scheduler-and-jobs

Scheduler & Background Jobs (Production)

Reference for making Frappe background jobs and scheduled tasks robust under real-world load. The server-scripts skill covers basic frappe.enqueue and scheduler_events syntax; this is its production-grade companion.

Queue selection

Frappe ships three RQ queues with different timeout defaults:

Queue	Default timeout	Use for
short	60 seconds	Quick fire-and-forget: send one email, update one cache key, recompute one denormalized field
default	300 seconds (5 min)	Most normal jobs: process one document, sync a small batch
long	1500 seconds (25 min)	Bulk imports, monthly reports, heavy data migrations

frappe.enqueue("my_app.tasks.send_one_email", queue="short", recipient="x@y.com")
frappe.enqueue("my_app.tasks.process_invoice", queue="default", invoice="SINV-001")
frappe.enqueue("my_app.tasks.monthly_aging_report", queue="long")

Mismatch consequences:

• Job on short that runs >60s: killed mid-execution with no rollback. Half-written state persists.
• Job on long that takes 5s: hogs a long-queue worker slot another truly-long job needs. Smaller blast radius.

When in doubt, set an explicit timeout and use default:

frappe.enqueue(
    "my_app.tasks.process_invoice",
    queue="default",
    timeout=180,
    invoice="SINV-001",
)

now=True, enqueue_after_commit, is_async

frappe.enqueue(method, queue="default", now=False, enqueue_after_commit=False, is_async=True, **kwargs)

Flag	What it does	When to use
now=True	Runs the function in the current process synchronously, no Redis	Tests only. In production it defeats the purpose of enqueueing — the request thread blocks.
enqueue_after_commit=True	Defers the enqueue until the current DB transaction commits	When you enqueue from a validate / before_save hook on a doc that's about to be saved. Without this flag, the worker may pick up the job and try to read the doc before the request thread commits — race condition.
is_async=False	Runs in the current process (similar to now=True)	Almost never. Use now=True for tests; otherwise let it queue.

The enqueue_after_commit rule: if you call frappe.enqueue from any document lifecycle hook (validate, before_save, on_update, on_submit, on_cancel) and the job needs the doc to exist or have its latest state, set enqueue_after_commit=True. Otherwise the job races the commit.

class SalesInvoice(Document):
    def on_submit(self):
        # Worker needs to see the submitted invoice — wait for commit
        frappe.enqueue(
            "my_app.tasks.send_invoice_email",
            queue="short",
            timeout=60,
            enqueue_after_commit=True,
            invoice_name=self.name,
        )

Cron expressions in scheduler_events

The team convention is to prefer cron: over the named keys (hourly, daily, weekly, monthly) — explicit cron expressions are more obvious to read and grep for. The named keys are still fine for "I genuinely want this every hour" but cron is preferred when the cadence is anything else or when you want to see the schedule at a glance.

# hooks.py — pattern from edu_quality
scheduler_events = {
    "all": [
        "edu_quality.api.student_application.get_and_schedule_pending_walkouts",
        "edu_quality.overrides_hooks.item.upload_all_imported_to_drive",
    ],
    "cron": {
        "0 * * * *":    ["edu_quality.tasks.cron"],
        "0 19 * * *":   ["edu_quality.tasks.send_bulk_notification_cmap_to_guardian"],
        "0 1 * * *":    ["edu_quality.edu_quality.scheduled_tasks.assessment_pdf_scheduler.attach_daily_assessment_pdfs"],
        "0 6 * * *":    [
            "edu_quality.tasks.schedule_birthday_greeting",
            "edu_quality.tasks.event_reminder",
        ],
        "* * * * *":    [
            "edu_quality.cmap_jobs.send_ptm_notifications_to_students",
            "edu_quality.cmap_jobs.notify_teacher_before_one_hour_job",
        ],
    },
}

Common patterns:

Cron	Meaning
*/15 * * * *	Every 15 minutes
0 * * * *	Top of every hour
0 9 * * 1-5	9 AM Monday–Friday
0 0 * * 0	Midnight Sunday
0 0 1 * *	Midnight on the 1st of every month
*/30 9-17 * * 1-5	Every 30 minutes, 9 AM–5 PM, weekdays

Scheduler runs in the default queue. For long scheduled jobs, the scheduler entry point should frappe.enqueue to the long queue rather than blocking the scheduler thread — this is the team's standard two-step pattern (see assessment_pdf_scheduler.py):

def attach_daily_assessment_pdfs():
    """Scheduler entry point — fast handoff."""
    frappe.enqueue(
        _process_daily_assessment_pdfs,
        queue="long",
        timeout=3600,  # 1 hour
        job_name="daily_assessment_pdf_attachment",
    )
    return {"success": True, "message": "Daily Assessment PDF attachment job has been queued"}


def _process_daily_assessment_pdfs():
    # The actual work
    ...

