Slide 1
Slide 1 text
Managing & Scaling
Asynchronous Workers
Justin Carmony - Lone Star PHP ‘14
( And Staying Sane! )
@JustinCarmony
Slide 2
Slide 2 text
Me
• Director of Development
@ Deseret Digital Media
• Utah PHP Usergroup
President
• I Make (and Break)
Web Stuff (~10 years)
• @JustinCarmony
[email protected]
Slide 3
Slide 3 text
This
Presentation
• Slides Posted Online
• Feel free to ask on-topic
question during presentation
• Q&A Session At the End
• Feel free to ask me any
questions afterwards
Slide 4
Slide 4 text
Warning:
This presentation contains materials that is
based off the opinions of the presenter. It is
not absolute truth (i.e. 1+1=2), and should
not be taken as such. Rather, it is just a
bunch of things I think are a good idea.
Slide 5
Slide 5 text
• Theory Behind Workers
• Why they can be difficult to manage
• Best Practices for Writing Workers
• Handling The Hiccups
Presentation Outline
Slide 6
Slide 6 text
The Theory
Of Workers
Slide 7
Slide 7 text
• You have a Job
• You put it in a queue
• Worker takes the Job from the queue
• Worker does the Job
• Repeat
The Theory of Workers
Slide 8
Slide 8 text
Job Queue
Worker
Done!
It’s Simple…
Right?
Slide 9
Slide 9 text
No content
Slide 10
Slide 10 text
Actually…
bit.ly/1hZ4JHl
Slide 11
Slide 11 text
Managing Workers
Can Be Complex
Slide 12
Slide 12 text
Why Workers Can Be Complex
Long Running
• Measured in Hours vs Milliseconds
• Most PHP isn’t written to be long
running
• Most Connections have timeouts
Keep The Running
• What Happens When a Worker
Dies?
• How to restart a worker?
Monitoring
• What is my worker doing?
• Is it frozen / hung?
• What happened to my Job?
Potentially Dangerous
• “My worker filled my disk with temp
files”
• “A bug accidentally deleted all of
photos.”
Slide 13
Slide 13 text
Best Practices for
Writing Workers
(aka how not to hate your life)
Slide 14
Slide 14 text
Simplicity
Slide 15
Slide 15 text
Avoid
Complexity
Slide 16
Slide 16 text
Use The
Right Tools
Slide 17
Slide 17 text
• Beanstalkd
Lightweight, Fast, Simple
• RabbitMQ
More Robust
• Redis
Very, very basic list queues
• Gearman
Full service Queue/Worker System
• Amazon SQS, IronMQ
Cloud-Based Servers that just Work
Use Better Tools - Queues
Better Queues
• Relational Databases
(MySQL, Postgres, etc)
• Other Relational / Document Data
Stores (Mongo, Couch)
• Flat Files"
• Any crazy hacky thing people come
up with
Poor Queues
Personal "
Recommendation
Slide 18
Slide 18 text
• Very Fast - In-Memory by default
• Dynamic queues which they call “tubes”
• Bury jobs that have an error, kick them back into queue when ready
• Jobs re-enter queue if not finished by timeout
• Can put jobs in via a delay
Why Beanstalkd
Slide 19
Slide 19 text
• PHP 5.5
• Beanstalkd - The queue
• Redis - Status Information
• Pheanstalk - The PHP Library to communicate w/ Beanstalkd
• Predis - Library to talk to Redis
Our Example Technology Stack
Slide 20
Slide 20 text
Typical Queue Life Cycle
Client
put
Job Worker
reserve
Poof!
delete
Slide 21
Slide 21 text
Possible Beanstalkd Life Cycle
Poof!
DELAYED
put with
delay
(time passes)
READY RESERVED
reserve
put
release with delay
release
delete
BURIED
bury
kick
Poof!
delete
Slide 22
Slide 22 text
Queueing Jobs
Slide 23
Slide 23 text
Queueing Jobs - JSON for Data
{!
