FAQ

Is execution exactly-once?

Execution is at-least-once: an attempt can run again if its worker crashes, misses heartbeats, or the leader fails over mid-attempt. Side effects can still be exactly-once: record the task id and attempt number (a fencing token) with each effect and let only the newest attempt write. See Exactly-once side effects. Task keys make submission idempotent.

Does a deadline cancel the task when it passes?

No. Deadlines raise urgency; they never cancel. Use timeout to bound an attempt, or cancel the task yourself.

What is the difference between a deadline and a fixed start?

A deadline says when work should be finished and lets cronoq reorder work to meet it. A fixed start says when work should begin; cronoq reserves capacity and holds a worker for it.

Can low-priority work starve?

Not behind normal traffic: aging lets a waiting low task overtake fresh normal work after about 45 seconds with the default settings. It can wait behind sustained high or critical load, by design. For stronger mixing, raise age_cap; it is configurable only from the Rust library (Builder::priority or Builder::configure), and the server and the Python package use the defaults.

How many tasks can the engine hold?

Memory per pending task is a few hundred bytes plus its payload. Millions of pending tasks are practical; use max_pending to bound them.

Can I use cronoq without tokio?

From Rust, use BlockingScheduler (it owns a private tokio runtime), or drive cronoq-core yourself. From Python, the bindings manage their own runtime.

Do I need Postgres?

No. Without a store, state is in memory. For durability on one host, use the SQLite store (a local file, with standby processes on the same host). Use Postgres to run replicas on several hosts.

How do workers find the leader in an HA setup?

Through a load balancer or Kubernetes Service that routes only to ready endpoints. Only the leader reports /readyz = 200.

Can tasks pass data to their dependents?

Not directly. Put references (object-store paths, database ids) in payloads, or read the parent's output with Wait / TaskHandle in the producer and submit the child afterwards.

How do I limit tokens per minute rather than requests?

Create a token bucket whose units are, for example, thousands of tokens, and have each task request ceil(estimated_tokens / 1000) units.

Why are my server tasks not running?

A server pool's capacity is its connected workers' slots. Check cronoq_workers_connected and cronoq_pool_workers{pool=...}; then check that the task's resources exist and are not exhausted.

Is there a UI?

There is no web UI. cronoq top gives a live view in the terminal; for the rest, use /v1/snapshot, the Admin service, Prometheus/Grafana and status events.

What happens when the database is unavailable?

The scheduler keeps running in memory and retries the commit with backoff for up to max_store_outage (5 minutes by default). Meanwhile, calls that need durability (submissions, cancellations, schedule changes) wait, and no new attempts start. If the outage lasts longer, the scheduler stops and those callers get a store error (UNAVAILABLE / 503 on the server); another replica takes over once the database is back. See Durability and HA.

Can several teams share one server?

Within one trust domain, yes: give each team tokens scoped to its pools (pools = [...]), so it can only submit into, serve, watch and look up its own pools. Keys and resources are shared across pools, so teams that must not learn about each other's work need separate servers. See Security hardening.

How do I upgrade a cluster with a durable store?

The new version applies its schema migrations when it connects, and an older build refuses to start against a newer schema, so upgrades roll forward only. When a release adds no migration, replace the standbys first and the leader last; accepted work survives the leader change and workers reconnect on their own. When the changelog lists a migration, follow its upgrade notes (or stop all replicas, then start the new version).