Note the deterministic job_name — running the scheduler twice in the same window is a no-op because frappe.enqueue dedups jobs with the same name in the queue.

Common patterns:

Cron	Meaning
*/15 * * * *	Every 15 minutes
0 * * * *	Top of every hour
0 9 * * 1-5	9 AM Monday–Friday
0 0 * * 0	Midnight Sunday
0 0 1 * *	Midnight on the 1st of every month
*/30 9-17 * * 1-5	Every 30 minutes, 9 AM–5 PM, weekdays

Scheduler runs in the default queue. For long scheduled jobs, the scheduler invocation should frappe.enqueue to the long queue rather than blocking the scheduler thread:

def daily_cleanup():
    """Scheduler entry point — fast: just enqueue the real work."""
    frappe.enqueue("my_app.tasks._do_cleanup", queue="long", timeout=1800)

def _do_cleanup():
    # The actual work
    ...

Idempotency

A job is idempotent if running it twice produces the same end state as running it once. Workers crash, retries happen, scheduler windows can overlap — assume your job WILL run twice.

Pattern 1: check before mutate

def sync_customer_to_external(customer_id):
    # Has this customer already been synced today?
    today = frappe.utils.nowdate()
    if frappe.db.exists("External Sync Log", {
        "customer": customer_id,
        "sync_date": today,
        "status": "Success",
    }):
        return  # Already done — exit cleanly

    # ... do the sync ...

    frappe.get_doc({
        "doctype": "External Sync Log",
        "customer": customer_id,
        "sync_date": today,
        "status": "Success",
    }).insert(ignore_permissions=True)
    frappe.db.commit()

Pattern 2: dedup via job_name

frappe.enqueue accepts a job_name. If a job with the same name is already in the queue or running, the new enqueue is dropped (returns None):

frappe.enqueue(
    "my_app.tasks.recompute_customer_balance",
    queue="short",
    job_name=f"recompute-balance-{customer_id}",
    customer=customer_id,
)

Use stable, deterministic job names. Random suffixes defeat the dedup.

Pattern 3: optimistic update with version check

def update_inventory(item_code, expected_qty, new_qty):
    rows = frappe.db.sql(
        """UPDATE `tabBin`
           SET actual_qty = %s
           WHERE item_code = %s AND actual_qty = %s""",
        (new_qty, item_code, expected_qty),
    )
    if not rows:
        # Someone else updated it first; this run is stale
        return

Distributed locks

When two workers might run the same critical section concurrently (e.g. two webhooks triggering the same recalculation), use Frappe's cache as a lock:

def with_lock(key, ttl=300):
    """Poor man's lock via Redis cache. Returns True if acquired."""
    cache = frappe.cache()
    if cache.get_value(key):
        return False
    cache.set_value(key, frappe.session.user or "unknown", expires_in_sec=ttl)
    return True

def release_lock(key):
    frappe.cache().delete_value(key)


def recompute_customer_balance(customer_id):
    lock_key = f"recompute-balance:{customer_id}"
    if not with_lock(lock_key, ttl=120):
        return  # another worker is doing it
    try:
        # ... heavy recompute ...
        frappe.db.commit()
    finally:
        release_lock(lock_key)

This is not atomic — there's a tiny race between get_value and set_value. For true atomicity, use Redis SETNX directly:

def acquire_lock(key, ttl=300):
    redis = frappe.cache()
    # SETNX returns 1 if set, 0 if already exists
    acquired = redis.execute_command("SET", key, "1", "NX", "EX", ttl)
    return acquired is not None

For most ERPNext workloads the cache-based lock is fine. Reserve SETNX for jobs that handle money or stock where double-execution matters.

