qless-with-throttles 0.12.0a7

Redis-based Queue Management

qless

This is a fork of seomoz/qless-py with support for throttles. Qless is a powerful Redis-based job queueing system inspired by resque, but built on a collection of Lua scripts, maintained in the qless-core repo. Be sure to check the changelog below.

Philosophy and Nomenclature

A job is a unit of work identified by a job id or jid. A queue can contain several jobs that are scheduled to be run at a certain time, several jobs that are waiting to run, and jobs that are currently running. A worker is a process on a host, identified uniquely, that asks for jobs from the queue, performs some process associated with that job, and then marks it as complete. When it's completed, it can be put into another queue.

Jobs can only be in one queue at a time. That queue is whatever queue they were last put in. So if a worker is working on a job, and you move it, the worker's request to complete the job will be ignored.

A job can be canceled, which means it disappears into the ether, and we'll never pay it any mind every again. A job can be dropped, which is when a worker fails to heartbeat or complete the job in a timely fashion, or a job can be failed, which is when a host recognizes some systematically problematic state about the job. A worker should only fail a job if the error is likely not a transient one; otherwise, that worker should just drop it and let the system reclaim it.

Features

1. Jobs don't get dropped on the floor -- Sometimes workers drop jobs. Qless automatically picks them back up and gives them to another worker
2. Tagging / Tracking -- Some jobs are more interesting than others. Track those jobs to get updates on their progress. Tag jobs with meaningful identifiers to find them quickly in the UI.
3. Job Dependencies -- One job might need to wait for another job to complete
4. Stats -- qless automatically keeps statistics about how long jobs wait to be processed and how long they take to be processed. Currently, we keep track of the count, mean, standard deviation, and a histogram of these times.
5. Job data is stored temporarily -- Job info sticks around for a configurable amount of time so you can still look back on a job's history, data, etc.
6. Priority -- Jobs with the same priority get popped in the order they were inserted; a higher priority means that it gets popped faster
7. Retry logic -- Every job has a number of retries associated with it, which are renewed when it is put into a new queue or completed. If a job is repeatedly dropped, then it is presumed to be problematic, and is automatically failed.
8. Web App -- With the advent of a Ruby client, there is a Sinatra-based web app that gives you control over certain operational issues
9. Scheduled Work -- Until a job waits for a specified delay (defaults to 0), jobs cannot be popped by workers
10. Recurring Jobs -- Scheduling's all well and good, but we also support jobs that need to recur periodically.
11. Notifications -- Tracked jobs emit events on pubsub channels as they get completed, failed, put, popped, etc. Use these events to get notified of progress on jobs you're interested in.

Interest piqued? Then read on!

Installation

Install from pip:

pip install qless-py

Alternatively, install qless-py from source by checking it out from github, and checking out the qless-core submodule:

git clone git://github.com/seomoz/qless-py.git
cd qless-py
# qless-core is a submodule
git submodule init
git submodule update
sudo python setup.py install

Business Time!

You've read this far -- you probably want to write some code now and turn them into jobs. Jobs are described essentially by two pieces of information -- a class and data. The class should have static methods that know how to process this type of job depending on the queue it's in. For those thrown for a loop by this example, it's in reference to a South Park episode where a group of enterprising gnomes set on world domination through three steps: 1) collect underpants, 2) ? 3) profit!

# In gnomes.py
class GnomesJob:
    # This would be invoked when a GnomesJob is popped off the 'underpants' queue
    @staticmethod
    def underpants(job):
        # 1) Collect Underpants
        ...
        # Complete and advance to the next step, 'unknown'
        job.complete('unknown')

    @staticmethod
    def unknown(job):
        # 2) ?
        ...
        # Complete and advance to the next step, 'profit'
        job.complete('profit')

    @staticmethod
    def profit(job):
        # 3) Profit
        ...
        # Complete the job
        job.complete()

This makes it easy to describe how a GnomesJob might move through a pipeline, first in the 'underpants' step, then 'unknown', and lastly 'profit.' Alternatively, you can define a single method process that knows how to complete the job, no matter what queue it was popped from. The above is just meant as a convenience for pipelines:

# Alternative gnomes.py
class GnomesJob:
    # This method would be invoked at every stage
    @staticmethod
    def process(job):
        if job['queue'] == 'underpants':
            ...
            job.complete('underpants')
        elif job['queue'] == 'unknown':
            ...
            job.complete('profit')
        elif job['queue'] == 'profit':
            ...
            job.complete()
        else:
            job.fail('unknown-stage', 'What what?')

Jobs have user data associated with them that can be modified as it goes through a pipeline. In general, you should make this data a dictionary, in which case it's accessible through __getitem__ and __setitem__. Otherwise, it's accessible through job.data. For example, you might update the data...

