Cut DB load: drop jobs realtime, remove dead index

CHANGELOG.md

@@ -41,6 +41,19 @@ On merge, CI will:
   previously-crawled domain ramps up from a few concurrent requests and widens
   on success. Tunable via `GNH_PACER_START_FLOOR_DELAY_MS` (`0` restores the
   prior behaviour).
+- Dashboard and Webflow job lists now refresh via adaptive polling (500 ms while
+  a job is active, 1 s when idle) instead of Supabase Realtime. `public.jobs`
+  was removed from the `supabase_realtime` publication because decoding its WAL
+  on every job-counter change was ~14% of total database time, billed
+  continuously regardless of how many dashboards were subscribed.
+  `notifications` realtime is unchanged. Reverse with
+  `ALTER PUBLICATION supabase_realtime ADD TABLE public.jobs;`.
+
+### Removed
+
+- Redundant `idx_domain_hosts_domain_id` index (0 recorded scans; the composite
+  `(domain_id, host)` unique index already serves every `domain_id`-prefix
+  lookup), removing write overhead on every `domain_hosts` upsert.
 
 ## Full changelog history
 

supabase/migrations/20260527224118_drop_dead_domain_hosts_index.sql

@@ -0,0 +1,14 @@
+-- Drop redundant index `idx_domain_hosts_domain_id`.
+--
+-- `pg_stat_user_indexes` shows 0 scans against this index, while the composite
+-- unique index `domain_hosts_domain_id_host_key (domain_id, host)` serves every
+-- `domain_id`-prefix lookup (12.6M scans). The standalone `(domain_id)` index is
+-- pure write overhead on every `domain_hosts` upsert.
+--
+-- `CONCURRENTLY` is intentionally omitted — Supabase's branch migration runner
+-- pipelines statements over the extended protocol, which Postgres forbids for
+-- `DROP INDEX CONCURRENTLY`. Dropping an index is a fast metadata operation;
+-- the brief lock on a 105k-row table is acceptable and there are no live
+-- customers yet. Guarded with `IF EXISTS` to remain idempotent across replays.
+
+DROP INDEX IF EXISTS public.idx_domain_hosts_domain_id;

supabase/migrations/20260527224119_disable_jobs_realtime.sql

@@ -0,0 +1,25 @@
+-- Remove `public.jobs` from the `supabase_realtime` publication.
+--
+-- `jobs` is the highest-churn table in the queue (per-task `running_tasks`
+-- bumps, `recalculate_job_stats`, the `update_job_queue_counters` trigger). The
+-- Realtime WAL-decode poll (`realtime.list_changes`) was ~14% of total database
+-- time, decoding every one of those changes regardless of how many dashboards
+-- were subscribed. The frontend already has an adaptive polling fallback
+-- (500ms while a job is active, 1-2s idle) reading the lane-isolated `/v1/jobs`
+-- API, so dropping `jobs` from the publication trades a continuous tenant-wide
+-- cost for cheap, viewer-scoped polling.
+--
+-- `notifications` stays on the publication. Reverse with
+-- `ALTER PUBLICATION supabase_realtime ADD TABLE public.jobs;` if needed.
+
+DO $$
+BEGIN
+    IF EXISTS (
+        SELECT 1 FROM pg_publication_tables
+        WHERE pubname = 'supabase_realtime'
+          AND schemaname = 'public'
+          AND tablename = 'jobs'
+    ) THEN
+        ALTER PUBLICATION supabase_realtime DROP TABLE public.jobs;
+    END IF;
+END $$;

web/static/app/pages/dashboard.js

@@ -100,8 +100,8 @@ async function init() {
     containerSelector: "#adminGroup",
   });
 
-  // Subscribe to realtime job updates (falls back to 10 s polling when
-  // Supabase realtime is unavailable, e.g. on preview branches).
+  // Poll for job updates (500 ms while a job is active, 1 s when idle).
+  // `jobs` is no longer on the Supabase Realtime publication.
   let unsubscribe = null;
   function startSubscription() {
     if (unsubscribe) unsubscribe();

web/static/app/pages/webflow-jobs.js

@@ -25,20 +25,15 @@ import {
   formatRelativeTime,
   formatDuration,
   formatCount,
-  formatStatus,
-  statusCategory,
 } from "/app/lib/formatters.js";
 import { createStatusPill } from "/app/components/hover-status-pill.js";
 import { createDataTable } from "/app/components/hover-data-table.js";
 
 // ── Constants ──────────────────────────────────────────────────────────────────
 
-const REALTIME_DEBOUNCE_MS = 250;
-const SUBSCRIBE_RETRY_INTERVAL_MS = 1000;
-const MAX_SUBSCRIBE_RETRIES = 15;
 // Match legacy gnh-auth-extension.js: 500 ms when jobs are active, 1 s when idle.
-const FALLBACK_POLLING_INTERVAL_ACTIVE_MS = 500;
-const FALLBACK_POLLING_INTERVAL_IDLE_MS = 1000;
+const POLLING_INTERVAL_ACTIVE_MS = 500;
+const POLLING_INTERVAL_IDLE_MS = 1000;
 
 // ── Data fetching ──────────────────────────────────────────────────────────────
 
@@ -154,152 +149,52 @@ export function renderErrorState(container, message = "Failed to load jobs.") {
   container.appendChild(div);
 }
 
-// ── Realtime subscription ──────────────────────────────────────────────────────
+// ── Adaptive polling ─────────────────────────────────────────────────────────
 