Retries

Frappe does not auto-retry failed jobs. The standard pattern: catch transient errors and re-enqueue with backoff.

def sync_with_retry(account_id, attempt=1, max_attempts=3):
    try:
        result = call_external_api(account_id)
        save_result(account_id, result)
        frappe.db.commit()
    except (ConnectionError, TimeoutError) as e:
        if attempt >= max_attempts:
            frappe.log_error(
                title=f"Sync failed after {max_attempts} attempts: {account_id}",
                message=frappe.get_traceback(),
            )
            raise

        # Exponential backoff: 1m, 2m, 4m, ...
        delay_minutes = 2 ** (attempt - 1)
        frappe.enqueue(
            "my_app.tasks.sync_with_retry",
            queue="default",
            timeout=300,
            enqueue_after_commit=True,
            account_id=account_id,
            attempt=attempt + 1,
            max_attempts=max_attempts,
            # Note: Frappe doesn't have native delayed enqueue; you'd need
            # to schedule via cron or use the scheduler hook for true delay.
        )

For true delayed retries (rather than re-enqueueing immediately), use a scheduled job that polls a "Retry Queue" DocType, or upgrade to a custom RQ scheduler.

Monitoring

# Pending jobs in all queues for a site
bench --site mysite show-pending-jobs

# Scheduler health (jobs per minute, last run, errors)
bench --site mysite scheduler-events
bench --site mysite doctor

# Restart workers (clears any stuck job state in worker memory)
bench restart

# Watch worker output live (bench start mode)
tail -F sites/mysite/logs/worker.log
tail -F sites/mysite/logs/scheduler.log

The frappe-debug skill in this plugin starts the bench-error-log monitor, which surfaces worker errors live during a debug session.

For deeper introspection, RQ has a CLI:

# Inside bench env
bench --site mysite execute "frappe.utils.background_jobs.get_workers"

# Or via rq-info if installed
rq info -u redis://localhost:11000

Debugging stuck jobs

Symptoms: a job is queued but never finishes, or worker process is alive but doing nothing.

1. Check pending-jobs — does the job appear?

   bench --site mysite show-pending-jobs
   

2. Check worker logs for the job's start line — did it begin?

   grep "<job_id>" sites/mysite/logs/worker.log
   

3. Worker process state — is it CPU-bound or stuck on I/O?

   ps aux | grep rq
   strace -p <pid>          # what syscall is it blocked on?
   py-spy dump --pid <pid>  # what Python frame is it in? (best tool)
   

4. Cancel and restart:

   bench --site mysite cancel-running-jobs
   bench restart-workers
   

5. Common causes:

   • Job acquired a lock and crashed before releasing → next run blocks forever
   • Job did frappe.db.commit() mid-loop with no rollback handling, leaving DB in odd state
   • Job is waiting on an external API with no timeout
   • Worker memory leak from a long-running job that loaded too much into memory; restart workers

Memory leaks

Workers don't restart between jobs. A job that leaks (e.g. holds a reference to a 100MB result set) leaves that memory allocated for the next job.

Mitigations:

• For known-heavy jobs, set RQ to recycle the worker after N jobs (bench config set worker.max_jobs_per_worker 100).
• Restart workers daily via cron: bench restart-workers.
• Avoid global caches in worker code; use frappe.cache() (Redis) instead.

Long-running jobs: chunking + progress

Jobs >5 minutes should chunk and publish progress so the user (or admin) can see they're alive:

def reindex_all_customers():
    customers = frappe.get_all("Customer", pluck="name", limit_page_length=0)
    total = len(customers)
    chunk_size = 200
    user = frappe.session.user

    for i in range(0, total, chunk_size):
        chunk = customers[i:i + chunk_size]
        for customer in chunk:
            try:
                _reindex_one(customer)
            except Exception:
                frappe.log_error(title=f"reindex failed: {customer}")
                # Continue with the rest — don't let one bad row kill the run

        frappe.db.commit()  # commit each chunk; partial progress is preserved on crash

        frappe.publish_realtime(
            "reindex_progress",
            {"current": i + len(chunk), "total": total},
            user=user,
        )

    frappe.publish_realtime("reindex_done", {"total": total}, user=user)

Why commit per chunk: if the worker dies on chunk 8 of 50, you lose chunk 8 only — chunks 1–7 are durably written. Without per-chunk commits, the entire run rolls back.

