Scheduling and cron

import gylo

app = gylo.Gylo()


@app.task
async def remind(user_id: int) -> None: ...


async def defer(conn) -> None:
    await remind.options(delay=3600.0).enqueue(conn, 7)


@app.cron("0 3 * * *", name="nightly-report", timezone="Europe/London")
async def nightly_report() -> None: ...


@app.cron("*/5 * * * *", name="sync-inventory", queue="sync")
async def sync_inventory() -> None: ...

Delayed jobs

delay= holds a job for that many seconds before it becomes runnable. The delay is applied in the database, so it survives worker restarts and needs no scheduler process.

Cron

@app.cron(expression, ...) registers a task that also runs on a schedule. Five-field expressions are standard cron; a six-field form adds leading seconds. timezone matters for anything coarser than hourly: 0 3 * * * in Europe/London is a different UTC instant in January than in July, and gylo resolves the zone — including skipped and repeated hours at DST transitions — rather than drifting twice a year.

The schedule lives with the code. When a worker starts, its children declare their schedules upward and gylo records them; deploying the code is deploying the schedule. Editing the expression takes effect on the next deploy; a rename is a new schedule.

There is no leader. Every worker checks for due schedules; a conditional UPDATE on the schedule row means exactly one of them wins each occurrence. No election, no lock service, no single process whose death silences every schedule. This was measured against the alternative: leader election added a failover stall and saved nothing.

Missed runs are skipped, not backfilled. A schedule due while the whole fleet was down advances to its next occurrence on recovery. If you need catch-up semantics, enqueue the backlog explicitly — silently replaying an unknown number of missed runs is the wrong default for jobs like "charge all customers".

Resolution is the worker's maintenance interval (10s by default): schedules are examined once per tick, so an every-second expression fires once per tick, not sixty times a minute.

Pausing

Set paused on the row in gylo_cron to stop a schedule without deploying. gylo never unsets it — a redeploy updates the schedule's definition but leaves an operator's pause standing.