Retries, persistence, and visibility for FastAPI's BackgroundTasks.
Production-grade background tasks. No workers, no brokers, drop-in.

---

You shipped background_tasks.add_task(send_email, address=email) and it worked in dev. Then you deployed it and a task failed. You had no idea. No retry happened. No log survived. The user never got their email. You found out three days later when they complained.

FastAPI's BackgroundTasks is fine for simple fire-and-forget work. But production has a way of finding the gaps.

Does this sound familiar?

• A task failed and you only found out when a user complained
• You restarted the server and lost every pending task in memory
• You have no idea how many tasks ran, which ones failed, or how long they took
• You added try/except and a print statement because there is no other option
• You considered adding Celery just to get retries and visibility, then saw the ops cost

fastapi-taskflow is a thin layer on top of BackgroundTasks. You keep your existing code. You get retries, persistence, a live dashboard, structured logging, and task history. No brokers, no workers, no new infrastructure.

Quick start

pip install fastapi-taskflow

from fastapi import BackgroundTasks, FastAPI
from fastapi_taskflow import TaskAdmin, TaskManager

task_manager = TaskManager(snapshot_db="tasks.db", snapshot_interval=30.0)
app = FastAPI()

# auto_install=True wires FastAPI's BackgroundTasks injection so existing
# route signatures work without any changes.
TaskAdmin(app, task_manager, auto_install=True)


@task_manager.task(retries=3, delay=1.0, backoff=2.0)
def send_email(address: str) -> None:
    ...


@app.post("/signup")
def signup(email: str, background_tasks: BackgroundTasks):
    task_id = background_tasks.add_task(send_email, address=email)
    return {"task_id": task_id}

uvicorn examples.basic_app:app --reload

curl -X POST "http://localhost:8000/signup?email=user@example.com"
curl "http://localhost:8000/tasks"
curl "http://localhost:8000/tasks/metrics"
open "http://localhost:8000/tasks/dashboard"

The route signature does not change. Tasks that fail are retried. If the server restarts before a task finishes, it is re-dispatched on startup. Every task has a UUID, a status, and a history entry you can query.

Features

• Automatic retries with configurable delay and exponential backoff
• Task IDs and full lifecycle tracking: PENDING, RUNNING, SUCCESS, FAILED, INTERRUPTED
• Live admin dashboard over SSE at /tasks/dashboard
• SQLite persistence out of the box; Redis, PostgreSQL, and MySQL as optional extras
• Pending task requeue: unfinished tasks at shutdown are re-dispatched on startup
• requeue_on_interrupt: opt-in requeue for idempotent tasks interrupted mid-execution
• Idempotency keys: prevent duplicate execution of the same logical operation
• Multi-instance support: atomic requeue claiming, shared task history across instances
• task_log(message, level=, **extra): structured log entries with level filtering and arbitrary extra fields
• get_task_context(): access task metadata (task_id, attempt, tags) from any code path inside a running task
• Tags: attach key/value labels at enqueue time, forwarded to every log and lifecycle event
• Pluggable observers: FileLogger, StdoutLogger, InMemoryLogger, and custom TaskObserver implementations
• Argument encryption: Fernet-based at-rest encryption for task args and kwargs
• Trace context propagation: OpenTelemetry spans flow from the request into background execution (Python 3.11+)
• Process executor: executor='process' routes CPU-bound tasks through a ProcessPoolExecutor, bypassing the GIL with true parallel workers
• Concurrency controls: opt-in semaphore for async tasks, dedicated thread pool for sync tasks, and configurable process worker count
• Priority queues: priority= on @task_manager.task() or add_task(), higher-priority tasks run first, equal-priority tasks are FIFO
• Eager dispatch: eager=True starts a task immediately via asyncio.create_task before the HTTP response is sent
• Scheduled tasks: @task_manager.schedule(every=) and cron= with distributed lock for multi-instance
• Zero-migration injection: keep your existing BackgroundTasks annotations
• Both sync and async task functions supported

Installation

pip install fastapi-taskflow

With all optional dependencies:

pip install "fastapi-taskflow[all]"

Or install only what you need:

Extra	Installs	Required for
redis	redis[asyncio]	Redis persistence backend
postgres	psycopg2-binary	PostgreSQL persistence backend
mysql	PyMySQL	MySQL / MariaDB persistence backend
scheduler	croniter	Cron-based scheduled tasks
encryption	cryptography	Argument encryption at rest
process	cloudpickle	Extended serialization for executor='process' tasks

pip install "fastapi-taskflow[redis]"
pip install "fastapi-taskflow[postgres]"
pip install "fastapi-taskflow[mysql]"
pip install "fastapi-taskflow[scheduler]"
pip install "fastapi-taskflow[encryption]"
pip install "fastapi-taskflow[process]"

Injection patterns

Three ways to get a ManagedBackgroundTasks instance into your routes:

