{"ok":true}`,
+	)
+	targetID := uuid.New().String()
+
+	d := iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusPending,
+	)
+
+	body := event.Body
+	task := iTask(
+		d, event, s.WebhookID, targetID,
+		"metrics-ok", iHTTPConfig(ts.URL), 3, 1, &body,
+	)
+
+	s.Engine.ExportProcessNewTask(context.TODO(), &task)
+
+	iAssertStatus(t, s.WebhookDB, d.ID,
+		database.DeliveryStatusDelivered,
+	)
+}
+
+// mExhaustRetries delivers to a target that answers 500 with a
+// two-attempt budget, driving both attempts so the delivery ends
+// terminally failed.
+func mExhaustRetries(t *testing.T, s iSetup) {
+	t.Helper()
+
+	ts := httptest.NewServer(http.HandlerFunc(
+		func(w http.ResponseWriter, _ *http.Request) {
+			w.WriteHeader(http.StatusInternalServerError)
+		},
+	))
+	defer ts.Close()
+
+	event := iSeedEvent(
+		t, s.WebhookDB, s.WebhookID, `{"ok":false}`,
+	)
+	targetID := uuid.New().String()
+
+	d := iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusPending,
+	)
+
+	body := event.Body
+	cfg := iHTTPConfig(ts.URL)
+
+	first := iTask(
+		d, event, s.WebhookID, targetID,
+		"metrics-fail", cfg, 2, 1, &body,
+	)
+
+	s.Engine.ExportProcessNewTask(context.TODO(), &first)
+
+	iAssertStatus(t, s.WebhookDB, d.ID,
+		database.DeliveryStatusRetrying,
+	)
+
+	// The engine's own scheduler would re-enqueue this after the
+	// backoff; driving the second attempt directly keeps the test
+	// deterministic and off the wall clock.
+	second := iTask(
+		d, event, s.WebhookID, targetID,
+		"metrics-fail", cfg, 2, 2, &body,
+	)
+
+	s.Engine.ExportProcessRetryTask(
+		context.TODO(), &second,
+	)
+
+	iAssertStatus(t, s.WebhookDB, d.ID,
+		database.DeliveryStatusFailed,
+	)
+}
+
+// TestDeliveryMetrics_CircuitBreakerGauge proves the open-breaker
+// gauge follows a breaker that trips.
+func TestDeliveryMetrics_CircuitBreakerGauge(t *testing.T) {
+	t.Parallel()
+
+	s := newISetup(t)
+	reg := mIsolate(t, s)
+
+	ts := httptest.NewServer(http.HandlerFunc(
+		func(w http.ResponseWriter, _ *http.Request) {
+			w.WriteHeader(http.StatusInternalServerError)
+		},
+	))
+	defer ts.Close()
+
+	event := iSeedEvent(
+		t, s.WebhookDB, s.WebhookID, `{"trip":true}`,
+	)
+	targetID := uuid.New().String()
+
+	d := iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusPending,
+	)
+
+	body := event.Body
+	cfg := iHTTPConfig(ts.URL)
+
+	// A retry budget above the failure threshold, so the breaker
+	// rather than the budget is what stops the delivery.
+	maxRetries := delivery.ExportDefaultFailureThreshold + 5
+
+	first := iTask(
+		d, event, s.WebhookID, targetID,
+		"metrics-trip", cfg, maxRetries, 1, &body,
+	)
+
+	s.Engine.ExportProcessNewTask(context.TODO(), &first)
+
+	assert.InDelta(t, 0.0,
+		mGauge(t, reg, mBreakers, mTypeHTTP), 0)
+
+	for attempt := 2; attempt <= delivery.
+		ExportDefaultFailureThreshold; attempt++ {
+		task := iTask(
+			d, event, s.WebhookID, targetID,
+			"metrics-trip", cfg, maxRetries, attempt, &body,
+		)
+
+		s.Engine.ExportProcessRetryTask(
+			context.TODO(), &task,
+		)
+	}
+
+	assert.InDelta(t, 1.0,
+		mGauge(t, reg, mBreakers, mTypeHTTP), 0)
+}
+
+// TestDeliveryMetrics_BreakerBlockedIsNotAnAttempt proves a delivery
+// an open circuit breaker refuses is neither counted as an attempt
+// nor observed in the duration histogram.
+//
+// It sends nothing and records no result row, so counting it would
+// climb the attempts counter with no traffic behind it and pull the
+// duration quantiles down with near-zero samples for as long as the
+// breaker stayed open — the metric moving the wrong way during the
+// outage it exists to reveal.
+func TestDeliveryMetrics_BreakerBlockedIsNotAnAttempt(
+	t *testing.T,
+) {
+	t.Parallel()
+
+	s := newISetup(t)
+	reg := mIsolate(t, s)
+
+	ts := httptest.NewServer(http.HandlerFunc(
+		func(w http.ResponseWriter, _ *http.Request) {
+			w.WriteHeader(http.StatusInternalServerError)
+		},
+	))
+	defer ts.Close()
+
+	event := iSeedEvent(
+		t, s.WebhookDB, s.WebhookID, `{"blocked":true}`,
+	)
+	targetID := uuid.New().String()
+
+	d := iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusPending,
+	)
+
+	body := event.Body
+	cfg := iHTTPConfig(ts.URL)
+	maxRetries := delivery.ExportDefaultFailureThreshold + 5
+
+	first := iTask(
+		d, event, s.WebhookID, targetID,
+		"metrics-blocked", cfg, maxRetries, 1, &body,
+	)
+
+	s.Engine.ExportProcessNewTask(context.TODO(), &first)
+
+	for attempt := 2; attempt <= delivery.
+		ExportDefaultFailureThreshold; attempt++ {
+		task := iTask(
+			d, event, s.WebhookID, targetID,
+			"metrics-blocked", cfg, maxRetries, attempt, &body,
+		)
+
+		s.Engine.ExportProcessRetryTask(context.TODO(), &task)
+	}
+
+	require.InDelta(t, 1.0,
+		mGauge(t, reg, mBreakers, mTypeHTTP), 0,
+		"breaker should be open before the blocked attempt")
+
+	threshold := float64(
+		delivery.ExportDefaultFailureThreshold,
+	)
+
+	assert.InDelta(t, threshold,
+		mCounter(t, reg, mAttempts, mTypeHTTP), 0)
+	assert.Equal(t, uint64(threshold),
+		mHTTPDurations(t, reg))
+
+	retriesBefore := mCounter(t, reg, mRetries, mTypeHTTP)
+
+	blocked := iTask(
+		d, event, s.WebhookID, targetID,
+		"metrics-blocked", cfg, maxRetries,
+		delivery.ExportDefaultFailureThreshold+1, &body,
+	)
+
+	s.Engine.ExportProcessRetryTask(context.TODO(), &blocked)
+
+	// The breaker refused it: rescheduled, so the retry counter
+	// moved, but nothing was attempted or timed.
+	assert.InDelta(t, retriesBefore+1,
+		mCounter(t, reg, mRetries, mTypeHTTP), 0)
+	assert.InDelta(t, threshold,
+		mCounter(t, reg, mAttempts, mTypeHTTP), 0)
+	assert.Equal(t, uint64(threshold),
+		mHTTPDurations(t, reg))
+}
+
+// TestDeliveryMetrics_OrphanedRetryFailureLabelled proves the
+// terminal failure of a delivery whose target no longer retries is
+// counted against the target's real type, not against unknown. The
+// type is threaded in as an argument because populating d.Target on
+// that path would write the target row into the per-webhook database
+// (https://git.eeqj.de/sneak/webhooker/issues/206).
+func TestDeliveryMetrics_OrphanedRetryFailureLabelled(
+	t *testing.T,
+) {
+	t.Parallel()
+
+	s := newISetup(t)
+	reg := mIsolate(t, s)
+
+	iCreateWebhook(
+		t, s.MainDB, s.WebhookID, "orphaned-label",
+	)
+
+	deliveryID := iSeedRetryingWithType(
+		t, s, database.TargetTypeLog,
+	)
+
+	s.Engine.ExportSweepWebhookRetries(
+		context.Background(), s.WebhookID,
+	)
+
+	iAssertStatus(t, s.WebhookDB, deliveryID,
+		database.DeliveryStatusFailed,
+	)
+
+	assert.InDelta(t, 1.0,
+		mCounter(t, reg, mFailed, mTypeLog), 0)
+}
+
+// TestDeliveryMetrics_QueueDepthGauges proves the sampler publishes
+// the queued deliveries it finds in the per-webhook databases, and
+// that a drained queue reads zero rather than keeping its last
+// value.
+func TestDeliveryMetrics_QueueDepthGauges(t *testing.T) {
+	t.Parallel()
+
+	s := newISetup(t)
+	reg := mIsolate(t, s)
+
+	iCreateWebhook(
+		t, s.MainDB, s.WebhookID, "queue-depth",
+	)
+
+	targetID := uuid.New().String()
+
+	iCreateTarget(t, s.MainDB, targetID, s.WebhookID,
+		"queue-depth-target", database.TargetTypeHTTP,
+		iHTTPConfig("https://example.com/hook"), 3,
+	)
+
+	event := iSeedEvent(
+		t, s.WebhookDB, s.WebhookID, `{"queued":true}`,
+	)
+
+	pending := iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusPending,
+	)
+
+	iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusPending,
+	)
+
+	retrying := iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusRetrying,
+	)
+
+	s.Engine.ExportSampleQueueDepths(context.Background())
+
+	assert.InDelta(t, 2.0,
+		mGauge(t, reg, mPending, mTypeHTTP), 0)
+	assert.InDelta(t, 1.0,
+		mGauge(t, reg, mRetrying, mTypeHTTP), 0)
+	assert.InDelta(t, 0.0,
+		mGauge(t, reg, mPending, mTypeLog), 0)
+
+	require.NoError(t, s.WebhookDB.
+		Model(&database.Delivery{}).
+		Where("id IN ?", []string{pending.ID, retrying.ID}).
+		Update(
+			"status", database.DeliveryStatusDelivered,
+		).Error)
+
+	s.Engine.ExportSampleQueueDepths(context.Background())
+
+	assert.InDelta(t, 1.0,
+		mGauge(t, reg, mPending, mTypeHTTP), 0)
+	assert.InDelta(t, 0.0,
+		mGauge(t, reg, mRetrying, mTypeHTTP), 0)
+}
+
+// TestDeliveryMetrics_QueueDepthDeletedTarget proves a backlog queued
+// against a target that has since been deleted stays visible, in the
+// unknown series, instead of being dropped. That backlog is the one
+// nobody is watching, so losing it would defeat the queue-depth
+// alerting this metric exists for.
+func TestDeliveryMetrics_QueueDepthDeletedTarget(t *testing.T) {
+	t.Parallel()
+
+	s := newISetup(t)
+	reg := mIsolate(t, s)
+
+	iCreateWebhook(
+		t, s.MainDB, s.WebhookID, "deleted-target",
+	)
+
+	// No target row is created: this is a delivery whose target was
+	// deleted out from under it.
+	targetID := uuid.New().String()
+
+	event := iSeedEvent(
+		t, s.WebhookDB, s.WebhookID, `{"orphan":true}`,
+	)
+
+	iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusPending,
+	)
+
+	iSeedDelivery(
+		t, s.WebhookDB, event.ID, targetID,
+		database.DeliveryStatusRetrying,
+	)
+
+	s.Engine.ExportSampleQueueDepths(context.Background())
+
+	assert.InDelta(t, 1.0,
+		mGauge(t, reg, mPending, mTypeUnknown), 0)
+	assert.InDelta(t, 1.0,
+		mGauge(t, reg, mRetrying, mTypeUnknown), 0)
+	assert.InDelta(t, 0.0,
+		mGauge(t, reg, mPending, mTypeHTTP), 0)
+}