 /**
- * Subscribe to job changes for an organisation via Supabase Realtime.
- * Falls back to polling if realtime fails.
+ * Poll for job changes for an organisation.
  *
- * @param {string} orgId
- * @param {() => void} onUpdate - called when jobs may have changed
- * @returns {() => void} unsubscribe / cleanup function
+ * `jobs` was removed from the Supabase Realtime publication — decoding the WAL
+ * for every job-counter tick was ~14% of total database time, billed whether
+ * or not anyone was watching. We poll the lane-isolated `/v1/jobs` API instead:
+ * 500 ms while a job is active, 1 s when idle. The interval is re-evaluated on
+ * every tick so it adapts as jobs start and finish.
+ *
+ * @param {string} _orgId - retained for call-site compatibility
+ * @param {() => (void | Promise<void>)} onUpdate - called when jobs may have changed
+ * @returns {() => void} cleanup function
  */
-export function subscribeToJobUpdates(orgId, onUpdate) {
-  let channel = null;
-  let retryCount = 0;
-  let retryTimer = null;
-  let fallbackTimer = null;
-  let lastUpdate = 0;
-  let debounceTimer = null;
-  let unsubscribed = false;
-
-  function throttledUpdate() {
-    const now = Date.now();
-    if (now - lastUpdate >= REALTIME_DEBOUNCE_MS) {
-      lastUpdate = now;
-      clearFallback();
-      onUpdate();
-      return;
-    }
-    if (!debounceTimer) {
-      debounceTimer = setTimeout(() => {
-        debounceTimer = null;
-        if (unsubscribed) return;
-        lastUpdate = Date.now();
-        clearFallback();
-        onUpdate();
-      }, REALTIME_DEBOUNCE_MS);
-    }
-  }
+export function subscribeToJobUpdates(_orgId, onUpdate) {
+  let timer = null;
+  let stopped = false;
 
-  // Adaptive interval: 500 ms while jobs are active, 1 s when idle.
-  // Matches the legacy gnh-auth-extension.js dual-interval behaviour.
-  function getFallbackInterval() {
+  function intervalMs() {
     return window.dataBinder?.hasRealtimeActiveJobs
-      ? FALLBACK_POLLING_INTERVAL_ACTIVE_MS
-      : FALLBACK_POLLING_INTERVAL_IDLE_MS;
-  }
-
-  let fallbackIntervalMs = null;
-
-  function startFallback() {
-    const nextMs = getFallbackInterval();
-    if (fallbackTimer && fallbackIntervalMs === nextMs) return;
-    if (fallbackTimer) {
-      clearInterval(fallbackTimer);
-    }
-    fallbackIntervalMs = nextMs;
-    fallbackTimer = setInterval(onUpdate, fallbackIntervalMs);
-  }
-
-  function clearFallback() {
-    if (fallbackTimer) {
-      clearInterval(fallbackTimer);
-      fallbackTimer = null;
-      fallbackIntervalMs = null;
-    }
-  }
-
-  function cleanup() {
-    unsubscribed = true;
-    if (retryTimer) {
-      clearTimeout(retryTimer);
-      retryTimer = null;
-    }
-    if (debounceTimer) {
-      clearTimeout(debounceTimer);
-      debounceTimer = null;
-    }
-    clearFallback();
-    if (channel && window.supabase) {
-      window.supabase.removeChannel(channel).catch(() => {});
-      channel = null;
-    }
+      ? POLLING_INTERVAL_ACTIVE_MS
+      : POLLING_INTERVAL_IDLE_MS;
   }
 
-  function subscribe() {
-    if (unsubscribed) return;
-    if (!orgId || !window.supabase?.channel) {
-      if (retryCount < MAX_SUBSCRIBE_RETRIES) {
-        retryCount++;
-        retryTimer = setTimeout(subscribe, SUBSCRIBE_RETRY_INTERVAL_MS);
-      } else {
-        startFallback();
+  function schedule() {
+    if (stopped) return;
+    timer = setTimeout(async () => {
+      if (stopped) return;
+      // Await so a slow refresh can't stack overlapping polls; reschedule in
+      // finally so a rejected update still keeps the loop alive.
+      try {
+        await onUpdate();
+      } finally {
+        schedule();
       }
-      return;
-    }
-
-    retryCount = 0;
+    }, intervalMs());
+  }
 
-    try {
-      channel = window.supabase
-        .channel(`hover-jobs:${orgId}`)
-        .on(
-          "postgres_changes",
-          {
-            event: "INSERT",
-            schema: "public",
-            table: "jobs",
-            filter: `organisation_id=eq.${orgId}`,
-          },
-          throttledUpdate
-        )
-        .on(
-          "postgres_changes",
-          {
-            event: "UPDATE",
-            schema: "public",
-            table: "jobs",
-            filter: `organisation_id=eq.${orgId}`,
-          },
-          throttledUpdate
-        )
-        .on(
-          "postgres_changes",
-          {
-            event: "DELETE",
-            schema: "public",
-            table: "jobs",
-            filter: `organisation_id=eq.${orgId}`,
-          },
-          throttledUpdate
-        )
-        .subscribe((status, err) => {
-          if (
-            (status === "CHANNEL_ERROR" || status === "TIMED_OUT" || err) &&
-            !unsubscribed
-          ) {
-            startFallback();
-          }
-        });
+  schedule();
 
-      // Start fallback immediately; clearFallback() stops it on first real event
-      startFallback();
-    } catch {
-      startFallback();
+  return function cleanup() {
+    stopped = true;
+    if (timer) {
+      clearTimeout(timer);
+      timer = null;
     }
-  }
-
-  subscribe();
-  return cleanup;
+  };
 }