Open-source monitoring for Celery

See every task.
Keep every byte.

Feloxi turns the events your Celery workers already emit into live dashboards, searchable history and alerts that fire. It runs on your servers. No agent, no SDK, and nothing leaves your network.

Get started

docker compose up

Redis or RabbitMQ · Docker Compose, Helm or plain manifests · Apache-2.0

TypeError ×17tasks.email.send_welcome_email09:33:02 · sent to Slack #ops-alerts

TypeError ×17send_welcome_email→ Slack

The Feloxi overview: throughput, failure rate, runtime and backlog readouts, tasks per minute by outcome, failures grouped by exception, and queue depths.

0
agents to install: 3
lines of Celery config: 2 GB
is enough to start: 10
alert conditions: Apache-2.0
license

01How it works

Your broker already knows. Feloxi just listens.

Celery publishes an event every time a task is sent, starts, retries, succeeds or fails. Feloxi subscribes to those events on the broker you already run, so there is nothing to install next to your code.

Turn on events

Three settings, then restart workers with --events. Your task code stays as it is.

# celeryconfig.py
worker_send_task_events = True
task_send_sent_event = True
task_track_started = True

02
Point Feloxi at your broker
Add a Redis or RabbitMQ URL under Brokers and test the connection. Add staging and production side by side.
redis://redis.internal:6379/0Connected
03
Watch
Tasks show up within seconds. History goes to ClickHouse on your disk, kept as long as your retention policy says.

Your network

Celery workersyour code, unchanged

Redis or RabbitMQthe broker you run

FeloxiRust API · WebSocket

Dashboardin your browser

ClickHousetask history

PostgreSQLusers, rules, config

Redislive worker state

Vendor cloudsomeone else's disk

not involved

Your network

Celery workersyour code, unchanged

Redis or RabbitMQthe broker you run

FeloxiRust API · WebSocket

ClickHousetask history

Postgresusers

Redislive worker state

Vendor cloud: not involved

02Search and failures

Find the task that broke.

Search every task by ID, name, arguments, result or exception, then open it for the traceback, runtime, retries and the chain it ran in.

Failures collapse into one row per fingerprint, with IDs and URLs stripped out, so forty-two declined payments read as one problem, not forty-two.

Retry from the UIRevoke through the broker

state isFailedlast 24h82 failures · 4 fingerprints

Failing nowgrouped by fingerprint

TypeErrorNew
setup_billing() missing 1 required positional argument: 'plan'tasks.email.send_welcome_email
17
ValueError
Payment declined for order order_id: insufficient fundstasks.payments.process_payment
42
OSError
Failed to fetch image: urltasks.media.resize_image
15
RuntimeError
Transcoding failed for url: codec errortasks.media.transcode_video
8

03Alerting

# ops-alerts2 messages from one incident

Feloxi09:33

FiringWelcome emails failing

Failure rate of tasks.email.send_welcome_email is 12.5% over 5m (threshold 10%).Likely cause: TypeError setup_billing() missing 1 required positional argument, 3 events.

Nothing in between. No repeats while it keeps firing.

Feloxi09:47

ResolvedWelcome emails failing

Back under threshold at 1.8%. Fired for 14m 06s.

One page when it breaks. One when it's fixed.

Alerts behave like incidents: a notification when a condition starts firing, a resolved notice with the duration when it clears, and nothing in between. Failed deliveries retry with backoff and land in the delivery log.

01Failure rate
02Slow tasks
03Worker offline
04Queue depth
05Throughput anomaly
06Latency anomaly
07Error spike
08Beat missed
09No events
10Task failed

Slack
Discord
PagerDuty
Email
Webhook

Silences for maintenance windows. Severity floors per channel, so warnings go to Slack and only critical pages wake someone up.

04And the rest of the queue

Workflows

Chains, groups and chords, drawn as they ran.

When step three of a chord fails, you see which branch broke and what was skipped because of it.

import_csv · sales2.6 s

import_csv · inventory3.6 s

import_csv · customers2.4 s

aggregate_report_dataskipped

Workers

Grouped the way you deploy them.