internal/delivery/queue_depth.go

@@ -0,0 +1,187 @@
+package delivery
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	"sneak.berlin/go/webhooker/internal/database"
+)
+
+// queueDepthSampleInterval is how often the pending and retrying
+// queue depths are counted and published as gauges.
+const queueDepthSampleInterval = 30 * time.Second
+
+// queueDepthSampler publishes the pending and retrying queue depths
+// on a timer for as long as the engine runs.
+//
+// The depths are counted out of the databases rather than tracked as
+// deltas alongside the status transitions. A delta counter would have
+// to be seeded correctly at startup from rows written by a previous
+// process, and would drift permanently on any transition that failed
+// to persist. Counting is the measurement that cannot go wrong, and
+// it is the same whole-database walk the retry sweep already makes.
+func (e *Engine) queueDepthSampler(ctx context.Context) {
+	defer e.wg.Done()
+
+	ticker := time.NewTicker(queueDepthSampleInterval)
+	defer ticker.Stop()
+
+	e.sampleQueueDepths(ctx)
+
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			e.sampleQueueDepths(ctx)
+		}
+	}
+}
+
+// sampleQueueDepths counts every queued delivery across all
+// per-webhook databases and publishes the result.
+func (e *Engine) sampleQueueDepths(ctx context.Context) {
+	if e.database == nil || e.dbManager == nil {
+		return
+	}
+
+	types, err := e.targetTypesByID()
+	if err != nil {
+		e.log.Error(
+			"queue depth sample: failed to load target types",
+			"error", err,
+		)
+
+		return
+	}
+
+	var webhookIDs []string
+
+	err = e.database.DB().
+		Model(&database.Webhook{}).
+		Pluck("id", &webhookIDs).Error
+	if err != nil {
+		e.log.Error(
+			"queue depth sample: failed to query webhook IDs",
+			"error", err,
+		)
+
+		return
+	}
+
+	pending := make(map[database.TargetType]int)
+	retrying := make(map[database.TargetType]int)
+
+	for _, webhookID := range webhookIDs {
+		select {
+		case <-ctx.Done():
+			return
+		default:
+		}
+
+		if !e.dbManager.DBExists(webhookID) {
+			continue
+		}
+
+		e.sampleWebhookQueueDepths(
+			webhookID, types, pending, retrying,
+		)
+	}
+
+	e.mtr.SetQueueDepths(pending, retrying)
+}
+
+// targetTypesByID maps every configured target id to its type. The
+// deliveries live in the per-webhook databases but carry only a
+// target id, so the type label has to come from the main database.
+func (e *Engine) targetTypesByID() (
+	map[string]database.TargetType, error,
+) {
+	var rows []struct {
+		ID   string
+		Type database.TargetType
+	}
+
+	err := e.database.DB().
+		Model(&database.Target{}).
+		Select("id", "type").
+		Scan(&rows).Error
+	if err != nil {
+		return nil, fmt.Errorf("loading targets: %w", err)
+	}
+
+	types := make(map[string]database.TargetType, len(rows))
+
+	for _, row := range rows {
+		types[row.ID] = row.Type
+	}
+
+	return types, nil
+}
+
+// sampleWebhookQueueDepths adds one webhook's queued deliveries into
+// the running totals.
+//
+// A delivery whose target has since been deleted is not in the type
+// map and so counts under the empty target type. Set.SetQueueDepths
+// folds that into the unknown series rather than dropping it: a
+// backlog stuck behind a deleted target is a backlog that still needs
+// to be alertable.
+func (e *Engine) sampleWebhookQueueDepths(
+	webhookID string,
+	types map[string]database.TargetType,
+	pending, retrying map[database.TargetType]int,
+) {
+	webhookDB, err := e.dbManager.GetDB(webhookID)
+	if err != nil {
+		e.log.Error(
+			"queue depth sample: failed to get webhook database",
+			"webhook_id", webhookID,
+			"error", err,
+		)
+
+		return
+	}
+
+	var rows []struct {
+		TargetID string
+		Status   database.DeliveryStatus
+		Depth    int
+	}
+
+	err = webhookDB.
+		Model(&database.Delivery{}).
+		Select("target_id", "status", "count(*) as depth").
+		Where("status IN ?", []database.DeliveryStatus{
+			database.DeliveryStatusPending,
+			database.DeliveryStatusRetrying,
+		}).
+		Group("target_id, status").
+		Scan(&rows).Error
+	if err != nil {
+		e.log.Error(
+			"queue depth sample: "+
+				"failed to count queued deliveries",
+			"webhook_id", webhookID,
+			"error", err,
+		)
+
+		return
+	}
+
+	for _, row := range rows {
+		targetType := types[row.TargetID]
+
+		switch row.Status {
+		case database.DeliveryStatusPending:
+			pending[targetType] += row.Depth
+		case database.DeliveryStatusRetrying:
+			retrying[targetType] += row.Depth
+		case database.DeliveryStatusDelivered,
+			database.DeliveryStatusFailed:
+			// Excluded by the query above: a delivery that has
+			// reached a terminal state is not queued.
+		}
+	}
+}

