Jobtastic makes your user-responsive long-running Celery jobs totally awesomer. Celery is the ubiquitous python job queueing tool and jobtastic is a python library that adds useful features to your Celery tasks. Specifically, these are features you probably want if the results of your jobs are expensive or if your users need to wait while they compute their results.
Jobtastic gives you goodies like:
- Easy progress estimation/reporting
- Job status feedback
- Helper methods for gracefully handling a dead task broker
(
delay_or_eageranddelay_or_fail) - Super-easy result caching
- Thundering herd avoidance
- Integration with a celery jQuery plugin for easy client-side progress display
- Memory leak detection in a task run
Make your Celery jobs more awesome with Jobtastic.
If you have user-facing tasks for which a user must wait, you should try Jobtastic. It's great for:
- Complex reports
- Graph generation
- CSV exports
- Any long-running, user-facing job
You could write all of the stuff yourself, but why?
- Install gcc and the python C headers so that you can build psutil.
On Ubuntu, that means running:
$ sudo apt-get install build-essential python-dev python2.7-dev python3.5-dev rabbitmq-server
On OS X, you'll need to run the "XcodeTools" installer.
-
Get the project source and install it
$ pip install jobtastic
Let's take a look at an example task using Jobtastic:
from time import sleep
from jobtastic import JobtasticTask
class LotsOfDivisionTask(JobtasticTask):
"""
Division is hard. Make Celery do it a bunch.
"""
# These are the Task kwargs that matter for caching purposes
significant_kwargs = [
('numerators', str),
('denominators', str),
]
# How long should we give a task before assuming it has failed?
herd_avoidance_timeout = 60 # Shouldn't take more than 60 seconds
# How long we want to cache results with identical ``significant_kwargs``
cache_duration = 0 # Cache these results forever. Math is pretty stable.
# Note: 0 means different things in different cache backends. RTFM for yours.
def calculate_result(self, numerators, denominators, **kwargs):
"""
MATH!!!
"""
results = []
divisions_to_do = len(numerators)
# Only actually update the progress in the backend every 10 operations
update_frequency = 10
for count, divisors in enumerate(zip(numerators, denominators)):
numerator, denominator = divisors
results.append(numerator / denominator)
# Let's let everyone know how we're doing
self.update_progress(
count,
divisions_to_do,
update_frequency=update_frequency,
)
# Let's pretend that we're using the computers that landed us on the moon
sleep(0.1)
return resultsThis task is very trivial, but imagine doing something time-consuming instead of division (or just a ton of division) while a user waited. We wouldn't want a double-clicker to cause this to happen twice concurrently, we wouldn't want to ever redo this work on the same numbers and we would want the user to have at least some idea of how long they'll need to wait. Just by setting those 3 member variables, we've done all of these things.
Basically, creating a Celery task using Jobtastic is a matter of:
- Subclassing
jobtastic.JobtasticTask - Defining some required member variables
- Writing your
calculate_resultmethod (instead of the normal Celeryrun()method) - Sprinkling
update_progress()calls in yourcalculate_result()method to communicate progress
Now, to use this task in your Django view, you'll do something like:
from django.shortcuts import render_to_response
from my_app.tasks import LotsOfDivisionTask
def lets_divide(request):
"""
Do a set number of divisions and keep the user up to date on progress.
"""
iterations = request.GET.get('iterations', 1000) # That's a lot. Right?
step = 10
# If we can't connect to the backend, let's not just 500. k?
result = LotsOfDivisionTask.delay_or_fail(
numerators=range(0, step * iterations * 2, step * 2),
denominators=range(1, step * iterations, step),
)
return render_to_response(
'my_app/lets_divide.html',
{'task_id': result.task_id},
)The my_app/lets_divide.html template will then use the task_id
to query the task result all asynchronous-like
and keep the user up to date with what is happening.
For Flask, you might do something like:
from flask import Flask, render_template
from my_app.tasks import LotsOfDivisionTask
app = Flask(__name__)
@app.route("/", methods=['GET'])
def lets_divide():
iterations = request.args.get('iterations', 1000)
step = 10
result = LotsOfDivisionTask.delay_or_fail(
numerators=range(0, step * iterations * 2, step * 2),
denominators=range(1, step * iterations, step),
)
return render_template('my_app/lets_divide.html', task_id=result.task_id)"But wait, Wes. What the heck do those member variables actually do?" You ask.
