This addon adds an integrated Job Queue to Odoo.
It allows to postpone method calls executed asynchronously.
Jobs are executed in the background by a Jobrunner
, in their own transaction.
Example:
from odoo import models, fields, api
class MyModel(models.Model):
_name = 'my.model'
def my_method(self, a, k=None):
_logger.info('executed with a: %s and k: %s', a, k)
class MyOtherModel(models.Model):
_name = 'my.other.model'
def button_do_stuff(self):
self.env['my.model'].with_delay().my_method('a', k=2)
In the snippet of code above, when we call button_do_stuff
, a job capturing
the method and arguments will be postponed. It will be executed as soon as the
Jobrunner has a free bucket, which can be instantaneous if no other job is
running.
Features:
- Views for jobs, jobs are stored in PostgreSQL
- Jobrunner: execute the jobs, highly efficient thanks to PostgreSQL's NOTIFY
- Channels: give a capacity for the root channel and its sub-channels and segregate jobs in them. Allow for instance to restrict heavy jobs to be executed one at a time while little ones are executed 4 at a times.
- Retries: Ability to retry jobs by raising a type of exception
- Retry Pattern: the 3 first tries, retry after 10 seconds, the 5 next tries, retry after 1 minutes, ...
- Job properties: priorities, estimated time of arrival (ETA), custom description, number of retries
- Related Actions: link an action on the job view, such as open the record concerned by the job
Table of contents
Be sure to have the requests
library.
- Using environment variables and command line:
- Adjust environment variables (optional):
ODOO_QUEUE_JOB_CHANNELS=root:4
or any other channels configuration. The default isroot:1
- if
xmlrpc_port
is not set:ODOO_QUEUE_JOB_PORT=8069
- Start Odoo with
--load=web,queue_job
and--workers
greater than 1. [1]
- Adjust environment variables (optional):
- Keep in mind that the number of workers should be greater than the number of
channels.
queue_job
will reuse normal Odoo workers to process jobs. It will not spawn its own workers. - Using the Odoo configuration file:
[options]
(...)
workers = 6
server_wide_modules = web,queue_job
(...)
[queue_job]
channels = root:2
- Environment variables have priority over the configuration file.
- Confirm the runner is starting correctly by checking the odoo log file:
...INFO...queue_job.jobrunner.runner: starting ...INFO...queue_job.jobrunner.runner: initializing database connections ...INFO...queue_job.jobrunner.runner: queue job runner ready for db <dbname> ...INFO...queue_job.jobrunner.runner: database connections ready
- Create jobs (eg using
base_import_async
) and observe they start immediately and in parallel. - Tip: to enable debug logging for the queue job, use
--log-handler=odoo.addons.queue_job:DEBUG
[1] | It works with the threaded Odoo server too, although this way of running Odoo is obviously not for production purposes. |
Be sure to check out Jobs Garbage Collector CRON and change enqueued_delta and started_delta parameters to your needs.
enqueued_delta
: Spent time in minutes after which an enqueued job is considered stuck. Set it to 0 to disable this check.started_delta
: Spent time in minutes after which a started job is considered stuck. This parameter should not be less than--limit-time-real // 60
parameter in your configuration. Set it to 0 to disable this check. Set it to -1 to automate it, based in the server's--limit-time-real
config parameter.
# `model` corresponds to 'queue.job' model model.requeue_stuck_jobs(enqueued_delta=1, started_delta=-1)
To use this module, you need to:
- Go to
Job Queue
menu
The fast way to enqueue a job for a method is to use with_delay()
on a record
or model:
def button_done(self):
self.with_delay().print_confirmation_document(self.state)
self.write({"state": "done"})
return True
Here, the method print_confirmation_document()
will be executed asynchronously
as a job. with_delay()
can take several parameters to define more precisely how
the job is executed (priority, ...).
All the arguments passed to the method being delayed are stored in the job and
passed to the method when it is executed asynchronously, including self
, so
the current record is maintained during the job execution (warning: the context
is not kept).