internal/delivery/target.go

@@ -27,6 +27,12 @@ type Scheduler interface {
 // own circuit breaker, and reschedules via the injected
 // Scheduler. Fire-and-forget targets simply record a single
 // attempt.
+//
+// An implementation reports each attempt it actually dispatches to
+// Engine.observeAttempt, alongside the DeliveryResult it records for
+// it. Deliver is also entered for attempts that never happen — an
+// open circuit breaker refuses one — so the count cannot be taken
+// from around this call.
 type Target interface {
 	Deliver(
 		ctx context.Context,
@@ -74,6 +80,12 @@ type attemptResult struct {
 	errMsg     string
 }
 
+// elapsed returns how long the attempt took. The field is stored in
+// milliseconds because that is what DeliveryResult persists.
+func (r attemptResult) elapsed() time.Duration {
+	return time.Duration(r.duration) * time.Millisecond
+}
+
 // initTargets builds the target registry, wiring each target
 // to the engine's persistence helpers and giving the HTTP and
 // Slack targets the shared SSRF-safe client. It is called by

internal/delivery/target_database.go

@@ -42,7 +42,14 @@ func (t *databaseTarget) Deliver(
 	_ *Task,
 	_ Scheduler,
 ) {
+	start := time.Now()
+
 	err := t.archive(d)
+
+	elapsed := time.Since(start)
+
+	t.eng.observeAttempt(d.Target.Type, elapsed)
+
 	if err != nil {
 		t.eng.log.Error(
 			"failed to archive event to database target",
@@ -53,22 +60,25 @@ func (t *databaseTarget) Deliver(
 
 		t.eng.recordResult(
 			webhookDB, d, 1, false, 0, "",
-			err.Error(), 0,
+			err.Error(), elapsed.Milliseconds(),
 		)
 
 		t.eng.updateDeliveryStatus(
-			webhookDB, d, database.DeliveryStatusFailed,
+			webhookDB, d, d.Target.Type,
+			database.DeliveryStatusFailed,
 		)
 
 		return
 	}
 
 	t.eng.recordResult(
-		webhookDB, d, 1, true, 0, "", "", 0,
+		webhookDB, d, 1, true, 0, "", "",
+		elapsed.Milliseconds(),
 	)
 
 	t.eng.updateDeliveryStatus(
-		webhookDB, d, database.DeliveryStatusDelivered,
+		webhookDB, d, d.Target.Type,
+		database.DeliveryStatusDelivered,
 	)
 }
 

internal/delivery/target_http.go

@@ -74,6 +74,8 @@ func (c *httpCore) fireAndForget(
 	d *database.Delivery,
 	res attemptResult,
 ) {
+	c.eng.observeAttempt(d.Target.Type, res.elapsed())
+
 	c.eng.recordResult(
 		webhookDB, d, 1, res.success,
 		res.statusCode, res.respBody, res.errMsg,
@@ -82,7 +84,7 @@ func (c *httpCore) fireAndForget(
 
 	if res.success {
 		c.eng.updateDeliveryStatus(
-			webhookDB, d,
+			webhookDB, d, d.Target.Type,
 			database.DeliveryStatusDelivered,
 		)
 
@@ -90,7 +92,8 @@ func (c *httpCore) fireAndForget(
 	}
 
 	c.eng.updateDeliveryStatus(
-		webhookDB, d, database.DeliveryStatusFailed,
+		webhookDB, d, d.Target.Type,
+		database.DeliveryStatusFailed,
 	)
 }
 
@@ -107,10 +110,17 @@ func (c *httpCore) withRetry(
 		return
 	}
 
+	// Allow may have moved the breaker to half-open, and the
+	// attempt below may open or close it, so the gauge is
+	// republished on every exit from here.
+	defer c.publishCircuitState(d.Target.Type)
+
 	attemptNum := task.AttemptNum
 
 	res := attempt()
 
+	c.eng.observeAttempt(d.Target.Type, res.elapsed())
+
 	c.eng.recordResult(
 		webhookDB, d, attemptNum, res.success,
 		res.statusCode, res.respBody, res.errMsg,
@@ -121,7 +131,7 @@ func (c *httpCore) withRetry(
 		cb.RecordSuccess()
 
 		c.eng.updateDeliveryStatus(
-			webhookDB, d,
+			webhookDB, d, d.Target.Type,
 			database.DeliveryStatusDelivered,
 		)
 
@@ -146,6 +156,8 @@ func (c *httpCore) circuitBreakerBlock(
 		return false
 	}
 
+	defer c.publishCircuitState(d.Target.Type)
+
 	remaining := cb.CooldownRemaining()
 
 	c.eng.log.Info(
@@ -157,7 +169,7 @@ func (c *httpCore) circuitBreakerBlock(
 	)
 
 	c.eng.updateDeliveryStatus(
-		webhookDB, d,
+		webhookDB, d, d.Target.Type,
 		database.DeliveryStatusRetrying,
 	)
 
@@ -177,7 +189,7 @@ func (c *httpCore) handleRetry(
 ) {
 	if attemptNum >= maxRetries {
 		c.eng.updateDeliveryStatus(
-			webhookDB, d,
+			webhookDB, d, d.Target.Type,
 			database.DeliveryStatusFailed,
 		)
 
@@ -185,7 +197,8 @@ func (c *httpCore) handleRetry(
 	}
 
 	c.eng.updateDeliveryStatus(
-		webhookDB, d, database.DeliveryStatusRetrying,
+		webhookDB, d, d.Target.Type,
+		database.DeliveryStatusRetrying,
 	)
 
 	backoff := calcBackoff(attemptNum)
@@ -215,6 +228,28 @@ func (c *httpCore) getCircuitBreaker(
 	return cb
 }
 
+// publishCircuitState recounts this core's open breakers and
+// publishes the gauge. Each core holds the breakers of exactly one
+// target type, so the recount is over that type's targets alone.
+// Counting rather than adjusting a delta keeps the gauge honest
+// however a breaker changed state.
+func (c *httpCore) publishCircuitState(
+	targetType database.TargetType,
+) {
+	open := 0
+
+	c.circuitBreakers.Range(func(_, val any) bool {
+		cb, ok := val.(*CircuitBreaker)
+		if ok && cb.State() == CircuitOpen {
+			open++
+		}
+
+		return true
+	})
+
+	c.eng.mtr.SetCircuitBreakersOpen(targetType, open)
+}
+
 // remainingBackoff returns how long remains of the backoff
 // window for the last attempt of a recovered retrying
 // delivery. It implements rescheduler.
@@ -302,7 +337,8 @@ func (t *httpTarget) Deliver(
 		)
 
 		t.eng.updateDeliveryStatus(
-			webhookDB, d, database.DeliveryStatusFailed,
+			webhookDB, d, d.Target.Type,
+			database.DeliveryStatusFailed,
 		)
 
 		return

internal/delivery/target_log.go

@@ -2,6 +2,7 @@ package delivery
 
 import (
 	"context"
+	"time"
 