@staticmethod
def underpants(job):
    # Record how many underpants we collected
    job['collected'] = ...

@staticmethod
def unknown(job):
    # Make some decision based on how many we've collected.
    if job['collected'] ...:
        ...

Great! With all this in place, let's put them in the queue so that they can get run

import qless
# Connecting to localhost on 6379
client = qless.Client()
# Connecting to a remote machine
client = qless.Client('redis://foo.bar.com:1234')

Now, reference a queue, and start putting your gnomes to work:

queue = client.queues['underpants']

import gnomes
for i in range(1000):
    queue.put(gnomes.GnomesJob, {})

Alternatively, if the job class is not importable from where you're adding jobs, you can use the full path of the job class as a string:

...
for i in range(1000):
    queue.put('gnomes.GnomesJob', {})

By way of a quick note, it's important that your job class can be imported -- you can't create a job class in an interactive prompt, for example. You can add jobs in an interactive prompt, but just can't define new job types.

Running

All that remains is to have workers actually run these jobs. This distribution comes with a script to help with this:

qless-py-worker -q underpants -q unknown -q profit

This script actually forks off several subprocesses that perform the work, and the original process keeps tabs on them to ensure that they are all up and running. In the future, the parent process might also perform other sanity checks, but for the time being, it's just that the process is still alive. You can specify the host and port you want to use for the qless server as well:

qless-py-worker --host foo.bar --port 1234 ...

In the absence of the --workers argument, qless will spawn as many workers as there are cores on the machine. The interval specifies how often to poll in seconds) for work items. Future versions may have a mechanism to support blocking pop.

qless-py-worker --workers 4 --interval 10

Because this works on a forked process model, it can be convenient to import large modules before subprocesses are forked. Specify these with --import:

qless-py-worker --import my.really.bigModule

Filesystem

Previous versions of qless-py included a feature to have each worker process run in its own sandbox directory. We've removed this feature because since greenlets can't run in their own directory, the 'regular' and greenlet workers behave differently.

In lieu of this behavior, each child process runs in its own sandboxed directory and each job is given a sandbox attribute which is the name of a directory for the sole use of that job. It's guaranteed to be clean by the time the job is performed, and it cleaned up afterwards.

For example, if you invoke:

qless-py-worker --workers 4 --greenlets 5 --workdir foo

Then four child processes will be spawned using the directories:

foo/qless-py-workers/sandbox-{0,1,2,3}

The jobs run by the greenlets in the first process are given their own sandboxes of the form:

foo/qless-py-workers/sandbox-0/greenlet-{0,1,2,3,4}

Gevent

Some jobs are I/O-bound, and might want to, say, make use of a greenlet pool. If you have a class where you've, say, monkey-patched socket, you can ask qless to create a pool of greenlets to run you job inside each process. To run 5 processes with 50 greenlets each:

qless-py-worker --workers 5 --greenlets 50

Signals

With a worker running, you can send signals to child processes to:

• USR1 - Get the current stack trace in that worker
• USR2 - Enter a debugger in that worker

So, for example, if one of the worker child processes is PID 1234, then you can invoke kill -USR1 1234 to get the backtrace in the logs (and console output).

Resuming Jobs

This is an experimental feature, but you can start workers --resume flag to have the worker begin its processing with the jobs it left off with. For instance, during deployments, it's common to restart the worker processes, and the --resume flag has the worker first perform a check with qless to see which jobs it had last been running (and still has locks for).

This flag should be used with some caution. In particular, if two workers are running with the same worker name, then this should not be used. The reason is that through the qless interface, it's impossible to differentiate the two, and currently-running jobs may be confused with jobs that were simply dropped when the worker was stopped.

Debugging / Developing

Whenever a job is processed, it checks to see if the file in which your job is defined has been updated since its last import. If it has, it automatically reimports it. We think of this as a feature.

With this in mind, when I start a new project and want to make use of qless, I first start up the web app locally (see qless for more), take a first pass, and enqueue a single job while the worker is running:

# Supposing that I have /my/awesome/project/awesomeproject.py
# In one terminal...
qless-py-worker --path /my/awesome/project --queue foo --workers 1 --interval 10 --verbose

# In another terminal...
>>> import qless
>>> import awesomeproject
>>> qless.Client().queues['foo'].put(awesomeproject.Job, {'key': 'value'))

From there, I watch the output on the worker, adjust my job class, save it, watch again, etc., but without restarting the worker -- in general it shouldn't be necessary to restart the worker.