Dependencies can be expressed between jobs. To start a graph of jobs, use delayable()
on a record or model. The following is the equivalent of with_delay()
but using the
long form:
def button_done(self):
delayable = self.delayable()
delayable.print_confirmation_document(self.state)
delayable.delay()
self.write({"state": "done"})
return True
Methods of Delayable objects return itself, so it can be used as a builder pattern, which in some cases allow to build the jobs dynamically:
def button_generate_simple_with_delayable(self):
self.ensure_one()
# Introduction of a delayable object, using a builder pattern
# allowing to chain jobs or set properties. The delay() method
# on the delayable object actually stores the delayable objects
# in the queue_job table
(
self.delayable()
.generate_thumbnail((50, 50))
.set(priority=30)
.set(description=_("generate xxx"))
.delay()
)
The simplest way to define a dependency is to use .on_done(job)
on a Delayable:
def button_chain_done(self):
self.ensure_one()
job1 = self.browse(1).delayable().generate_thumbnail((50, 50))
job2 = self.browse(1).delayable().generate_thumbnail((50, 50))
job3 = self.browse(1).delayable().generate_thumbnail((50, 50))
# job 3 is executed when job 2 is done which is executed when job 1 is done
job1.on_done(job2.on_done(job3)).delay()
Delayables can be chained to form more complex graphs using the chain()
and
group()
primitives.
A chain represents a sequence of jobs to execute in order, a group represents
jobs which can be executed in parallel. Using chain()
has the same effect as
using several nested on_done()
but is more readable. Both can be combined to
form a graph, for instance we can group [A] of jobs, which blocks another group
[B] of jobs. When and only when all the jobs of the group [A] are executed, the
jobs of the group [B] are executed. The code would look like:
from odoo.addons.queue_job.delay import group, chain
def button_done(self):
group_a = group(self.delayable().method_foo(), self.delayable().method_bar())
group_b = group(self.delayable().method_baz(1), self.delayable().method_baz(2))
chain(group_a, group_b).delay()
self.write({"state": "done"})
return True
When a failure happens in a graph of jobs, the execution of the jobs that depend on the
failed job stops. They remain in a state wait_dependencies
until their "parent" job is
successful. This can happen in two ways: either the parent job retries and is successful
on a second try, either the parent job is manually "set to done" by a user. In these two
cases, the dependency is resolved and the graph will continue to be processed. Alternatively,
the failed job and all its dependent jobs can be canceled by a user. The other jobs of the
graph that do not depend on the failed job continue their execution in any case.
Note: delay()
must be called on the delayable, chain, or group which is at the top
of the graph. In the example above, if it was called on group_a
, then group_b
would never be delayed (but a warning would be shown).
- priority: default is 10, the closest it is to 0, the faster it will be executed
- eta: Estimated Time of Arrival of the job. It will not be executed before this date/time
- max_retries: default is 5, maximum number of retries before giving up and set the job state to 'failed'. A value of 0 means infinite retries.
- description: human description of the job. If not set, description is computed from the function doc or method name
- channel: the complete name of the channel to use to process the function. If specified it overrides the one defined on the function
- identity_key: key uniquely identifying the job, if specified and a job with the same key has not yet been run, the new job will not be created
In earlier versions, jobs could be configured using the @job
decorator.
This is now obsolete, they can be configured using optional queue.job.function
and queue.job.channel
XML records.
Example of channel:
<record id="channel_sale" model="queue.job.channel">
<field name="name">sale</field>
<field name="parent_id" ref="queue_job.channel_root" />
</record>
Example of job function:
<record id="job_function_sale_order_action_done" model="queue.job.function">
<field name="model_id" ref="sale.model_sale_order" />
<field name="method">action_done</field>
<field name="channel_id" ref="channel_sale" />
<field name="related_action" eval='{"func_name": "custom_related_action"}' />
<field name="retry_pattern" eval="{1: 60, 2: 180, 3: 10, 5: 300}" />
</record>
The general form for the name
is: <model.name>.method
.
The channel, related action and retry pattern options are optional, they are documented below.
When writing modules, if 2+ modules add a job function or channel with the same name (and parent for channels), they'll be merged in the same record, even if they have different xmlids. On uninstall, the merged record is deleted when all the modules using it are uninstalled.
Job function: model
If the function is defined in an abstract model, you can not write
<field name="model_id" ref="xml_id_of_the_abstract_model"</field>
but you have to define a function for each model that inherits from the abstract model.
Job function: channel
The channel where the job will be delayed. The default channel is root
.
Job function: related action
The Related Action appears as a button on the Job's view. The button will execute the defined action.