 	"gorm.io/gorm"
 	"sneak.berlin/go/webhooker/internal/database"
@@ -34,6 +35,8 @@ func (t *logTarget) Deliver(
 	_ *Task,
 	_ Scheduler,
 ) {
+	start := time.Now()
+
 	t.eng.log.Info(
 		"webhook event delivered to log target",
 		"delivery_id", d.ID,
@@ -48,11 +51,17 @@ func (t *logTarget) Deliver(
 		"body", d.Event.Body,
 	)
 
+	elapsed := time.Since(start)
+
+	t.eng.observeAttempt(d.Target.Type, elapsed)
+
 	t.eng.recordResult(
-		webhookDB, d, 1, true, 0, "", "", 0,
+		webhookDB, d, 1, true, 0, "", "",
+		elapsed.Milliseconds(),
 	)
 
 	t.eng.updateDeliveryStatus(
-		webhookDB, d, database.DeliveryStatusDelivered,
+		webhookDB, d, d.Target.Type,
+		database.DeliveryStatusDelivered,
 	)
 }

internal/delivery/target_slack.go

@@ -101,7 +101,8 @@ func (t *slackTarget) failConfig(
 	)
 
 	t.eng.updateDeliveryStatus(
-		webhookDB, d, database.DeliveryStatusFailed,
+		webhookDB, d, d.Target.Type,
+		database.DeliveryStatusFailed,
 	)
 }
 

internal/handlers/handlers.go

@@ -18,6 +18,7 @@ import (
 	"sneak.berlin/go/webhooker/internal/globals"
 	"sneak.berlin/go/webhooker/internal/healthcheck"
 	"sneak.berlin/go/webhooker/internal/logger"
+	"sneak.berlin/go/webhooker/internal/metrics"
 	"sneak.berlin/go/webhooker/internal/middleware"
 	"sneak.berlin/go/webhooker/internal/session"
 	"sneak.berlin/go/webhooker/templates"
@@ -73,6 +74,7 @@ type Handlers struct {
 	mw        *middleware.Middleware
 	notifier  delivery.Notifier
 	evictor   delivery.WebhookEvictor
+	mtr       *metrics.Set
 	templates map[string]*template.Template
 
 	// dummyVerifications counts the equivalent-cost verifications
@@ -114,6 +116,7 @@ func New(
 	s.mw = params.Middleware
 	s.notifier = params.Notifier
 	s.evictor = params.Evictor
+	s.mtr = metrics.Default()
 
 	// Parse all page templates once at startup
 	s.templates = map[string]*template.Template{

internal/handlers/webhook.go

@@ -217,6 +217,11 @@ func (h *Handlers) createAndDeliverEvent(
 		return
 	}
 
+	// Counted here, after the commit: an event is received once it
+	// is durably stored, which is what the delivery counters are
+	// compared against on a dashboard.
+	h.mtr.EventReceived()
+
 	h.finishWebhookResponse(w, event, entrypoint, tasks)
 }
 