Internals and Additional Features

While in many cases the above is sufficient, there are also many cases where you may need something more. Hopefully after this section many of your questions will be answered.

Priority

Jobs can optionally have priority associated with them. Jobs of equal priority are popped in the order in which they were put in a queue. The higher the priority, the sooner it will be processed. If, for example, you get a new job to collect some really valuable underpants:

queue.put(qless.gnomes.GnomesJob, {'address': '123 Brief St.'}, priority = 10)

You can also adjust a job's priority while it's waiting:

job = client.jobs['83da4d32a0a811e1933012313b062cf1']
job.priority = 25

Scheduled Jobs

Jobs can also be scheduled for the future with a delay (in seconds). If for example, you just learned of an underpants heist opportunity, but you have to wait until later:

queue.put(qless.gnomes.GnomesJob, {}, delay=3600)

It's worth noting that it's not guaranteed that this job will run at that time. It merely means that this job will only be considered valid after the delay has passed, at which point it will be subject to the normal constraints. If you want it to be processed very soon after the delay expires, you could also boost its priority:

queue.put(qless.gnomes.GnomesJob, {}, delay=3600, priority=100)

Recurring Jobs

Whether it's nightly maintainence, or weekly customer updates, you can have a job of a certain configuration set to recur. Recurring jobs still support priority, and tagging, and are attached to a queue. Let's say, for example, I need some global maintenance to run, and I don't care what machine runs it, so long as someone does:

client.queues['maintenance'].recur(myJob, {'tasks': ['sweep', 'mop', 'scrub']}, interval=60 * 60 * 24)

That will spawn a job right now, but it's possible you'd like to have it recur, but maybe the first job should wait a little bit:

client.queues['maintenance'].recur(..., interval=86400, offset=3600)

You can always update the tags, priority and even the interval of a recurring job:

job = client.jobs['83da4d32a0a811e1933012313b062cf1']
job.priority = 20
job.tag('foo', 'bar')
job.untag('hello')
job.interval = 7200

These attributes aren't attached to the recurring jobs, per se, but it's used as the template for the job that it creates. In the case where more than one interval passes before a worker tries to pop the job, more than one job is created. The thinking is that while it's completely client-managed, the state should not be dependent on how often workers are trying to pop jobs.

# Recur every minute
queue.recur(..., {'lots': 'of jobs'}, 60)
# Wait 5 minutes
len(queue.pop(10))
# => 5 jobs got popped

Configuration Options

You can get and set global (read: in the context of the same Redis instance) configuration to change the behavior for heartbeating, and so forth. There aren't a tremendous number of configuration options, but an important one is how long job data is kept around. Job data is expired after it has been completed for jobs-history seconds, but is limited to the last jobs-history-count completed jobs. These default to 50k jobs, and 30 days, but depending on volume, your needs may change. To only keep the last 500 jobs for up to 7 days:

client.config['jobs-history'] = 7 * 86400
client.config['jobs-history-count'] = 500

Tagging / Tracking

In qless, 'tracking' means flagging a job as important. Tracked jobs have a tab reserved for them in the web interface, and they also emit subscribable events as they make progress (more on that below). You can flag a job from the web interface, or the corresponding code:

client.jobs['b1882e009a3d11e192d0b174d751779d'].track()

Jobs can be tagged with strings which are indexed for quick searches. For example, jobs might be associated with customer accounts, or some other key that makes sense for your project.

queue.put(qless.gnomes.GnomesJob, {'tags': 'aplenty'}, tags=['12345', 'foo', 'bar'])

This makes them searchable in the web interface, or from code:

jids = client.jobs.tagged('foo')

You can add or remove tags at will, too:

job = client.jobs['b1882e009a3d11e192d0b174d751779d']
job.tag('howdy', 'hello')
job.untag('foo', 'bar')

Job Dependencies

Jobs can be made dependent on the completion of another job. For example, if you need to buy eggs, and buy a pan before making an omelete, you could say:

eggs_jid = client.queues['buy_eggs'].put(myJob, {'count': 12})
pan_jid  = client.queues['buy_pan' ].put(myJob, {'coating': 'non-stick'})
client.queues['omelete'].put(myJob, {'toppings': ['onions', 'ham']}, depends=[eggs_jid, pan_jid])

That way, the job to make the omelete can't be performed until the pan and eggs purchases have been completed.

Notifications

Tracked jobs emit events on specific pubsub channels as things happen to them. Whether it's getting popped off of a queue, completed by a worker, etc. The jist of it goes like this, though:

def callback(evt, jid):
    print '%s => %s' % (jid, evt)

from functools import partial


for evt in ['canceled', 'completed', 'failed', 'popped', 'put', 'stalled', 'track', 'untrack']:
    client.events.on(evt, partial(callback, evt))
client.events.listen()

If you're interested in, say, getting growl or campfire notifications, you should check out the qless-growl and qless-campfire ruby gems.