The default one is to open the view of the record related to the job (form view when there is a single record, list view for several records). In many cases, the default related action is enough and doesn't need customization, but it can be customized by providing a dictionary on the job function:
{
"enable": False,
"func_name": "related_action_partner",
"kwargs": {"name": "Partner"},
}
enable
: whenFalse
, the button has no effect (default:True
)func_name
: name of the method onqueue.job
that returns an actionkwargs
: extra arguments to pass to the related action method
Example of related action code:
class QueueJob(models.Model):
_inherit = 'queue.job'
def related_action_partner(self, name):
self.ensure_one()
model = self.model_name
partner = self.records
action = {
'name': name,
'type': 'ir.actions.act_window',
'res_model': model,
'view_type': 'form',
'view_mode': 'form',
'res_id': partner.id,
}
return action
Job function: retry pattern
When a job fails with a retryable error type, it is automatically retried later. By default, the retry is always 10 minutes later.
A retry pattern can be configured on the job function. What a pattern represents is "from X tries, postpone to Y seconds". It is expressed as a dictionary where keys are tries and values are seconds to postpone as integers:
{
1: 10,
5: 20,
10: 30,
15: 300,
}
Based on this configuration, we can tell that:
- 5 first retries are postponed 10 seconds later
- retries 5 to 10 postponed 20 seconds later
- retries 10 to 15 postponed 30 seconds later
- all subsequent retries postponed 5 minutes later
Job Context
The context of the recordset of the job, or any recordset passed in arguments of a job, is transferred to the job according to an allow-list.
The default allow-list is ("tz", "lang", "allowed_company_ids", "force_company", "active_test"). It can
be customized in Base._job_prepare_context_before_enqueue_keys
.
Bypass jobs on running Odoo
When you are developing (ie: connector modules) you might want to bypass the queue job and run your code immediately.
To do so you can set QUEUE_JOB__NO_DELAY=1 in your enviroment.
Bypass jobs in tests
When writing tests on job-related methods is always tricky to deal with delayed recordsets. To make your testing life easier you can set queue_job__no_delay=True in the context.
Tip: you can do this at test case level like this
@classmethod
def setUpClass(cls):
super().setUpClass()
cls.env = cls.env(context=dict(
cls.env.context,
queue_job__no_delay=True, # no jobs thanks
))
Then all your tests execute the job methods synchronously without delaying any jobs.
Asserting enqueued jobs
The recommended way to test jobs, rather than running them directly and synchronously is to split the tests in two parts:
- one test where the job is mocked (trap jobs with
trap_jobs()
and the test only verifies that the job has been delayed with the expected arguments- one test that only calls the method of the job synchronously, to validate the proper behavior of this method only
Proceeding this way means that you can prove that jobs will be enqueued properly at runtime, and it ensures your code does not have a different behavior in tests and in production (because running your jobs synchronously may have a different behavior as they are in the same transaction / in the middle of the method). Additionally, it gives more control on the arguments you want to pass when calling the job's method (synchronously, this time, in the second type of tests), and it makes tests smaller.
The best way to run such assertions on the enqueued jobs is to use
odoo.addons.queue_job.tests.common.trap_jobs()
.
Inside this context manager, instead of being added in the database's queue,
jobs are pushed in an in-memory list. The context manager then provides useful
helpers to verify that jobs have been enqueued with the expected arguments. It
even can run the jobs of its list synchronously! Details in
odoo.addons.queue_job.tests.common.JobsTester
.
A very small example (more details in tests/common.py
):
# code
def my_job_method(self, name, count):
self.write({"name": " ".join([name] * count)
def method_to_test(self):
count = self.env["other.model"].search_count([])
self.with_delay(priority=15).my_job_method("Hi!", count=count)
return count
# tests
from odoo.addons.queue_job.tests.common import trap_jobs
# first test only check the expected behavior of the method and the proper
# enqueuing of jobs
def test_method_to_test(self):
with trap_jobs() as trap:
result = self.env["model"].method_to_test()
expected_count = 12
trap.assert_jobs_count(1, only=self.env["model"].my_job_method)
trap.assert_enqueued_job(
self.env["model"].my_job_method,
args=("Hi!",),
kwargs=dict(count=expected_count),
properties=dict(priority=15)
)
self.assertEqual(result, expected_count)
# second test to validate the behavior of the job unitarily
def test_my_job_method(self):
record = self.env["model"].browse(1)
record.my_job_method("Hi!", count=12)
self.assertEqual(record.name, "Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi!")
