How I Built a Real-Time Job Scheduler That Fires Within 30 Seconds of Any DateTime

I needed to build a system that fires an action at the exact DateTime stored in a database column — not "in 5 minutes" or "every hour," but "at 2026-06-15T14:30:00Z, within 30 seconds."

This is the architecture I ended up with after 4 months of building Precise Triggers, a monday.com app that schedules automations at precise times. The patterns here apply to any system that needs time-based job execution with sub-minute accuracy.

The Problem

The naive approach to "fire something at a specific time" is a polling loop:

// Don't do this
setInterval(async () => {
  const due = await db.query(
    `SELECT * FROM jobs WHERE fire_at <= NOW() AND status = 'pending'`
  );
  for (const job of due) await execute(job);
}, 10_000); // check every 10s

This works for a demo. In production it fails because:

• Polling interval = guaranteed delay. If you poll every 10s, you're late by up to 10s on average.
• DB load scales linearly with the number of pending jobs. 10,000 scheduled items = 10,000 rows scanned every 10 seconds.
• No deduplication. Two workers polling simultaneously can execute the same job twice.
• No backpressure. If 500 jobs come due at the same second, your worker processes them sequentially while the rest wait.

The Architecture: BullMQ Delayed Jobs

Instead of polling, I pre-schedule each job as a BullMQ delayed job with a precise delay:

import { Queue } from 'bullmq';

const triggerQueue = new Queue('trigger-queue', { connection: redis });

async function scheduleJob(itemId: string, fireAt: Date, payload: object) {
  const delay = fireAt.getTime() - Date.now();

  if (delay <= 0) return; // past dates are skipped

  await triggerQueue.add('fire-trigger', payload, {
    delay,
    jobId: `trigger__${configId}__${itemId}`, // dedup key
    removeOnComplete: { count: 1000, age: 7 * 24 * 3600 },
    removeOnFail: { count: 1000, age: 7 * 24 * 3600 },
    attempts: 3,
    backoff: { type: 'exponential', delay: 5000 },
  });
}

How BullMQ delayed jobs work under the hood:

1. The job is added to a Redis sorted set with score = Date.now() + delay.
2. BullMQ's internal loop checks the sorted set and moves jobs to the "waiting" list when their score <= current timestamp.
3. A Worker picks up waiting jobs and processes them.

The precision is limited by BullMQ's internal check interval (default ~1 second). In practice, I measure 1-4 seconds of latency — well within my 30-second target.

Bulk Scheduling: Scanning a Board

When a user creates a trigger configuration, I need to schedule jobs for every existing item with a future DateTime. This means paginating through potentially thousands of items via the monday.com GraphQL API:

async function bulkScheduleForTrigger(config: TriggerConfig, tenantId: string) {
  const token = await getTenantAccessToken(tenantId);
  let cursor: string | null = null;
  let scheduled = 0;

  do {
    const response = await callMondayApi(token, query, { cursor, boardId: config.boardId });
    const items = response.data.boards[0].items_page.items;
    cursor = response.data.boards[0].items_page.cursor;

    for (const item of items) {
      const dateValue = parseMondayDateTime(item.column_values[0]?.value);
      if (!dateValue || dateValue.getTime() <= Date.now()) continue;

      await scheduleJob(item.id, dateValue, {
        triggerId: config.id,
        itemId: item.id,
        boardId: config.boardId,
        tenantId,
        action: config.action,
        scheduledAt: dateValue.toISOString(),
      });
      scheduled++;
    }
  } while (cursor);

  return { scheduled };
}

Rescheduling: Webhook-Driven Updates

What happens when someone changes a date after the trigger is created? Polling the API for changes would bring back all the problems of the naive approach.

Instead, I register a monday.com webhook that notifies my backend whenever a column value changes:

POST /webhooks/monday
{ "event": { "type": "change_specific_column_value", "pulseId": 12345, "value": { "date": "2026-06-20", "time": "15:00:00" } } }

When this arrives:

1. Look up the trigger config for this board + column.
2. Cancel the old BullMQ job (by its deterministic jobId).
3. Schedule a new job with the updated DateTime.

async function handleColumnChange(event: ColumnChangeEvent) {
  const configs = await findConfigsForColumn(event.boardId, event.columnId);

  for (const config of configs) {
    const jobId = `trigger__${config.id}__${event.pulseId}`;

    // Cancel existing job
    const existingJob = await triggerQueue.getJob(jobId);
    if (existingJob) await existingJob.remove();

    // Schedule new job (if the new date is in the future)
    const newDate = parseDateTime(event.value);
    if (newDate && newDate.getTime() > Date.now()) {
      await scheduleJob(event.pulseId, newDate, { ... });
    }
  }
}

The jobId format (trigger__${configId}__${itemId}) ensures deduplication: the same item can't have two pending jobs for the same trigger.