internal/metrics/metrics.go

@@ -0,0 +1,344 @@
+// Package metrics defines the Prometheus collectors describing
+// webhooker's delivery pipeline: how many events arrive, how many
+// deliveries are attempted, how they end, how long they take, how
+// deep the queues are, and how many circuit breakers are open.
+//
+// The inbound HTTP metrics come from the go-http-metrics recorder in
+// internal/middleware and land on prometheus.DefaultRegisterer. These
+// collectors register there too, so both surfaces are gathered by the
+// one promhttp handler mounted on the authenticated /metrics route.
+package metrics
+
+import (
+	"sync"
+	"time"
+
+	"github.com/prometheus/client_golang/prometheus"
+	"github.com/prometheus/client_golang/prometheus/promauto"
+	"sneak.berlin/go/webhooker/internal/database"
+)
+
+// namespace prefixes every collector defined here.
+const namespace = "webhooker"
+
+// targetTypeLabel is the only label any delivery metric carries, and
+// cardinality is the whole reason for that.
+//
+// A target type is one of four compile-time constants, so the label
+// domain is bounded by construction. Target ids, event ids and
+// entrypoint ids are not: they are UUIDs minted per operator action
+// or per inbound request, a series is never reclaimed once it exists,
+// and labelling by any of them makes /metrics a memory leak that
+// grows with traffic. normalizeTargetType enforces the bound at every
+// call site — a type the registry does not know collapses into
+// unknownTargetType rather than minting a series of its own.
+const targetTypeLabel = "target_type"
+
+// unknownTargetType is the bucket for a target type outside the known
+// set, so an unrecognised value cannot mint a new series.
+const unknownTargetType = "unknown"
+
+// Delivery duration buckets, exponential from 5ms so the last bucket
+// (about 98s) sits above the 30s outbound HTTP client timeout.
+const (
+	durationBucketStart  = 0.005
+	durationBucketFactor = 3
+	durationBucketCount  = 10
+)
+
+// knownTargetTypes is the fixed label domain: the target types the
+// delivery engine implements.
+//
+//nolint:gochecknoglobals // the label domain, built once per process
+var knownTargetTypes = []database.TargetType{
+	database.TargetTypeHTTP,
+	database.TargetTypeDatabase,
+	database.TargetTypeLog,
+	database.TargetTypeSlack,
+}
+
+// defaultSet is the process-wide metric set, registered on the same
+// registry the HTTP middleware and the /metrics handler already use.
+// It is built on first use rather than in an init so that a test
+// binary that never touches metrics never registers them.
+//
+//nolint:gochecknoglobals // one process-wide registration, by design
+var defaultSet = sync.OnceValue(func() *Set {
+	return New(prometheus.DefaultRegisterer)
+})
+
+// Default returns the process-wide metric set.
+func Default() *Set {
+	return defaultSet()
+}
+
+// Set is one registered group of webhooker's delivery collectors.
+// Production uses the single Default set; tests build their own
+// against a private registry so assertions are not disturbed by
+// deliveries other tests are making concurrently.
+type Set struct {
+	eventsReceived      prometheus.Counter
+	deliveryAttempts    *prometheus.CounterVec
+	deliveriesSucceeded *prometheus.CounterVec
+	deliveriesFailed    *prometheus.CounterVec
+	deliveryRetries     *prometheus.CounterVec
+	deliveryDuration    *prometheus.HistogramVec
+	deliveriesPending   *prometheus.GaugeVec
+	deliveriesRetrying  *prometheus.GaugeVec
+	circuitBreakersOpen *prometheus.GaugeVec
+}
+
+// New registers a full set of delivery collectors on reg and returns
+// it. It panics if reg already holds them, which is the intended
+// behaviour for a duplicate registration.
+func New(reg prometheus.Registerer) *Set {
+	factory := promauto.With(reg)
+
+	s := &Set{
+		eventsReceived: factory.NewCounter(
+			prometheus.CounterOpts{
+				Namespace: namespace,
+				Name:      "events_received_total",
+				Help: "Webhook events received and " +
+					"stored, so the receive and deliver " +
+					"sides can be compared.",
+			},
+		),
+		deliveryDuration: factory.NewHistogramVec(
+			prometheus.HistogramOpts{
+				Namespace: namespace,
+				Name:      "delivery_duration_seconds",
+				Help: "Wall time of a single delivery " +
+					"attempt, by target type.",
+				Buckets: prometheus.ExponentialBuckets(
+					durationBucketStart,
+					durationBucketFactor,
+					durationBucketCount,
+				),
+			},
+			[]string{targetTypeLabel},
+		),
+	}
+
+	s.registerCounters(factory)
+	s.registerGauges(factory)
+	s.initSeries()
+
+	return s
+}
+
+// EventReceived counts one inbound webhook event stored.
+func (s *Set) EventReceived() {
+	s.eventsReceived.Inc()
+}
+
+// DeliveryAttempted counts one delivery attempt dispatched to a
+// target.
+func (s *Set) DeliveryAttempted(t database.TargetType) {
+	s.deliveryAttempts.
+		WithLabelValues(normalizeTargetType(t)).
+		Inc()
+}
+
+// ObserveDeliveryDuration records how long one delivery attempt took.
+func (s *Set) ObserveDeliveryDuration(
+	t database.TargetType, d time.Duration,
+) {
+	s.deliveryDuration.
+		WithLabelValues(normalizeTargetType(t)).
+		Observe(d.Seconds())
+}
+
+// DeliveryStatusChanged counts a delivery's transition into a new
+// status. The mapping from status to counter lives here, next to the
+// collectors, so the engine has a single call for every transition it
+// persists. A move back to pending is not an outcome and counts
+// nothing.
+func (s *Set) DeliveryStatusChanged(
+	t database.TargetType, status database.DeliveryStatus,
+) {
+	label := normalizeTargetType(t)
+
+	switch status {
+	case database.DeliveryStatusDelivered:
+		s.deliveriesSucceeded.WithLabelValues(label).Inc()
+	case database.DeliveryStatusFailed:
+		s.deliveriesFailed.WithLabelValues(label).Inc()
+	case database.DeliveryStatusRetrying:
+		s.deliveryRetries.WithLabelValues(label).Inc()
+	case database.DeliveryStatusPending:
+	}
+}
+
+// SetQueueDepths publishes the pending and retrying queue depths from
+// one sample. Every label in the queue domain is written on every
+// call, so a type whose queue has drained reads zero instead of
+// holding its last value forever.
+func (s *Set) SetQueueDepths(
+	pending, retrying map[database.TargetType]int,
+) {
+	pendingByLabel := foldToLabels(pending)
+	retryingByLabel := foldToLabels(retrying)
+
+	for _, label := range queueDepthLabels() {
+		s.deliveriesPending.WithLabelValues(label).
+			Set(float64(pendingByLabel[label]))
+		s.deliveriesRetrying.WithLabelValues(label).
+			Set(float64(retryingByLabel[label]))
+	}
+}
+
+// queueDepthLabels is the label domain of the two queue-depth gauges:
+// the known target types plus unknown.
+//
+// Unknown is a real bucket here, not a safety net. A delivery queued
+// against a target that has since been deleted carries a target id no
+// longer in the targets table, so the sample resolves it to the empty
+// type; folding it into unknown is what keeps that backlog visible.
+// Dropping it would hide the one queue nobody is watching.
+func queueDepthLabels() []string {
+	labels := make([]string, 0, len(knownTargetTypes)+1)
+
+	for _, t := range knownTargetTypes {
+		labels = append(labels, string(t))
+	}
+
+	return append(labels, unknownTargetType)
+}
+
+// foldToLabels collapses a per-target-type count onto the bounded
+// label domain, summing everything outside the known set into
+// unknown.
+func foldToLabels(
+	counts map[database.TargetType]int,
+) map[string]int {
+	byLabel := make(map[string]int, len(counts))
+
+	for t, n := range counts {
+		byLabel[normalizeTargetType(t)] += n
+	}
+
+	return byLabel
+}
+
+// SetCircuitBreakersOpen publishes how many of a target type's
+// circuit breakers are currently open.
+func (s *Set) SetCircuitBreakersOpen(
+	t database.TargetType, open int,
+) {
+	s.circuitBreakersOpen.
+		WithLabelValues(normalizeTargetType(t)).
+		Set(float64(open))
+}
+
+func (s *Set) registerCounters(factory promauto.Factory) {
+	s.deliveryAttempts = factory.NewCounterVec(
+		prometheus.CounterOpts{
+			Namespace: namespace,
+			Name:      "delivery_attempts_total",
+			Help: "Delivery attempts dispatched to a " +
+				"target, by target type.",
+		},
+		[]string{targetTypeLabel},
+	)
+
+	s.deliveriesSucceeded = factory.NewCounterVec(
+		prometheus.CounterOpts{
+			Namespace: namespace,
+			Name:      "deliveries_succeeded_total",
+			Help: "Deliveries that reached the delivered " +
+				"state, by target type.",
+		},
+		[]string{targetTypeLabel},
+	)
+
+	s.deliveriesFailed = factory.NewCounterVec(
+		prometheus.CounterOpts{
+			Namespace: namespace,
+			Name:      "deliveries_failed_total",
+			Help: "Deliveries that failed terminally and " +
+				"will not be retried, by target type.",
+		},
+		[]string{targetTypeLabel},
+	)
+
+	s.deliveryRetries = factory.NewCounterVec(
+		prometheus.CounterOpts{
+			Namespace: namespace,
+			Name:      "delivery_retries_total",
+			Help: "Deliveries put back into the retrying " +
+				"state, by target type.",
+		},
+		[]string{targetTypeLabel},
+	)
+}
+
+func (s *Set) registerGauges(factory promauto.Factory) {
+	s.deliveriesPending = factory.NewGaugeVec(
+		prometheus.GaugeOpts{
+			Namespace: namespace,
+			Name:      "deliveries_pending",
+			Help: "Deliveries currently in the pending " +
+				"state, by target type.",
+		},
+		[]string{targetTypeLabel},
+	)
+
+	s.deliveriesRetrying = factory.NewGaugeVec(
+		prometheus.GaugeOpts{
+			Namespace: namespace,
+			Name:      "deliveries_retrying",
+			Help: "Deliveries currently in the retrying " +
+				"state, by target type.",
+		},
+		[]string{targetTypeLabel},
+	)
+
+	s.circuitBreakersOpen = factory.NewGaugeVec(
+		prometheus.GaugeOpts{
+			Namespace: namespace,
+			Name:      "circuit_breakers_open",
+			Help: "Delivery circuit breakers currently " +
+				"open, by target type.",
+		},
+		[]string{targetTypeLabel},
+	)
+}
+
+// initSeries materialises every known-target-type series at zero, so
+// a dashboard and an alert rule see a target type that has not
+// delivered yet rather than a missing series.
+//
+// The queue-depth gauges additionally get their unknown series, which
+// holds deliveries queued against a deleted target. That backlog can
+// predate the process — it is read out of the databases, not counted
+// from transitions — so its series has to exist from the first scrape
+// rather than appearing only once a backlog has already built up.
+func (s *Set) initSeries() {
+	for _, t := range knownTargetTypes {
+		label := string(t)
+
+		s.deliveryAttempts.WithLabelValues(label)
+		s.deliveriesSucceeded.WithLabelValues(label)
+		s.deliveriesFailed.WithLabelValues(label)
+		s.deliveryRetries.WithLabelValues(label)
+		s.deliveriesPending.WithLabelValues(label)
+		s.deliveriesRetrying.WithLabelValues(label)
+		s.circuitBreakersOpen.WithLabelValues(label)
+	}
+
+	s.deliveriesPending.WithLabelValues(unknownTargetType)
+	s.deliveriesRetrying.WithLabelValues(unknownTargetType)
+}
+
+// normalizeTargetType maps a target type onto the bounded label
+// domain, collapsing anything outside it to unknownTargetType.
+func normalizeTargetType(t database.TargetType) string {
+	for _, known := range knownTargetTypes {
+		if t == known {
+			return string(known)
+		}
+	}
+
+	return unknownTargetType
+}