If you prefer, you can still test the whole thing in a single test, by calling
jobs_tester.perform_enqueued_jobs()
in your test.
def test_method_to_test(self):
with trap_jobs() as trap:
result = self.env["model"].method_to_test()
expected_count = 12
trap.assert_jobs_count(1, only=self.env["model"].my_job_method)
trap.assert_enqueued_job(
self.env["model"].my_job_method,
args=("Hi!",),
kwargs=dict(count=expected_count),
properties=dict(priority=15)
)
self.assertEqual(result, expected_count)
trap.perform_enqueued_jobs()
record = self.env["model"].browse(1)
record.my_job_method("Hi!", count=12)
self.assertEqual(record.name, "Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi! Hi!")
Execute jobs synchronously when running Odoo
When you are developing (ie: connector modules) you might want to bypass the queue job and run your code immediately.
To do so you can set QUEUE_JOB__NO_DELAY=1
in your environment.
Warning
Do not do this in production
Execute jobs synchronously in tests
You should use trap_jobs
, really, but if for any reason you could not use it,
and still need to have job methods executed synchronously in your tests, you can
do so by setting queue_job__no_delay=True
in the context.
Tip: you can do this at test case level like this
@classmethod
def setUpClass(cls):
super().setUpClass()
cls.env = cls.env(context=dict(
cls.env.context,
queue_job__no_delay=True, # no jobs thanks
))
Then all your tests execute the job methods synchronously without delaying any jobs.
In tests you'll have to mute the logger like:
@mute_logger('odoo.addons.queue_job.models.base')
Note
in graphs of jobs, the queue_job__no_delay
context key must be in at
least one job's env of the graph for the whole graph to be executed synchronously
- Idempotency (https://www.restapitutorial.com/lessons/idempotency.html): The queue_job should be idempotent so they can be retried several times without impact on the data.
- The job should test at the very beginning its relevance: the moment the job will be executed is unknown by design. So the first task of a job should be to check if the related work is still relevant at the moment of the execution.
Through the time, two main patterns emerged:
- For data exposed to users, a model should store the data and the model should be the creator of the job. The job is kept hidden from the users
- For technical data, that are not exposed to the users, it is generally alright to create directly jobs with data passed as arguments to the job, without intermediary models.
- After creating a new database or installing
queue_job
on an existing database, Odoo must be restarted for the runner to detect it. - When Odoo shuts down normally, it waits for running jobs to finish.
However, when the Odoo server crashes or is otherwise force-stopped,
running jobs are interrupted while the runner has no chance to know
they have been aborted. In such situations, jobs may remain in
started
orenqueued
state after the Odoo server is halted. Since the runner has no way to know if they are actually running or not, and does not know for sure if it is safe to restart the jobs, it does not attempt to restart them automatically. Such stale jobs therefore fill the running queue and prevent other jobs to start. You must therefore requeue them manually, either from the Jobs view, or by running the following SQL statement before starting Odoo:
update queue_job set state='pending' where state in ('started', 'enqueued')
- [ADD] Run jobrunner as a worker process instead of a thread in the main process (when running with --workers > 0)
- [REF]
@job
and@related_action
deprecated, any method can be delayed, and configured usingqueue.job.function
records - [MIGRATION] from 13.0 branched at rev. e24ff4b
Bugs are tracked on GitHub Issues. In case of trouble, please check there if your issue has already been reported. If you spotted it first, help us to smash it by providing a detailed and welcomed feedback.
Do not contact contributors directly about support or help with technical issues.
- Camptocamp
- ACSONE SA/NV
- Guewen Baconnier <[email protected]>
- Stéphane Bidoul <[email protected]>
- Matthieu Dietrich <[email protected]>
- Jos De Graeve <[email protected]>
- David Lefever <[email protected]>
- Laurent Mignon <[email protected]>
- Laetitia Gangloff <[email protected]>
- Cédric Pigeon <[email protected]>
- Tatiana Deribina <[email protected]>
- Souheil Bejaoui <[email protected]>
- Eric Antones <[email protected]>
- Simone Orsi <[email protected]>
This module is maintained by the OCA.
OCA, or the Odoo Community Association, is a nonprofit organization whose mission is to support the collaborative development of Odoo features and promote its widespread use.
Current maintainer:
This module is part of the OCA/queue project on GitHub.
You are welcome to contribute. To learn how please visit https://odoo-community.org/page/Contribute.