"job": "process_image",!
"data": {!
"source_url": "http://example.com/image.jpg",!
"save_as": "/uploads/some_file.jpg",!
"width": 100,!
"height": 100!
},!
"success": {!
"callback": "http://api.example.com/callback/success/process_image"!
},!
"error": {!
"bury": true,!
"callback": "http://api.example.com/callback/error/process_image"!
}!
}!
Common Data Structure for all Jobs
Data For Specific Job
Slide 24
Slide 24 text
• Create a Queue (or “Tube”) for each type of “job”
• Similar Jobs only in a Queue
• Use Priorities Sparingly, Keep it a simple FIFO: First In First Out
• Workers can listen got multiple tubes, use different tubes for
different priorities.
• “Dedicated Workers” for particular tubes (i.e priority or bulk)
Naming Queues
Slide 25
Slide 25 text
• email.priority — For things like reset password, account creation
• email.regular — For normal emails like friend notifications
• email.bulk — For mass emails
Naming Queues - Examples
Slide 26
Slide 26 text
• Data - The data for the job
• Priority - Lowest to Highest Priority
(default: 1024, min: 0, max: 4,294,967,295)
• Delay - # of Seconds before Job is ready to be reserved
(default: 0)
• TTR - How long to wait for job to be completed
(default: 60)
Options for Queueing in Beanstalkd
Better to be Explicit & Always Set These Values
Slide 27
Slide 27 text
Queueing Job - Job Object
job = $job_name;!
$this->data = $data;!
$this->success = $success;!
$this->error = $error;!
}!
}!
Slide 28
Slide 28 text
Queueing Job - Sending Job
"[email protected]",!
"from" => "[email protected]",!
"subject" => "So-and-so wants to friend your face!",!
"body" => "...."!
];!
!
$job = new \DDM\Awesome\Job('send_email', $data);!
!
!
$pheanstalk!
->useTube('email.regular')!
->put(json_encode($job), 500, 0, 120);!
Slide 29
Slide 29 text
Writing Workers
Slide 30
Slide 30 text
• Ran from the Command Line
(i.e. php path/to/worker.php)"
• Bootstrapped & Auto-loaded
(its like it is a real part of your application)
• Blocked Listening to Queue
(If queue is empty, wait for new job)"
• Takes CLI Arguments
(php worker.php --logLevel=debug --run=60)"
• Can Run Multiple on Same Server
(php worker.php 1; php worker.php 2)
Use Better Tools - Workers
Better Workers
• Ran some other way
(i.e. curl web request via cron job)
• One-off random scripts"
• Polling Queue via a Loop"
• Complex"
Poor Workers
Slide 31
Slide 31 text
Benefits:
• No Complexity of a Web Server
• Easier to Keep Running Continuously
• Easier to Manage Logs & Monitor
Common Problems:
• PHP CLI config different from
Apache’s or PHP-FPM’s
Run From the Command Line
Slide 32
Slide 32 text
• Use Modern PHP Bootstrapping & Autoloading
• Workers should exist as a part of your Application
• The PHP file to run your worker should be minimal
• Use all the same best practices (writing tests, using OOP, etc)
• Worker code will evolve over time like your Web App’s code, treat it
as a first-class citizen and not some quick one-off
Bootstrapped & Autoloaded
Slide 33
Slide 33 text
Use Coding
Best Practices
Slide 34
Slide 34 text
I’m Serious!
Slide 35
Slide 35 text
No content
Slide 36
Slide 36 text
Slide 37
Slide 37 text
build(new ImageWorker('worker_1'));!
!
// Have the worker Run!
$worker->run();!
Worker Script - Good Example
Autoloading
Bootstrapping
Create & Setup Worker
Run
Namespaces!
Slide 38
Slide 38 text
Setup
Slide 39
Slide 39 text
• Typically by default PHP will timeout after a
certain amount of time.