# Pattern 1: keep the native annotation (requires auto_install=True)
def route(background_tasks: BackgroundTasks):
    task_id = background_tasks.add_task(my_func, arg)

# Pattern 2: explicit managed type (also requires auto_install=True)
from fastapi_taskflow import ManagedBackgroundTasks

def route(background_tasks: ManagedBackgroundTasks):
    task_id = background_tasks.add_task(my_func, arg)

# Pattern 3: explicit Depends (no install() required)
from fastapi import Depends

def route(tasks=Depends(task_manager.get_tasks)):
    task_id = tasks.add_task(my_func, arg)

Decorator options

Parameter	Type	Default	Description
retries	int	0	Additional attempts after the first failure
delay	float	0.0	Seconds before the first retry
backoff	float	1.0	Multiplier applied to delay on each retry
persist	bool	False	Save this task for requeue on restart
name	str	function name	Override the name shown in the dashboard
requeue_on_interrupt	bool	False	Requeue this task if it was mid-execution at shutdown. Only set for idempotent tasks.
eager	bool	False	Start the task via asyncio.create_task immediately when add_task() is called, before the response is sent. Per-call eager on add_task() overrides this.
priority	int | None	None	Route through the priority queue. Higher integers run first. Conventional range 1 (lowest) to 10 (highest). Per-call priority on add_task() overrides this.

Idempotency keys

Pass an idempotency_key to add_task() to prevent the same logical operation from running twice. If a non-failed task with the same key already exists, the original task_id is returned and the task is not enqueued again.

task_id = tasks.add_task(send_notification, order_id, idempotency_key="order-42-notified")

Useful for handling retried HTTP requests, duplicate webhook deliveries, or double-clicks.

API endpoints

Method	Path	Description
GET	/tasks	List all tasks
GET	/tasks/{task_id}	Single task detail
GET	/tasks/metrics	Aggregated stats
GET	/tasks/dashboard	Live HTML dashboard
POST	/tasks/{task_id}/retry	Retry a failed or interrupted task

Multi-instance deployments

fastapi-taskflow supports running multiple instances behind a load balancer when a shared backend is configured.

Same host, multiple processes -- use SQLite. All instances share the same file. Requeue claiming is atomic so only one instance picks up each task on restart.

Multiple hosts -- use Redis, PostgreSQL, or MySQL. All instances share the same backend. Idempotency keys, requeue claiming, and completed task history all work across hosts.

from fastapi_taskflow.backends import RedisBackend

task_manager = TaskManager(
    snapshot_backend=RedisBackend("redis://localhost:6379/0"),
    requeue_pending=True,
)

Dashboard in multi-instance deployments -- the dashboard shows live tasks for the instance it is connected to. Completed tasks from all instances are visible via the shared backend (with a short cache window). For accurate live task visibility, route dashboard traffic to a single instance using sticky sessions at the load balancer.

Known caveats:

• Live PENDING and RUNNING tasks from other instances are not visible in real time. Each instance only holds its own in-memory state.
• SQLite multi-instance only works when all processes share the same file path on the same host. It does not work across separate machines.
• Tasks in RUNNING state at the time of a hard crash (SIGKILL, OOM) cannot be recovered. Only clean shutdowns trigger the pending store write.

Structured task logging

task_log() accepts an optional level= and arbitrary **extra keyword fields. Extras are forwarded to observers as structured fields rather than embedded in the message string.

from fastapi_taskflow import get_task_context, task_log

@task_manager.task(retries=3)
def process_order(order_id: int) -> None:
    ctx = get_task_context()
    task_log("Processing order", order_id=order_id, attempt=ctx.attempt if ctx else 0)
    task_log("Payment gateway error", level="warning", order_id=order_id, code=503)

get_task_context() returns a TaskContext with task_id, func_name, attempt, and tags from any code path inside a running task.

Observability

Pass one or more observers to loggers= to receive structured LogEvent and LifecycleEvent objects for every task_log() call and status transition:

from fastapi_taskflow import FileLogger, StdoutLogger, TaskManager

task_manager = TaskManager(
    snapshot_db="tasks.db",
    loggers=[
        FileLogger("tasks.log", log_lifecycle=True),
        StdoutLogger(log_lifecycle=True, min_level="warning"),
    ],
)

Built-in observers:

Observer	Description
FileLogger	Rotating plain text file. Works with tail -f, grep, and log shippers.
StdoutLogger	Prints to stdout. Suitable for containers with a log agent.
InMemoryLogger	Captures events in memory for test assertions.

Tags

Attach key/value labels to a task at enqueue time. Tags flow through to every log and lifecycle event.

task_id = tasks.add_task(
    send_email,
    address=email,
    tags={"user_id": str(user_id), "source": "signup"},
)

Argument encryption

When tasks carry sensitive data, encrypt args and kwargs at rest with Fernet:

import os
from fastapi_taskflow import TaskManager

task_manager = TaskManager(
    snapshot_db="tasks.db",
    encrypt_args_key=os.environ["TASK_ENCRYPTION_KEY"],
)

Generate a key:

python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"

Args and kwargs are encrypted at add_task() time and decrypted only inside the executor. They are never stored in plain text in the task store, the database, or any log file. Requires pip install "fastapi-taskflow[encryption]".