Retries

Workers sometimes die. That's an unfortunate reality of life. We try to mitigate the effects of this by insisting that workers heartbeat their jobs to ensure that they do not get dropped. That said, qless will automatically requeue jobs that do get 'stalled' up to the provided number of retries (default is 5). Since underpants profit can sometimes go awry, maybe you want to retry a particular heist several times:

queue.put(qless.gnomes.GnomesJob, {}, retries=10)

Pop

A client pops one or more jobs from a queue:

# Get a single job
job = queue.pop()
# Get 20 jobs
jobs = queue.pop(20)

Heartbeating

Each job object has a notion of when you must either check in with a heartbeat or turn it in as completed. You can get the absolute time until it expires, or how long you have left:

# When I have to heartbeat / complete it by (seconds since epoch)
job.expires_at
# How long until it expires
job.ttl

If your lease on the job will expire before you have a chance to complete it, then you should heartbeat it to make sure that no other worker gets access to it. Or, if you are done, you should complete it so that the job can move on:

# I call stay-offsies!
job.heartbeat()
# I'm done!
job.complete()
# I'm done with this step, but need to go into another queue
job.complete('anotherQueue')

Stats

One of the selling points of qless is that it keeps stats for you about your underpants hijinks. It tracks the average wait time, number of jobs that have waited in a queue, failures, retries, and average running time. It also keeps histograms for the number of jobs that have waited x time, and the number that took x time to run.

Frankly, these are best viewed using the web app.

Lua

Qless is a set of client language bindings, but the majority of the work is done in a collection of Lua scripts that comprise the core functionality. These scripts run on the Redis 2.6+ server atomically and allow for portability with the same functionality guarantees. Consult the documentation for qless-core to learn more about its internals.

Web App

Qless also comes with a web app for administrative tasks, like keeping tabs on the progress of jobs, tracking specific jobs, retrying failed jobs, etc. It's available in the qless library as a mountable Sinatra app. The web app is language agnostic and was one of the major desires out of this project, so you should consider using it even if you're not planning on using the Ruby client.

Changelog

Things that have changed over time.

v0.10.0

The major change was the switch to unified qless. This change is semi-incompatibile. In particular, it changes the job history format but the new version knows how to convert the old format forward. Upgrades to your workers should be made from the end of pipelines towards the start. It will also be necessary to upgrade your qless-web install if you're using it.

• Preempts workers running jobs for which they've lost their lock
• Improved coverage (98%, up from 71%), all of which was worker code
• Debugging signals
• Resumable workers
• Redis URL interface. When specifying a qless client, the default is still to point to localhost:6379, but rather than specify host and port, you should provide a single host argument of a Redis URL format. For example, redis://user:auth@host:port/db. Many of these paremeters are optional, but it seems to be the convention recently.

Upgrading to qless-py 0.10.0

Some notes, instructions and potential road blocks to the upgrade. This version has much better coverage, and a few added features, including stalled job preemption, pauseable queues, unified sandboxing and the ability to use the cleaner web interface.

Road Blocks

Before we talk about how to install the updated client, here are a couple potential road blocks that will need to be addressed before you can make the switch.

Sandboxes

If you were using sandboxes (if using the non-greenlet client) and relying on using the current working directory as the sandbox, that interface has been done away with. The replacement is that each job comes with a sandbox attribute which is guaranteed to be a directory that exists and empty at the start of the job, and which is cleaned up after the job. It's a great place for temporary files. This only applies if you are running a qless-worker, and not if you are using the qless client directly to work on jobs.

The directories are made up of subdirectories under the directory provided as --workdir, defaulting to the current directory.

Client Rename

If you are using the qless client directly, all instances of qless.client will have to change to qless.Client. It was an unfortunate mistake that it was ever named client to begin with, but hopefully this change won't be painful.

Redis Server Spec

There was a feature request to be able to provide redis auth credentials, and rather than support any new attributes to the redis client that might come along, we'll now use a redis url.

For example:

# Instead of this
client = qless.Client(host='foo', port=6380)
# Now it's this
client = qless.Client(url='redis://foo:6380')

This allows users to provide auth, select a database, etc. Remember to change this in worker invocations and config files.

Installation

With an existing copy of qless-py checked out

# Get the most recent version
git fetch
git checkout v0.10.0

# Checkout, update and build the submodule
git submodule init
git submodule update
make -C qless/qless-core

# Install dependencies and then qless
sudo pip install -r requirements.txt
sudo python setup.py install