• Use set_time_limit(0); to disable this limit.
PHP Timeouts
Slide 40
Slide 40 text
• I recommend assigning each worker two values:
• workerId — A unique identifier for that worker (i.e.
srv1_imgworker2) that is persistent through instances.
• instanceHash — A unique hash (i.e. random md5) for
that particular run. Useful for telling when a worker
restarts.
Setting Up Your Worker
Slide 41
Slide 41 text
Settings IDs & Instance Hash
class Worker!
{!
public $workerId = '';!
public $instanceHash = '';!
!
public function __constructor($worker_id)!
{!
$this->workerId = $worker_id;!
$this->instanceHash = md5(uniqid(rand(), true));!
}!
}!
Slide 42
Slide 42 text
• Connections dying / timing out is the
#1 cause for errors in workers!
• Check connections before each job.
(Example: $mysqli->ping())
• Close & Re-open connections for
infrequent Jobs
Creating Connections
Slide 43
Slide 43 text
Getting Jobs
From The Queue
Slide 44
Slide 44 text
The Old Way:
Polling
getJob();!
!
// Check to see if I got a job!
if($job)!
{!
// I did, yay!!
$job->doJob();!
}!
else!
{!
// Sleep for 5 seconds!
sleep(5);!
}!
// loop and do it all over again !
}!
Slide 45
Slide 45 text
The New Way:
Blocking
getJobOrWait(60);!
!
// Check to see if I got a job!
if($job)!
{!
// I did, yay!!
$job->doJob();!
}!
!
// loop and do it all over again !
}!
Slide 46
Slide 46 text
Beanstalkd Example
getPheanstalk();!
!
$pheanstalk->ignore('default')!
->watch('mail.priority')!
->watch('mail.regular')!
->watch('mail.bulk');!
!
while($this->run)!
{!
$job = $pheanstalk->reserve(60 * 5);!
if($job)!
{!
/* ... do work ... */!
}!
}!
}!
!
/* ... */!
}!
Get Beanstalkd Connection
Ignore Default Tube
Wait for Job for 5 Minutes
Slide 47
Slide 47 text
Doing the Work
Slide 48
Slide 48 text
Processing a Job - Poor Example
public function Work()!
{!
/* ... */!
$job = $pheanstalkd->reserve(60 * 5);!
!
if($job)!
{!
$job_data = json_decode($job->getData());!
if($job_data->job == 'send_mail')!
{!
$mail = new MailClass();!
$mail->setSubject($job_data->data->subject);!
/* ... more mail code ... */!
}!
else if($job_data->job == 'another_job')!
{!
/* ... even more code ... */!
}!
else if($job_data->job == 'even_another_job')!
{!
/* ... even MORE! code ... */!
}!
}!
/* ... */!
}!
Multiple If / Else If Statements
Doing the Job In-Line
Even More Code
This gets long & messy quick
Slide 49
Slide 49 text
Processing a Job - Better Example
$job = $pheanstalk->reserve(60 * 5);!
!
if($job)!
{!
$success = false;!
!
try {!
$success = $this->processJob($job); !
} catch (\Exception $ex)!
{!
$success = false;!
}!
!
if($success)!
{!
$pheanstalk->delete($job);!
}!
else!
{!
$pheanstalk->bury($job);!
}!
}!
Slide 50
Slide 50 text
Processing a Job - Better Example
class Worker {!
!
public $available_jobs = [!
"send_mail" => "\\DDM\\Awesome\\Job\\SendMailJob"!
];!
!
/* .... */!
!
function processJob($job_data)!
{!
$succes = false;!
!
if(isset($this->available_jobs[$job_data->job]))!
{!
$job_class = $this->available_jobs[$job_data->job];!
$class = new $job_class($this->getPheanstalk(), $job_data);!
$success = $class->process();!
}!
!
return $success;!
} !
!
/* .... */!
