What it does not do yet
Guides
- No periodic jobs. A schedule that runs once across several processes
needs a leader or a unique slot per instant; it is the first thing a
second process asks for.
- No limits across processes. Concurrency is per runner: two runners of
one queue run twice its concurrency, and there is no global rate.
- No dependencies between jobs, batches or workflows.
- No pausing a queue, and no fairness between keys: one key’s many jobs
run before another’s few, in the order the claim reads.
- A job’s lines carry no request id or trace of the enqueuer’s. The
runner’s statements reach the application’s observer, but nothing ties a
job to the request that made it owed.
- A queue’s counts are exported nowhere.
counts answers them, and
nothing turns them into metrics.
- A handler that never yields cannot be stopped. A timeout, a lost lease
and stopping cancel a fiber at its next effect, so a handler busy on the
CPU without one runs on until it ends.
- Two clocks that disagree by more than a quarter of a lease can run one
job in two places: a lease is an instant from the claiming runner’s
now, compared against another’s when it rescues.