Firstly. How the heck did you know my name?
And B, why don't I tell you!?
This is key to your caching magic. It's a list of 2-tuples containing the name of a kwarg plus a function to turn that kwarg in to a string. Jobtastic uses these to determine if your task should have an identical result to another task run. In our division example, any task with the same numerators and denominators can be considered identical, so Jobtastic can do smart things.
significant_kwargs = [
('numerators', str),
('denominators', str),
]If we were living in bizzaro world, and only the numerators mattered for division results, we could do something like:
significant_kwargs = [
('numerators', str),
]Now tasks called with an identical list of numerators will share a result.
This is the max number of seconds for which Jobtastic will wait for identical task results to be determined. You want this number to be on the very high end of the amount of time you expect to wait (after a task starts) for the result. If this number is hit, it's assumed that something bad happened to the other task run (a worker failed) and we'll start calculating from the start.
These let you tweak the default behavior.
Most often, you'll just be setting the cache_duration
to enable result caching.
If you want your results cached,
set this to a non-negative number of seconds.
This is the number of seconds for which identical jobs
should try to just re-use the cached result.
The default is -1,
meaning don't do any caching.
Remember,
JobtasticTask uses your significant_kwargs to determine what is identical.
This is an optional string used to represent tasks
that should share cache results and thundering herd avoidance.
You should almost never set this yourself,
and instead should let Jobtastic use the module.class name.
If you have two different tasks that should share caching,
or you have some very-odd cache key conflict,
then you can change this yourself.
You probably don't need to.
Set this value to monitor your tasks for any runs that increase the memory usage by more than this number of Megabytes (the SI definition). Individual task runs that increase resident memory by more than this threshold get some extra logging in order to help you debug the problem. By default, it logs the following via standard Celery logging:
- The memory increase
- The memory starting value
- The memory ending value
- The task's kwargs
You then grep for Jobtastic:memleak memleak_detected in your logs
to identify offending tasks.
If you'd like to customize this behavior,
you can override the warn_of_memory_leak method in your own Task.
Other than tweaking the member variables, you'll probably want to actually, you know, do something in your task.
This is where your magic happens. Do work here and return the result.
You'll almost definitely want to
call update_progress periodically in this method
so that your users get an idea of for how long they'll be waiting.
This is the guy you'll want to call to provide nice progress feedback and estimation.
In your calculate_result,
you'll want to periodically make calls like:
self.update_progress(work_done, total_work_to_do)Jobtastic takes care of handling timers to give estimates, and assumes that progress will be roughly uniform across each work item.
Most of the time,
you really don't need ultra-granular progress updates
and can afford to only give an update every N items completed.
Since every update would potentially hit your
CELERY_RESULT_BACKEND,
and that might cause a network trip,
it's probably a good idea to use the optional update_frequency argument
so that Jobtastic doesn't swamp your backend
with updated estimates no user will ever see.
In our division example, we're only actually updating the progress every 10 division operations:
# Only actually update the progress in the backend every 10 operations
update_frequency = 10
for count, divisors in enumerate(zip(numerators, denominators)):
numerator, denominator = divisors
results.append(numerator / denominator)
# Let's let everyone know how we're doing
self.update_progress(count, divisions_to_do, update_frequency=10)Sometimes,
your Task Broker
just up and dies
(I'm looking at you, old versions of RabbitMQ).
In production,
calling straight up delay() with a dead backend
will throw an error that varies based on what backend you're actually using.
You probably don't want to just give your user a generic 500 page
if your broker is down,
and it's not fun to handle that exception every single place
you might use Celery.
Jobtastic has your back.
Included are delay_or_eager and delay_or_fail methods
that handle a dead backend
and do something a little more production-friendly.
Note: One very important caveat with JobtasticTask is that
all of your arguments must be keyword arguments.
Note: This is a limitation of the current significant_kwargs implementation,
and totally fixable if someone wants to submit a pull request.
If your broker is behaving itself,
this guy acts just like delay().
In the case that your broker is down,
though,
it just goes ahead and runs the task in the current process
and skips sending the task to a worker.
You get back a nice shiny EagerResult object,
which behaves just like the AsyncResult you were expecting.