Slots in use, heartbeat gaps, load and throughput per deployment, with remote shutdown when a worker misbehaves.

worker-payments@prod-1
worker-payments@prod-2
worker-media@prod-1
worker-media@prod-2

Schedules

Every Beat schedule, and the one that missed.

Last and next run for each periodic task, plus an alert when a schedule stops running on time.

check-system-healthevery 30s8s ago
sync-all-inventoriesevery 2m74s ago
nightly-report0 0 * * *missed
send-digestsevery 5m3m ago

05Self-hosted

Nothing crosses the line.

Task arguments, results and tracebacks often hold customer data. Feloxi keeps them in your own ClickHouse, behind your own network rules. No vendor account, no telemetry, no per-seat bill.

Scale	Tasks/day	Workers	ClickHouse
Small	< 100K	< 10	1 CPU, 1 GB
Medium	100K–1M	10–100	2 CPU, 4 GB
Large	1M–10M	100–500	4 CPU, 16 GB

From the sizing guide, which also covers the API, Redis and Postgres.

Docker Compose

docker compose up -d

A laptop or a 2 GB VM. Boots the whole stack with a seeded demo.

Helm

helm install feloxi oci://ghcr.io/thesaadmirza/charts/feloxi

Kubernetes, with bundled Postgres, ClickHouse and Redis, or point it at managed ones.

Manifests

kubectl apply -f deploy/k8s/

Plain Kubernetes manifests for GitOps with Argo CD or Flux.

06Compare

As cheap as DIY. As complete as SaaS.

Flower is where most Celery teams start. Hosted APMs cover more ground, and bill for it. Feloxi is the task-queue piece, done properly and kept on your side.

	Feloxi	Flower	Hosted APM	Prometheus + Grafana
Task history	ClickHouse, kept as long as you set	In memory, lost on restart unless persistent mode is on	By plan and price tier	Metrics only, no per-task detail
Alerting	10 conditions, built in	None	Yes	Alertmanager rules you write
Chains and chords	Drawn as DAGs	No	Varies	No
Where data lives	Your infrastructure	Your infrastructure	Vendor cloud	Your infrastructure
In your app	Nothing	Nothing	Agent and SDK	An exporter
Price	Free	Free	Per host, seat or event	Free, plus your time
License	Apache-2.0	BSD-3-Clause	Proprietary	Apache-2.0, AGPL

Task history: ClickHouse, kept as long as you set; Flower: In memory, lost on restart unless persistent mode is on · Hosted APM: By plan and price tier · Prometheus: Metrics only, no per-task detail
Alerting: 10 conditions, built in; Flower: None · Hosted APM: Yes · Prometheus: Alertmanager rules you write
Chains and chords: Drawn as DAGs; Flower: No · Hosted APM: Varies · Prometheus: No
Where data lives: Your infrastructure; Flower: Your infrastructure · Hosted APM: Vendor cloud · Prometheus: Your infrastructure
In your app: Nothing; Flower: Nothing · Hosted APM: Agent and SDK · Prometheus: An exporter
Price: Free; Flower: Free · Hosted APM: Per host, seat or event · Prometheus: Free, plus your time
License: Apache-2.0; Flower: BSD-3-Clause · Hosted APM: Proprietary · Prometheus: Apache-2.0, AGPL

07Roadmap

Task queues first. Then the rest of the stack.

The same self-hosted core grows into logs, access control and dashboards, without a hosted tier that gets the features first. Request a feature

Shipped

Redis and RabbitMQ brokers
Workflow DAGs
Celery Beat tracking
Alert incidents, silences and severity routing
Slack and Discord in one click
Google sign-in and magic links
Scoped API keys
Prometheus /metrics
Helm chart

Building

End-to-end alert encryption

Log search, joined to tasks
SAML and OIDC single sign-on
Audit log
Custom dashboards
SLA tracking
Distributed tracing

Clone it, run it,
watch your queues.

One docker compose up boots the whole stack with a seeded demo. Free, with the source on GitHub.

Get started Star on GitHub

See every task.Keep every byte.