internal/metrics/metrics_test.go

@@ -0,0 +1,285 @@
+package metrics_test
+
+import (
+	"testing"
+
+	"github.com/prometheus/client_golang/prometheus"
+	dto "github.com/prometheus/client_model/go"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"sneak.berlin/go/webhooker/internal/database"
+	"sneak.berlin/go/webhooker/internal/metrics"
+)
+
+// knownLabels is the target_type label domain built from the target
+// types the delivery engine implements.
+func knownLabels() []string {
+	return []string{"http", "database", "log", "slack"}
+}
+
+// labelValues returns the target_type label values a metric family
+// currently carries.
+func labelValues(
+	t *testing.T, reg *prometheus.Registry, name string,
+) []string {
+	t.Helper()
+
+	families, err := reg.Gather()
+	require.NoError(t, err)
+
+	var values []string
+
+	for _, fam := range families {
+		if fam.GetName() != name {
+			continue
+		}
+
+		for _, m := range fam.GetMetric() {
+			for _, label := range m.GetLabel() {
+				if label.GetName() == "target_type" {
+					values = append(
+						values, label.GetValue(),
+					)
+				}
+			}
+		}
+	}
+
+	return values
+}
+
+func gaugeValue(
+	t *testing.T,
+	reg *prometheus.Registry,
+	name, targetType string,
+) float64 {
+	t.Helper()
+
+	families, err := reg.Gather()
+	require.NoError(t, err)
+
+	for _, fam := range families {
+		if fam.GetName() != name {
+			continue
+		}
+
+		for _, m := range fam.GetMetric() {
+			if hasTargetType(m, targetType) {
+				return m.GetGauge().GetValue()
+			}
+		}
+	}
+
+	t.Fatalf(
+		"gauge %s{target_type=%q} not found",
+		name, targetType,
+	)
+
+	return 0
+}
+
+func hasTargetType(m *dto.Metric, targetType string) bool {
+	for _, label := range m.GetLabel() {
+		if label.GetName() == "target_type" &&
+			label.GetValue() == targetType {
+			return true
+		}
+	}
+
+	return false
+}
+
+// TestUnknownTargetTypeCollapses is the cardinality guard: a target
+// type outside the known set must not mint a series of its own, or
+// /metrics grows without bound.
+func TestUnknownTargetTypeCollapses(t *testing.T) {
+	t.Parallel()
+
+	reg := prometheus.NewRegistry()
+	set := metrics.New(reg)
+
+	for _, bogus := range []string{
+		"a1b2c3d4-0000-0000-0000-000000000001",
+		"a1b2c3d4-0000-0000-0000-000000000002",
+		"webhook-forwarder",
+	} {
+		set.DeliveryAttempted(database.TargetType(bogus))
+	}
+
+	values := labelValues(
+		t, reg, "webhooker_delivery_attempts_total",
+	)
+
+	assert.ElementsMatch(t,
+		append(knownLabels(), "unknown"),
+		values,
+	)
+}
+
+// TestSetQueueDepthsZeroesDrainedTypes proves a queue that has
+// drained reads zero rather than holding its last sample forever.
+func TestSetQueueDepthsZeroesDrainedTypes(t *testing.T) {
+	t.Parallel()
+
+	reg := prometheus.NewRegistry()
+	set := metrics.New(reg)
+
+	set.SetQueueDepths(
+		map[database.TargetType]int{
+			database.TargetTypeHTTP: 7,
+		},
+		map[database.TargetType]int{
+			database.TargetTypeSlack: 2,
+		},
+	)
+
+	assert.InDelta(t, 7.0, gaugeValue(
+		t, reg, "webhooker_deliveries_pending", "http",
+	), 0)
+	assert.InDelta(t, 2.0, gaugeValue(
+		t, reg, "webhooker_deliveries_retrying", "slack",
+	), 0)
+
+	set.SetQueueDepths(
+		map[database.TargetType]int{},
+		map[database.TargetType]int{},
+	)
+
+	assert.InDelta(t, 0.0, gaugeValue(
+		t, reg, "webhooker_deliveries_pending", "http",
+	), 0)
+	assert.InDelta(t, 0.0, gaugeValue(
+		t, reg, "webhooker_deliveries_retrying", "slack",
+	), 0)
+}
+
+// TestKnownSeriesExistBeforeAnyDelivery proves every known target
+// type is published at zero from registration, so an alert rule does
+// not have to cope with a missing series.
+func TestKnownSeriesExistBeforeAnyDelivery(t *testing.T) {
+	t.Parallel()
+
+	reg := prometheus.NewRegistry()
+	metrics.New(reg)
+
+	for _, name := range []string{
+		"webhooker_delivery_attempts_total",
+		"webhooker_deliveries_succeeded_total",
+		"webhooker_deliveries_failed_total",
+		"webhooker_delivery_retries_total",
+		"webhooker_circuit_breakers_open",
+	} {
+		assert.ElementsMatch(t,
+			knownLabels(),
+			labelValues(t, reg, name),
+			"metric %s", name,
+		)
+	}
+
+	// The queue gauges additionally publish unknown from
+	// registration: a backlog queued against a deleted target lands
+	// there, and it can predate the process, so the series has to
+	// exist before the first sample rather than appearing only once
+	// something is already stuck.
+	for _, name := range []string{
+		"webhooker_deliveries_pending",
+		"webhooker_deliveries_retrying",
+	} {
+		assert.ElementsMatch(t,
+			append(knownLabels(), "unknown"),
+			labelValues(t, reg, name),
+			"metric %s", name,
+		)
+	}
+}
+
+// TestSetQueueDepthsFoldsUnknownTypes proves a queued delivery whose
+// target type is not a known one — a target deleted out from under it
+// resolves to the empty type — is summed into the unknown series
+// instead of being dropped, and that the fold is a sum rather than a
+// last-writer-wins.
+func TestSetQueueDepthsFoldsUnknownTypes(t *testing.T) {
+	t.Parallel()
+
+	reg := prometheus.NewRegistry()
+	set := metrics.New(reg)
+
+	set.SetQueueDepths(
+		map[database.TargetType]int{
+			database.TargetTypeHTTP:             1,
+			database.TargetType(""):             4,
+			database.TargetType("retired-type"): 3,
+		},
+		map[database.TargetType]int{
+			database.TargetType(""): 2,
+		},
+	)
+
+	assert.InDelta(t, 7.0, gaugeValue(
+		t, reg, "webhooker_deliveries_pending", "unknown",
+	), 0)
+	assert.InDelta(t, 2.0, gaugeValue(
+		t, reg, "webhooker_deliveries_retrying", "unknown",
+	), 0)
+	assert.InDelta(t, 1.0, gaugeValue(
+		t, reg, "webhooker_deliveries_pending", "http",
+	), 0)
+
+	set.SetQueueDepths(
+		map[database.TargetType]int{},
+		map[database.TargetType]int{},
+	)
+
+	assert.InDelta(t, 0.0, gaugeValue(
+		t, reg, "webhooker_deliveries_pending", "unknown",
+	), 0)
+}
+
+// TestDeliveryStatusChangedCounts maps each persisted status onto the
+// counter it moves.
+func TestDeliveryStatusChangedCounts(t *testing.T) {
+	t.Parallel()
+
+	reg := prometheus.NewRegistry()
+	set := metrics.New(reg)
+
+	set.DeliveryStatusChanged(
+		database.TargetTypeLog,
+		database.DeliveryStatusDelivered,
+	)
+	set.DeliveryStatusChanged(
+		database.TargetTypeLog,
+		database.DeliveryStatusFailed,
+	)
+	set.DeliveryStatusChanged(
+		database.TargetTypeLog,
+		database.DeliveryStatusRetrying,
+	)
+	set.DeliveryStatusChanged(
+		database.TargetTypeLog,
+		database.DeliveryStatusPending,
+	)
+
+	families, err := reg.Gather()
+	require.NoError(t, err)
+
+	counts := map[string]float64{}
+
+	for _, fam := range families {
+		for _, m := range fam.GetMetric() {
+			if hasTargetType(m, "log") {
+				counts[fam.GetName()] =
+					m.GetCounter().GetValue()
+			}
+		}
+	}
+
+	assert.InDelta(t, 1.0,
+		counts["webhooker_deliveries_succeeded_total"], 0)
+	assert.InDelta(t, 1.0,
+		counts["webhooker_deliveries_failed_total"], 0)
+	assert.InDelta(t, 1.0,
+		counts["webhooker_delivery_retries_total"], 0)
+	assert.InDelta(t, 0.0,
+		counts["webhooker_delivery_attempts_total"], 0)
+}