If you have a task that realistically only takes a few seconds to run,
this might be better than giving yours users an error message.
This method uses async_or_eager() under the hood.
Like delay_or_eager,
this helps you handle a dead broker.
Instead of running your task in the current process,
this actually generates a task result representing the failure.
This means that your client-side code can handle it
like any other failed task
and do something nice for the user.
Maybe send them a fruit basket?
For tasks that might take a while
or consume a lot of RAM,
you're probably better off using this than delay_or_eager
because you don't want to make a resource problem worse.
This method uses async_or_fail() under the hood.
This is a version of delay_or_eager() that exposes the calling signature
of apply_async().
This is a version of delay_or_fail() that exposes the calling signature
of apply_async().
That's all well and good on the server side, but the biggest benefit of Jobtastic is useful user-facing feedback. That means handling status checks using AJAX in the browser.
The easiest way to get rolling is to use our sister project, jquery-celery. It contains jQuery plugins that help you:
- Poll for task status and handle the result
- Display a progress bar using the info from the
PROGRESSstate. - Display tabular data using DataTables.
If you want to roll your own, the general pattern is to poll a URL (such as the django-celery task_status view ) with your taskid to get JSON status information and then handle the possible states to keep the user informed.
The jquery-celery jQuery plugin might still be useful as reference, even if you're rolling your own. In general, you'll want to handle the following cases:
Your task is still waiting for a worker process. It's generally useful to display something like "Waiting for your task to begin".
Your task has started and you've got a JSON object like:
{
"progress_percent": 0,
"time_remaining": 300
}progress_percent is a number between 0 and 100.
It's a good idea to give a different message if the percent is 0,
because the time remaining estimate might not yet be well-calibrated.
time_remaining is the number of seconds estimated to be left.
If there's no good estimate available, this value will be -1.
You've got your data. It's time to display the result.
Something went wrong and the worker reported a failure. This is a good time to either display a useful error message (if the user can be expected to correct the problem), or to ask the user to retry their task.
There are occasions where requesting the task status itself might error out. This isn't a reflection on the worker itself, as it could be caused by any number of application errors. In general, you probably want to try again if this happens, but if it persists, you'll want to give your user feedback.
We use tox
to run our tests against various combinations
of python/Django/Celery.
We only officially support
the combinations listed in our .travis.yml file,
but we're working on
(Issue 33)
supporting everything defined in tox.ini.
Until then,
you can run tests against supported combos with:
$ pip install tox
$ tox -e py27-django1.8.X-djangocelery3.1.X-celery3.1.X
Our test suite currently only tests usage with Django, which is definitely a bug. Especially if you use Jobtastic with Flask, we would love a pull request.
Have tasks whose duration is difficult to estimate or that doesn't have smooth progress? JobtasticMixins to the rescue!
JobtasticMixins provides an AVGTimeRedis mixin
that stores duration date in a Redis backend.
It then automatically uses this stored historical data
to calculate an estimate.
For more details,
check out JobtasticMixins
on github.
Yes. Increasingly so.
Jobtastic is currently known to work with Django 1.6+ and Celery 3.1.X The goal is to support those versions and newer. Please file issues if there are problems with newer versions of Django/Celery.
At this time of this writing, the latest supported version of kombu with celery 4.x is 4.0.2. This is due to an issue with invalid or temporarily broken brokers with the newer versions of kombu.
Also, RabbitMQ should be running in the background while running tests.
Previously,
if you were using Flask instead of Django,
then the only currently-supported way to work with Jobtastic
was with Memcached as your CELERY_RESULT_BACKEND.
Thanks to @rhunwicks this is no longer the case!
A cache is now selected with the following priority:
- If the Celery appconfig has a
JOBTASTIC_CACHEsetting and it is a valid cache, use it - If Django is installed, then:
- If the setting is a valid Django cache entry, then use that.
- If the setting is empty use the default cache
- If Werkzeug is installed, then:
- If the setting is a valid Celery Memcache or Redis Backend, then use that.
- If the setting is empty and the default Celery Result Backend is Memcache or Redis, then use that
This project isn't affiliated with the awesome folks at the Celery Project (unless having a huge crush counts as affiliation). It's a library that the folks at PolicyStat have been using internally and decided to open source in the hopes it is useful to others.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent