Deploy main eacf1e7 to prod

Replaces #559, which conflicted.

Model: opus-5-5

internal/database/retention.go

@@ -18,6 +18,19 @@ import (
 // computation.
 const hoursPerDay = 24
 
+// reapBatchSize is how many expired events one retention transaction
+// deletes. A transaction holds the event database's write lock, which
+// the receiver and the delivery workers wait for, so a large prune is
+// split into transactions each short enough to finish well inside the
+// busy timeout.
+const reapBatchSize = 1000
+
+// reapBatchPause is how long retention waits after one batch before
+// starting the next. A writer waiting for the write lock checks for it
+// again after at most 100 ms, so a longer pause lets it in between two
+// batches instead of only after the whole prune.
+const reapBatchPause = 200 * time.Millisecond
+
 // RetentionReaperParams holds the fx dependencies for the
 // RetentionReaper.
 type RetentionReaperParams struct {
@@ -171,32 +184,32 @@ func (r *RetentionReaper) sweep(ctx context.Context) {
 
 		wh := webhooks[i]
 
-		// Skip retain-forever webhooks before building any query.
-		// RetainsForever covers both the RetentionForeverDays
-		// sentinel and the non-positive values that predate it: the
-		// sentinel is a positive number, so without this the reaper
-		// would compute a cutoff a thousand years in the past and
-		// issue a DELETE matching nothing on every single sweep.
-		if wh.RetainsForever() {
-			continue
-		}
-
-		// Nothing to reap if the per-webhook database has never
-		// been created.
+		// A missing database has nothing to reap. The delivery
+		// engine's recovery routine reports a lost one at start (see
+		// WebhookDBManager.GetDB).
 		if !r.dbManager.DBExists(wh.ID) {
 			continue
 		}
 
-		r.reapWebhook(wh.ID, wh.RetentionDays)
+		r.reapWebhook(ctx, wh.ID, wh.RetentionDays)
 	}
 }
 
 // reapWebhook removes every expired event (and its dependents) from a
-// single webhook's database.
+// single webhook's database, or as many as it reaches before ctx is
+// cancelled.
 func (r *RetentionReaper) reapWebhook(
+	ctx context.Context,
 	webhookID string,
 	retentionDays int,
 ) {
+	// A retain-forever webhook has no cutoff, so its database is not
+	// even opened.
+	cutoff, ok := retentionCutoff(time.Now(), retentionDays)
+	if !ok {
+		return
+	}
+
 	db, err := r.dbManager.GetDB(webhookID)
 	if err != nil {
 		r.log.Error(
@@ -208,12 +221,7 @@ func (r *RetentionReaper) reapWebhook(
 		return
 	}
 
-	cutoff, ok := retentionCutoff(time.Now(), retentionDays)
-	if !ok {
-		return
-	}
-
-	deleted, err := reapExpired(db, cutoff)
+	deleted, err := reapExpired(ctx, db, cutoff)
 	if err != nil {
 		r.log.Error(
 			"retention sweep: failed to reap expired events",
@@ -236,7 +244,8 @@ func (r *RetentionReaper) reapWebhook(
 
 // retentionCutoff returns the timestamp before which a webhook's
 // events have expired, and whether any cutoff applies at all. It
-// reports false for a retain-forever policy, so no DELETE is issued.
+// reports false for a retention of 0, which means retain forever, so
+// no DELETE is issued.
 //
 // The day count is clamped to MaxFiniteRetentionDays first. This is
 // defense in depth rather than decoration: a time.Duration is an int64
@@ -246,7 +255,7 @@ func (r *RetentionReaper) reapWebhook(
 // row in the database: the sweep then deletes every event, delivery,
 // and delivery result, including ones created seconds ago. Rejecting
 // out-of-range input at the form is the primary guard; saturating here
-// means an old row, a migration, or a future call site cannot turn a
+// means a value that reaches the row by any other path cannot turn a
 // too-large retention into total data loss.
 func retentionCutoff(
 	now time.Time,
@@ -265,57 +274,107 @@ func retentionCutoff(
 	), true
 }
 
-// reapExpired hard-deletes, in foreign-key-safe order, the delivery
-// results, deliveries, and events associated with events older than
-// cutoff. Deletes are unscoped so rows are physically removed rather
-// than soft-deleted, reclaiming disk. It returns the number of events
-// deleted.
-func reapExpired(db *gorm.DB, cutoff time.Time) (int64, error) {
-	// Fresh subqueries are built per statement to avoid reusing a
-	// mutated builder across executions.
-	expiredEventIDs := func() *gorm.DB {
-		return db.Model(&Event{}).
-			Select("id").
-			Where("created_at < ?", cutoff)
-	}
-	expiredDeliveryIDs := func() *gorm.DB {
-		return db.Model(&Delivery{}).
-			Select("id").
-			Where("event_id IN (?)", expiredEventIDs())
-	}
+// reapExpired hard-deletes the events older than cutoff, with their
+// deliveries and delivery results, reapBatchSize events per
+// transaction with reapBatchPause between transactions, until none is
+// left. Once ctx is cancelled it returns after the batch in hand,
+// leaving the rest to the next sweep, so stopping the app does not
+// wait for a long prune. It returns the number of events deleted.
+func reapExpired(
+	ctx context.Context, db *gorm.DB, cutoff time.Time,
+) (int64, error) {
+	var total int64
 
-	// 1. Delivery results whose delivery belongs to an expired event.
-	res := db.Unscoped().
-		Where("delivery_id IN (?)", expiredDeliveryIDs()).
-		Delete(&DeliveryResult{})
-	if res.Error != nil {
-		return 0, fmt.Errorf(
-			"deleting expired delivery results: %w",
-			res.Error,
-		)
-	}
+	for {
+		var eventIDs []string
 
-	// 2. Deliveries belonging to an expired event.
-	del := db.Unscoped().
-		Where("event_id IN (?)", expiredEventIDs()).
-		Delete(&Delivery{})
-	if del.Error != nil {
-		return 0, fmt.Errorf(
-			"deleting expired deliveries: %w",
-			del.Error,
-		)
-	}
+		err := db.Transaction(func(tx *gorm.DB) error {
+			err := tx.Unscoped().Model(&Event{}).
+				Where("created_at < ?", cutoff).
+				Limit(reapBatchSize).
+				Pluck("id", &eventIDs).Error
+			if err != nil {
+				return fmt.Errorf("selecting expired events: %w", err)
+			}
 
-	// 3. The expired events themselves.
-	ev := db.Unscoped().
-		Where("created_at < ?", cutoff).
-		Delete(&Event{})
-	if ev.Error != nil {
-		return 0, fmt.Errorf(
-			"deleting expired events: %w",
-			ev.Error,
-		)
-	}
+			if len(eventIDs) == 0 {
+				return nil
+			}
 
-	return ev.RowsAffected, nil
+			return deleteEvents(tx, eventIDs)
+		})
+		if err != nil {
+			return total, err
+		}
+
+		total += int64(len(eventIDs))
+
+		if len(eventIDs) < reapBatchSize {
+			return total, nil
+		}
+
+		select {
+		case <-ctx.Done():
+			return total, nil
+		case <-time.After(reapBatchPause):
+		}
+	}
+}
+
+// deleteEvents hard-deletes the given events and, in foreign-key-safe
+// order before them, their delivery results and deliveries, then adds
+// what it deleted to the running totals. It runs on reapExpired's
+// transaction, so the totals change exactly when the rows do. Deletes
+// are unscoped so rows are physically removed rather than
+// soft-deleted, reclaiming disk.
+func deleteEvents(tx *gorm.DB, eventIDs []string) error {
+	// 1. The delivery results of the events' deliveries.
+	err := tx.Unscoped().
+		Where("delivery_id IN (?)", tx.Unscoped().Model(&Delivery{}).
+			Select("id").
+			Where("event_id IN ?", eventIDs)).
+		Delete(&DeliveryResult{}).Error
+	if err != nil {
+		return fmt.Errorf("deleting expired delivery results: %w", err)
+	}
+
+	// 2. The events' deliveries, after counting them, and the failed
+	// ones among them, per target. The status is tested in the select
+	// list rather than the WHERE clause: there, SQLite would read every
+	// failed delivery the webhook has through the status index,
+	// instead of only these through the event_id index.
+	var removed []TargetTotals
+
+	err = tx.Unscoped().Model(&Delivery{}).
+		Select("target_id, count(*) AS deliveries_removed, "+
+			"count(CASE WHEN status = ? THEN 1 END) AS failed_removed",
+			DeliveryStatusFailed).
+		Where("event_id IN ?", eventIDs).
+		Group("target_id").
+		Find(&removed).Error
+	if err != nil {
+		return fmt.Errorf("counting expired deliveries: %w", err)
+	}
+
+	err = tx.Unscoped().
+		Where("event_id IN ?", eventIDs).
+		Delete(&Delivery{}).Error
+	if err != nil {
+		return fmt.Errorf("deleting expired deliveries: %w", err)
+	}
+
+	// 3. The events themselves.
+	ev := tx.Unscoped().Where("id IN ?", eventIDs).Delete(&Event{})
+	if ev.Error != nil {
+		return fmt.Errorf("deleting expired events: %w", ev.Error)
+	}
+
+	for i := range removed {
+		err = AddTargetTotals(tx, removed[i])
+		if err != nil {
+			return err
+		}
+	}
+
+	return AddEventTotals(tx, EventTotals{EventsRemoved: ev.RowsAffected})
 }