internal/server/http.go

@@ -43,10 +43,7 @@ func (s *Server) serveUntilShutdown() {
 	err := s.httpServer.ListenAndServe()
 	if err != nil && !errors.Is(err, http.ErrServerClosed) {
 		s.log.Error("listen error", "error", err)
-
-		if s.cancelFunc != nil {
-			s.cancelFunc()
-		}
+		s.shutdownOnListenFailure()
 	}
 }
 

internal/server/listen_failure_test.go

@@ -0,0 +1,95 @@
+package server_test
+
+import (
+	"context"
+	"net"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+	"go.uber.org/fx"
+	"sneak.berlin/go/webhooker/internal/globals"
+	"sneak.berlin/go/webhooker/internal/server"
+)
+
+// listenFailureDeadline is how long the app gets to give up after a
+// listen it cannot satisfy. The defect this pins left the process
+// reporting RUNNING for 183 seconds with nothing bound; a bind error
+// is known instantly, so anything past a moment here is that defect
+// back.
+const listenFailureDeadline = 2 * time.Second
+
+// lifecycleTimeout bounds the app's start and stop sequences so a
+// wedged hook fails the test instead of hanging it.
+const lifecycleTimeout = 15 * time.Second
+
+// TestListenFailure_ShutsDownTheApp pins that a listener the server
+// cannot bind terminates the application with a non-zero status.
+//
+// The fx OnStart hook returns as soon as the serving goroutine is
+// spawned, so a bind failure is discovered after fx has already
+// reported RUNNING. Nothing else in the graph observes it, and the
+// process used to stay alive with no listener: down, but indis-
+// tinguishable from healthy to systemd's Restart=on-failure and to
+// Docker's restart policies, which is the state this test exists to
+// keep from returning.
+//
+// The port is occupied by a listener this test holds open, on a
+// kernel-chosen port, so the failure is the real EADDRINUSE the
+// operator hits when a second instance starts. Loopback is enough to
+// collide with the server's wildcard bind: a listening socket on a
+// specific address blocks the wildcard from claiming the same port.
+func TestListenFailure_ShutsDownTheApp(t *testing.T) {
+	t.Parallel()
+
+	var listenCfg net.ListenConfig
+
+	occupied, err := listenCfg.Listen(
+		t.Context(), "tcp", "127.0.0.1:0",
+	)
+	require.NoError(t, err)
+
+	t.Cleanup(func() { _ = occupied.Close() })
+
+	addr, ok := occupied.Addr().(*net.TCPAddr)
+	require.True(t, ok, "listener is not TCP")
+
+	// The collaborators come from the wired graph rather than stubs,
+	// so the Server under test is the one that ships. Only the port
+	// is test-specific.
+	env := newTestEnv(t)
+	env.cfg.Port = addr.Port
+
+	app := fx.New(
+		fx.NopLogger,
+		fx.Supply(env.log, env.cfg, env.mw, env.hnd),
+		fx.Provide(globals.New, server.New),
+		fx.Invoke(func(*server.Server) {}),
+	)
+
+	startCtx, cancelStart := context.WithTimeout(
+		context.Background(), lifecycleTimeout,
+	)
+	defer cancelStart()
+
+	require.NoError(t, app.Start(startCtx))
+
+	select {
+	case sig := <-app.Wait():
+		require.Equal(
+			t, server.ListenFailureExitCode, sig.ExitCode,
+			"listen failure must exit non-zero",
+		)
+	case <-time.After(listenFailureDeadline):
+		t.Fatal("listen failure left the app running")
+	}
+
+	// The stop sequence still has to complete: the fix must reach
+	// shutdown through fx rather than around it.
+	stopCtx, cancelStop := context.WithTimeout(
+		context.Background(), lifecycleTimeout,
+	)
+	defer cancelStop()
+
+	require.NoError(t, app.Stop(stopCtx))
+}

internal/server/routes.go

@@ -55,8 +55,11 @@ func (s *Server) setupGlobalMiddleware() {
 	s.router.Use(s.mw.SecurityHeaders())
 	s.router.Use(s.mw.Logging())
 
-	// Metrics middleware (only if credentials are configured)
-	if s.params.Config.MetricsUsername != "" {
+	// Metrics recording middleware, registered only when the
+	// endpoint that exposes what it records is served. The
+	// condition is the same MetricsAuthEnabled the /metrics mount
+	// in setupRoutes reads.
+	if s.params.Config.MetricsAuthEnabled() {
 		s.router.Use(s.mw.Metrics())
 	}
 
@@ -103,8 +106,14 @@ func (s *Server) setupRoutes() {
 		s.h.HandleHealthCheck(),
 	)
 
-	// set up authenticated /metrics route:
-	if s.params.Config.MetricsUsername != "" {
+	// Authenticated /metrics route. The condition is
+	// Config.MetricsAuthEnabled and never the username alone: a
+	// username with an empty password would otherwise mount the
+	// endpoint behind a credential map that accepts an empty
+	// password. Config rejects that combination at startup, and
+	// this reads the same value the startup log reports, so the
+	// two cannot disagree about whether the route exists.
+	if s.params.Config.MetricsAuthEnabled() {
 		s.router.Group(func(r chi.Router) {
 			r.Use(s.mw.MetricsAuth())
 			r.Get(

internal/server/routes_test.go

@@ -34,6 +34,13 @@ import (
 // the CSRF middleware executed.
 const csrfCookieName = "_gorilla_csrf"
 
+const (
+	// metricsUser and metricsAuthValue are the /metrics basic-auth
+	// credentials the metrics routing tests below configure.
+	metricsUser      = "metrics"
+	metricsAuthValue = "s3cret"
+)
+
 type noopNotifier struct{}
 
 func (n *noopNotifier) Notify([]delivery.Task) {}
@@ -69,9 +76,23 @@ type testEnv struct {
 func newTestEnv(t *testing.T) *testEnv {
 	t.Helper()
 