Common anti-patterns

Anti-pattern	Why it bites
Job mutates frappe.local.user and doesn't restore	Next job in the worker runs as that user
Job sets frappe.flags.ignore_permissions = True and doesn't reset	Subsequent jobs bypass perms unintentionally
Job depends on frappe.session.user	session.user is the user who enqueued — usually not what you want for scheduled jobs
Job calls frappe.db.commit() from inside a validate() controller hook (because it called enqueue(now=True))	Breaks the calling request's transaction
Job has no frappe.db.commit() at the end	Changes lost on worker restart
Job catches all exceptions and silently swallows	Real failures invisible until the error log fills
Scheduled job enqueues itself with now=False	Self-enqueueing loops can stack; use now=True only if synchronous is intended
Cron * * * * * (every minute) doing real work	Use the scheduler_events: "all" slot or queue to default

Example: robust nightly sync

A nightly job that pulls from an external API, processes per-customer, is idempotent, retries on transient failures, locks against concurrent runs, and reports progress:

# my_app/hooks.py
scheduler_events = {
    "cron": {
        "0 2 * * *": ["my_app.tasks.nightly_sync_kickoff"],
    },
}

# my_app/tasks.py
import frappe
from frappe.utils import nowdate, add_to_date


def nightly_sync_kickoff():
    """Scheduler entry — fast handoff to the long queue."""
    frappe.enqueue(
        "my_app.tasks._nightly_sync",
        queue="long",
        timeout=3600,
        job_name=f"nightly-sync-{nowdate()}",  # dedup on the date
    )


def _nightly_sync():
    lock_key = f"nightly-sync:{nowdate()}"

    # Acquire run lock — prevents overlap if yesterday's run is still going
    if not frappe.cache().execute_command("SET", lock_key, "1", "NX", "EX", 7200):
        frappe.log_error(title="nightly_sync skipped: previous run still active")
        return

    try:
        accounts = frappe.get_all(
            "External Account",
            filters={"sync_enabled": 1},
            pluck="name",
        )
        total = len(accounts)
        succeeded = failed = 0

        for i, account_id in enumerate(accounts):
            # Idempotency: skip if already synced today
            today_log = frappe.db.exists("External Sync Log", {
                "account": account_id,
                "sync_date": nowdate(),
                "status": "Success",
            })
            if today_log:
                succeeded += 1
                continue

            try:
                _sync_one_account(account_id)
                _log_sync(account_id, status="Success")
                succeeded += 1
            except (ConnectionError, TimeoutError) as e:
                # Transient — schedule a retry tomorrow morning
                _log_sync(account_id, status="Retry", error=str(e))
                _enqueue_retry(account_id)
                failed += 1
            except Exception:
                # Permanent — log and move on
                frappe.log_error(
                    title=f"nightly_sync permanent failure: {account_id}",
                    message=frappe.get_traceback(),
                )
                _log_sync(account_id, status="Failed", error=frappe.get_traceback())
                failed += 1

            # Commit per-account so partial progress is durable
            frappe.db.commit()

            if (i + 1) % 50 == 0:
                # Periodic progress log
                frappe.logger().info(f"nightly_sync: {i + 1}/{total} processed")

        frappe.logger().info(
            f"nightly_sync done: {succeeded} succeeded, {failed} failed of {total}"
        )

    finally:
        frappe.cache().delete_value(lock_key)


def _sync_one_account(account_id):
    # ... actual API call + DB writes ...
    pass


def _log_sync(account_id, status, error=None):
    frappe.get_doc({
        "doctype": "External Sync Log",
        "account": account_id,
        "sync_date": nowdate(),
        "status": status,
        "error_message": error,
    }).insert(ignore_permissions=True)


def _enqueue_retry(account_id):
    """Re-enqueue this account for a one-off retry in 30 min."""
    frappe.enqueue(
        "my_app.tasks._sync_one_account",
        queue="default",
        timeout=300,
        job_name=f"sync-retry-{account_id}-{nowdate()}",  # dedup on (account, date)
        account_id=account_id,
    )

See also

• server-scripts skill — basic frappe.enqueue and scheduler_events syntax
• bench-commands skill — bench show-pending-jobs, bench restart-workers, bench doctor
• frappe-debug skill — bench error log monitor catches worker exceptions live
• /frappe-perf command — runs bench doctor + show-pending-jobs and surfaces stuck jobs
• references/frappe-python.md (in frappe-review skill) — review-time checks for queue mismatches, missing idempotency, now=True in production