Concurrency controls

By default, tasks share the event loop and thread pool with request handlers. Under burst task load this can increase request latency. Both controls are opt-in and do not change existing behaviour when not set.

max_concurrent_tasks -- caps how many async tasks hold event loop time simultaneously via an asyncio.Semaphore. Tasks waiting for a slot are parked without blocking the loop.

max_sync_threads -- runs sync task functions in a dedicated ThreadPoolExecutor, isolated from the default pool used by sync request handlers.

task_manager = TaskManager(
    snapshot_db="tasks.db",
    max_concurrent_tasks=10,
    max_sync_threads=8,
)

Both default to None. When not set, execution is identical to previous versions.

Priority queues

Pass priority= to control execution order. Higher-priority tasks run before lower-priority ones regardless of arrival order. Equal-priority tasks execute in arrival (FIFO) order.

@task_manager.task(retries=2, priority=9)
async def send_otp(phone: str) -> None:
    ...

@task_manager.task(priority=1)
def generate_report(user_id: int) -> None:
    ...

@app.post("/otp")
async def otp(phone: str, tasks=Depends(task_manager.get_tasks)):
    task_id = tasks.add_task(send_otp, phone)
    return {"task_id": task_id}

Set priority= at the decorator level as a default, then override it per call:

task_id = tasks.add_task(process_item, item_id, priority=10)  # rush job
task_id = tasks.add_task(process_item, item_id, priority=1)   # defer
task_id = tasks.add_task(process_item, item_id)               # use decorator default

Tasks with no priority route through Starlette's normal background task list unchanged.

Eager dispatch

Set eager=True to start a task via asyncio.create_task the moment add_task() is called, before FastAPI sends the response. Useful for batch endpoints where multiple tasks are added in a single request handler and you want them to run concurrently rather than queued sequentially.

@app.post("/batch")
async def batch(items: list[str], tasks=Depends(task_manager.get_tasks)):
    task_ids = []
    for item in items:
        task_id = tasks.add_task(process_item, item, eager=True)
        task_ids.append(task_id)
    return {"task_ids": task_ids}

Override per call with tasks.add_task(func, arg, eager=True). When priority= is also set, the priority queue governs dispatch and eager is ignored.

Scheduled tasks

Run functions automatically at a fixed interval or on a cron expression. Scheduled tasks go through the same execution path as manually enqueued tasks, so retries, logging, persistence, and the dashboard all work without any extra setup.

from fastapi import FastAPI
from fastapi_taskflow import TaskAdmin, TaskManager, task_log

task_manager = TaskManager(snapshot_db="tasks.db")
app = FastAPI()
TaskAdmin(app, task_manager)


@task_manager.schedule(every=300, retries=1)
async def health_check() -> None:
    task_log("Running health check")
    ...


@task_manager.schedule(cron="0 2 * * *")
def nightly_cleanup() -> None:
    task_log("Starting cleanup")
    ...


@task_manager.schedule(cron="0 9 * * *", timezone="America/New_York")
async def morning_report() -> None:
    ...

Cron expressions require pip install "fastapi-taskflow[scheduler]". Interval-based schedules have no extra dependencies. Cron expressions default to UTC. Pass any IANA timezone name with timezone= to evaluate in local time.

In multi-instance deployments, a distributed lock ensures only one instance fires each scheduled entry per interval.

Custom dashboard title

Replace the "fastapi-taskflow" label in the dashboard header and login page with your own app name:

TaskAdmin(app, task_manager, title="My App")

File logging

In addition to the observer system, a plain text log file can be configured directly on TaskManager:

task_manager = TaskManager(
    snapshot_db="tasks.db",
    log_file="tasks.log",
    log_lifecycle=True,
)

Each line has the format [task_id] [func_name] 2026-01-01T12:00:00 message. For multi-process or multi-host deployments see the file logging guide.

What this is not

fastapi-taskflow does not compete with Celery, ARQ, Taskiq, or Dramatiq. Those tools are built for distributed workers, message brokers, and high-throughput task routing across separate machines.

This library is for teams using FastAPI's native BackgroundTasks who want retries, visibility, and resilience without adding worker infrastructure. If your tasks need to run on dedicated worker processes completely separate from your web application, use a proper task queue.

Contributing

Bug reports and pull requests are welcome. Open an issue on GitHub for anything unexpected, a feature request, or a question about intended behaviour before submitting a PR.

Contact

Quaicoe Richard (Attakay)

For questions, feedback, or anything related to the project:

• Email: richardquaicoe78@gmail.com
• LinkedIn: linkedin.com/in/richard-quaicoe-545ba211b
• X: @richman_khay

License

MIT