+	return newTestEnvWithConfig(t, &config.Config{
+		DataDir:     t.TempDir(),
+		Environment: config.EnvironmentDev,
+	})
+}
+
+// newTestEnvWithConfig is newTestEnv over a caller-supplied Config,
+// for the routes whose existence the configuration decides. The same
+// pointer reaches the router and every middleware, so a test cannot
+// accidentally configure one and not the other.
+func newTestEnvWithConfig(
+	t *testing.T, cfg *config.Config,
+) *testEnv {
+	t.Helper()
+
 	var (
 		log   *logger.Logger
-		cfg   *config.Config
 		mw    *middleware.Middleware
 		hnd   *handlers.Handlers
 		sess  *session.Session
@@ -84,12 +105,7 @@ func newTestEnv(t *testing.T) *testEnv {
 		fx.Provide(
 			globals.New,
 			logger.New,
-			func() *config.Config {
-				return &config.Config{
-					DataDir:     t.TempDir(),
-					Environment: config.EnvironmentDev,
-				}
-			},
+			func() *config.Config { return cfg },
 			database.New,
 			database.NewWebhookDBManager,
 			healthcheck.New,
@@ -99,7 +115,7 @@ func newTestEnv(t *testing.T) *testEnv {
 			middleware.New,
 			handlers.New,
 		),
-		fx.Populate(&log, &cfg, &mw, &hnd, &sess, &db, &dbMgr),
+		fx.Populate(&log, &mw, &hnd, &sess, &db, &dbMgr),
 	)
 	app.RequireStart()
 	t.Cleanup(app.RequireStop)
@@ -657,3 +673,119 @@ func TestSourceLogsBody_OtherUser404s(t *testing.T) {
 	assert.Equal(t, http.StatusSeeOther, anon.Code)
 	assert.Equal(t, "/pages/login", anon.Header().Get("Location"))
 }
+
+// metricsConfig is a Config differing from the routing default only
+// in the two /metrics credentials.
+func metricsConfig(
+	t *testing.T, username, password string,
+) *config.Config {
+	t.Helper()
+
+	return &config.Config{
+		DataDir:         t.TempDir(),
+		Environment:     config.EnvironmentDev,
+		MetricsUsername: username,
+		MetricsPassword: password,
+	}
+}
+
+// metricsRequest asks the real router for /metrics with the given
+// basic-auth credentials, or with no Authorization header when
+// username is empty.
+func (e *testEnv) metricsRequest(
+	username, password string,
+) *httptest.ResponseRecorder {
+	req := httptest.NewRequestWithContext(
+		context.Background(), http.MethodGet, "/metrics", nil,
+	)
+
+	if username != "" {
+		req.SetBasicAuth(username, password)
+	}
+
+	w := httptest.NewRecorder()
+	e.router.ServeHTTP(w, req)
+
+	return w
+}
+
+// TestMetricsRouteUnmountedWithoutCredentials pins that with neither
+// credential configured the route does not exist, which is the
+// documented behaviour and the only valid way for /metrics to be
+// absent.
+func TestMetricsRouteUnmountedWithoutCredentials(t *testing.T) {
+	t.Parallel()
+
+	env := newTestEnvWithConfig(t, metricsConfig(t, "", ""))
+
+	assert.Equal(
+		t, http.StatusNotFound,
+		env.metricsRequest("", "").Code,
+	)
+}
+
+// TestMetricsRouteRequiresCredentials pins that with both credentials
+// configured the route exists and every request that does not carry
+// the configured pair is refused — including the empty password that
+// a half-set configuration used to make sufficient.
+func TestMetricsRouteRequiresCredentials(t *testing.T) {
+	t.Parallel()
+
+	env := newTestEnvWithConfig(
+		t, metricsConfig(t, metricsUser, metricsAuthValue),
+	)
+
+	assert.Equal(
+		t, http.StatusUnauthorized,
+		env.metricsRequest("", "").Code,
+		"no credentials must not reach the metrics handler",
+	)
+	assert.Equal(
+		t, http.StatusUnauthorized,
+		env.metricsRequest(metricsUser, "").Code,
+		"an empty password must not reach the metrics handler",
+	)
+	assert.Equal(
+		t, http.StatusUnauthorized,
+		env.metricsRequest(metricsUser, "wrong").Code,
+	)
+
+	ok := env.metricsRequest(metricsUser, metricsAuthValue)
+	assert.Equal(t, http.StatusOK, ok.Code)
+	assert.Contains(t, ok.Body.String(), "go_goroutines")
+}
+
+// TestMetricsRouteUnmountedOnHalfSetConfig pins the defect from
+// https://git.eeqj.de/sneak/webhooker/issues/205 at the routing
+// layer. Config rejects a half-set pair at startup, so this Config
+// cannot be reached from the environment; the assertion is that the
+// route tree does not publish an endpoint accepting an empty
+// password even when handed one anyway, because the mount and the
+// startup log's hasMetricsAuth read the same value.
+func TestMetricsRouteUnmountedOnHalfSetConfig(t *testing.T) {
+	t.Parallel()
+
+	for _, tc := range []struct {
+		name     string
+		username string
+		password string
+	}{
+		{name: "username only", username: metricsUser},
+		{name: "password only", password: metricsAuthValue},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			cfg := metricsConfig(t, tc.username, tc.password)
+			env := newTestEnvWithConfig(t, cfg)
+
+			assert.False(t, cfg.MetricsAuthEnabled())
+			assert.Equal(
+				t, http.StatusNotFound,
+				env.metricsRequest(
+					tc.username, tc.password,
+				).Code,
+			)
+		})
+	}
+}

internal/server/server.go

@@ -50,6 +50,13 @@ const (
 	minSentryFlush = 250 * time.Millisecond
 )
 
+// ListenFailureExitCode is the status the process exits with when the
+// HTTP listener cannot be established, or dies for a reason other
+// than a requested shutdown. It must stay non-zero: systemd
+// `Restart=on-failure` and Docker's restart policies key off it, and a
+// zero exit would read as a deliberate stop.
+const ListenFailureExitCode = 1
+
 // SentryFlushBudget reports how long the Sentry flush may run when
 // remaining is the time left on the fx stop context after the HTTP
 // drain. sentry.Flush takes a bare duration and honours no context,
@@ -75,13 +82,13 @@ type ServerParams struct {
 	Config     *config.Config
 	Middleware *middleware.Middleware
 	Handlers   *handlers.Handlers
+	Shutdowner fx.Shutdowner
 }
 
 // Server is the main HTTP server that wires up routes and manages
 // graceful shutdown.
 type Server struct {
 	startupTime   time.Time
-	exitCode      int
 	sentryEnabled bool
 	log           *slog.Logger
 	cancelFunc    context.CancelFunc
@@ -159,7 +166,12 @@ func (s *Server) enableSentry() {
 	s.sentryEnabled = true
 }
 
-func (s *Server) serve() int {
+// serve installs the signal watcher, starts the listener and blocks
+// until the server's context is cancelled. The process exit status is
+// fx's to decide — from a signal, or from the code
+// shutdownOnListenFailure hands the Shutdowner — so this reports
+// nothing back to its caller.
+func (s *Server) serve() {
 	ctx, cancelFunc := context.WithCancel(context.Background())
 	s.cancelFunc = cancelFunc
 
@@ -185,7 +197,30 @@ func (s *Server) serve() int {
 	<-ctx.Done()
 	// Shutdown is handled by the fx OnStop hook (cleanShutdown).
 	// Do not call cleanShutdown() here to avoid double invocation.
-	return s.exitCode
+}
+
+// shutdownOnListenFailure ends the application after the HTTP
+// listener failed. The fx OnStart hook returns as soon as the serving
+// goroutine is spawned, so nothing downstream of it ever learns that
+// the listen failed: fx reports RUNNING and the process sits alive
+// with nothing bound, which is invisible to systemd and Docker
+// restart policies. Asking the Shutdowner to stop the app with a
+// non-zero code is what turns that into a visible failure.
+//
+// The context cancel that follows only unwinds serve()'s own wait.
+// The shutdown itself runs through fx's normal stop sequence, so the
+// clean-shutdown drain in cleanShutdown is reached unchanged.
+func (s *Server) shutdownOnListenFailure() {
+	err := s.params.Shutdowner.Shutdown(
+		fx.ExitCode(ListenFailureExitCode),
+	)
+	if err != nil {
+		s.log.Error("shutdown request failed", "error", err)
+	}
+
+	if s.cancelFunc != nil {
+		s.cancelFunc()
+	}
 }
 
 func (s *Server) cleanupForExit() {
@@ -193,9 +228,6 @@ func (s *Server) cleanupForExit() {
 }
 
 func (s *Server) cleanShutdown(ctx context.Context) {
-	// initiate clean shutdown
-	s.exitCode = 0
-
 	ctxShutdown, shutdownCancel := context.WithTimeout(
 		ctx, ShutdownTimeout,
 	)