GitLab

"running" is a state, not a transition.

Closes #17057660.

This test the fix in cw-celerytask-helpers: https://www.cubicweb.org/patch/17056309

Previous cset was rebased and published.
Also remove the '#egg=cw-celerytask-helpers' since it's only used for git or hg
urls.

PROGRESS is a custom state used to handle progress bar in views.
When we receive PROGRESS, fire transition 'running' (if possible).

Convention is to return a dict with a key "celerytask_subtasks".
So we can remove a "except Exception" that could hide errors or future bugs.

Set CW_CONNECTIONS_POOL_SIZE to 1 (by default) to only use one connection to
the database (no more is needed since it's single threaded).

Setting the parameter in the "config" object could not work since CubicWeb
reload it from config files when loading cubes.

When multiple instance of celery-monitor are running, we could have an
integrity errors raised if two instance are working on the same task_id or
temporary (network, host) failures. In this case we want to retry handling the
task_id later.

Put pending processed task_id in a "pending queue" and each minutes and if the
monitor queue is empty, requeue pending items.

This change require to handle the "timeout" parameter of loop (only used in
tests) in a different way to ensure not blocking forever in redis "brpoplpush".

bump version to 0.4.0.dev0

IMHO this looks better than "if rset".
Also log created entity eid.

We should not stop celery-monitor in case of temporary unavailability of redis
or cubicweb database.
Handle errors and automatically retry each 5 seconds.

This usecase (sending a serialized task signature as result of a task)  is
exceptional, so avoid flooding the logs with such messages.

Previously we handle monitoring celery task by listening to celery event bus
(celery.events.EventReceiver) that was not persistent. In this case we used a
dedicated routine (on_monitor_start) to synchronize non finished tasks, but
this wasn't working in case of an untracked task (eg. not started with
start_async_task). Also this was a single point of failure because it cannot
run in multiple instance without concurrency issues (events are sent to all
instances)

Now we use a redis queue where worker put task_id and task_name to be
synchronized and celery-monitor use brpop (https://redis.io/commands/brpop) to
process the queue.

We don't require CELERY_SEND_EVENTS to be enabled anymore (-E or --events in
worker options).
We require to add 'cw_celerytask_helpers.helpers' to CELERY_IMPORTS.

- strict policy on sql transaction (commit all or nothing for each task)
- make it work with on_monitor_start (used to synchronize task states when
  monitor starts in case of missing events)
- Use serializable() in "spawn" task so it force celery to use the json task
  serializer instead of pickle
- Don't always update task_name when creating subtasks outside of celerytask
  (eg. by using start_async_task), use a fixed identifier "<unknown>" as task
  name instead and only update these.

The view was called once and then being redirected to incontext view by
javascript code when task state changed.
Give the current vid to call on state change instead.

Add a method entity_content() to implement difference between incontext and outofcontext views.

Do not create a CeleryTask if no worker has aknowledge the said task.

celery.result.from_serializable is no more present in 4.x

it may occur that a transition cannot be fired due to a hook triggered by a
previous transition. Especially the on_monitor_start() was prone to this kind
of error since it was doing all the transitions in a single transaction.

[entities] improve workflow management of CeleryTasks (by celery-monitor) robustness (related to #16640842)

the transition may fail (if for some race-like reasons it has already been fired).

Also ensure some CeleryTask changes are commited (celery-monitor)
since the transaction may be rolled back during the WF management part of the
on_event() method.

including all the children subtasks' progress values

the celery configuration stuff for tests is not needed there

that returns an iterator on all parents recursively