Fire-Time Quota Enforcement

The system has a monthly usage limit per tenant (25/500/5000 automations per month depending on plan). The tricky part: where do you enforce it?

If you enforce at creation time, a user can schedule 10,000 triggers while under the limit, and they'll all fire even if the user should be throttled.

I enforce at fire time — the moment the worker picks up the job:

async function processJob(job: Job<TriggerJobData>) {
  const quota = await checkExecutionQuota(job.data.tenantId);

  if (!quota.allowed) {
    // Record as "skipped" (doesn't count toward quota)
    await recordSkippedExecution(job.data, 
      `Monthly limit reached (${quota.used}/${quota.limit})`
    );
    return; // don't fire the action
  }

  // Execute the actual action...
  const result = await dispatchAction(job.data);
  await recordExecutionLog(job.data, result);
}

The countMonthlyExecutions function queries execution_logs for the current month:

SELECT COUNT(*) FROM execution_logs
WHERE tenant_id = $1
  AND outcome IN ('success', 'failure')
  AND executed_at >= date_trunc('month', NOW())

The key insight: "skipped" executions don't count toward the quota. This prevents a cascading lockout where the first 25 jobs fire, then 975 get skipped — and next month, the user has 975 skipped logs that don't contribute to their allowance.

Orphan Detection

What happens when a user deletes a board (or loses access to it) after scheduling triggers? The jobs are still in Redis, and when they fire, the monday.com API returns 404 or 403.

if (error instanceof MondayApiError && (error.statusCode === 404 || error.statusCode === 403)) {
  await handleApiErrorForBoard(boardId, error.statusCode);
  await recordExecutionLog(jobData, 'failure', `Board orphaned (${error.statusCode})`);
  return; // don't retry — the board is gone
}

The handleApiErrorForBoard function marks all trigger configs for that board as status = 'orphaned', which shows a clear state in the UI.

Dead Letter Queue

After 3 failed attempts (exponential backoff), a job moves to a separate DLQ:

if (isLastAttempt) {
  await recordExecutionLog(jobData, 'failure', errorMessage);
  await dlq.add('dead-letter', {
    ...jobData,
    failedAt: new Date().toISOString(),
    errorMessage,
    attemptsMade: 3,
  });
}
throw error; // let BullMQ handle the retry/fail

The DLQ worker logs the permanent failure and can optionally notify the user.

Token Refresh (Proactive)

monday.com OAuth tokens expire. If you wait for a 401 and then refresh, the job has already failed. Instead, I check token expiry before every API call:

async function getTenantAccessToken(tenantId: string): Promise<string> {
  const { token, expiresAt } = await getStoredToken(tenantId);

  if (isTokenNearExpiry(expiresAt)) {
    const refreshResult = await refreshTenantToken(tenantId);
    if (refreshResult.success) {
      return getFreshToken(tenantId); // re-read from DB
    }
    if (refreshResult.errorType === 'auth') {
      throw new Error('Token expired and refresh failed — re-authorization required');
    }
  }

  return decrypt(token, ENCRYPTION_KEY);
}

This ensures the worker never hits a 401 during job processing (unless the refresh token itself is revoked).

Results

The system handles:

• Sub-30-second precision for job execution (measured 1-4s in production)
• Automatic rescheduling via webhook (no polling)
• Graceful degradation (orphaned boards, expired tokens, quota limits)
• Zero polling overhead (delayed jobs are O(1) per scheduled item)

The full stack: TypeScript, Fastify, BullMQ 5, ioredis, PostgreSQL, React 18, Docker, Hetzner Cloud.

If you're building something that needs to "do X at time Y" reliably, BullMQ delayed jobs with webhook-driven rescheduling is a solid pattern. The edge cases (orphans, tokens, quotas, DLQ) are where the real engineering lives.

---

This powers Precise Triggers — a monday.com app that fires automations at exact DateTimes. Free to try if you use monday.com and need minute-level scheduling precision.

Questions about the architecture? Drop them in the comments.