}!
Slide 51
Slide 51 text
Logging
Slide 52
Slide 52 text
• PSR-3 is a standard interface for logging defined by FIG
• Keep it simple, use Monolog
https://github.com/Seldaek/monolog
• Use the different levels: Debug, Info, Notice, Warning, Error, Critical,
Alert, Emergency
Logging - PSR-3 Logger
Slide 53
Slide 53 text
Monolog Example
// Monolog!
!
use Monolog\Logger;!
use Monolog\Handler\StreamHandler;!
use Monolog\Handler\StreamHandler;!
!
// Create the logger!
$logger = new Logger('worker');!
// Now add some handlers!
$logger->pushHandler(new StreamHandler(‘/tmp/workers.log', Logger::DEBUG));!
$logger->pushHandler(new RedisHandler(Predis\Client("tcp://localhost:6379"), 'ddm.awesome.worker.log'));!
!
$worker->setLogger($logger);!
Slide 54
Slide 54 text
Monolog Example
$job = $pheanstalk->reserve(60 * 5);!
!
if($job)!
{!
$this->logger->debug('Job found');!
$success = false;!
!
try {!
$success = $this->processJob($job); !
} catch (\Exception $ex)!
{!
$success = false;!
}!
!
if($success)!
{!
$this->logger->debug('Job Finished, Deleting');!
$pheanstalk->delete($job);!
}!
else!
{!
$this->logger->warning('Job failed, burying');!
$pheanstalk->bury($job);!
}!
}!
Slide 55
Slide 55 text
• Set up different logging handlers for different levels.
• Keep Performance & Volume in mind
• Examples:
• StdOut / StdErr - Debug (aka All)
• File - Notices & Higher
• Redis - Warnings & Higher
• Email - Critical & Higher
Multiple Logging Sources
Slide 56
Slide 56 text
Reporting Status
Slide 57
Slide 57 text
• Insight into what workers are
doing is crucial.
• Important for debugging &
monitoring.
• Hook your Monitoring &
Alerting tools to these
statuses.
Reporting Status - General Concept
• Thing to Report On:
• Runtime
• Last Heartbeat
• Currently Doing
• # of Jobs
• # of Errors
Slide 58
Slide 58 text
• Store in something fast, scalable.
• High Volume of Reads & Writes
• My recommendation: Redis
• Do Not Use:
• Primary Database
• Avoid Data Stores using Replication
Reporting Status - Storage
Slide 59
Slide 59 text
• A Heartbeat is a regular
“I’m still running!”
• Typically ran often before,
during, and after a job.
Reporting Heartbeat
while($this->run)!
{!
$this->heartbeat('idle');!
$job = $pheanstalk->reserve(60 * 5);!
!
if($job)!
{!
$success = false;!
!
$job_data = json_decode($job->getData());!
!
$this->heartbeat('processing_job_'.$job_data->job);!
!
try {!
$success = $this->processJob($job); !
} catch (\Exception $ex)!
{!
$success = false;!
}!
!
if($success)!
{!
$pheanstalk->delete($job);!
}!
else!
{!
$pheanstalk->bury($job);!
}!
}!
}!
Slide 60
Slide 60 text
Reporting Heartbeat
public function heartbeat($status)!
{!
$data = [!
'timestamp' => now(),!
'status' => $status,!
'workerId' => $this->workerId,!
'instanceHash' => $this->instanceHash,!
'jobs' => $this->jobCount,!
'errors' => $this->errorCount!
];!
!
$predis = $this->getPredisClient();!
$predis->hset('workers.heartbeat', $this->workerId, json_encode($data));!
}!
Slide 61
Slide 61 text
Shutdown / Restart
Slide 62
Slide 62 text
• For the love of everything good, do not just exit() / die();
• $worker->run = false; when I want to stop the worker.
• Allows for cleanup and a clean stop.
• Make maintaining your workers so, much, easier.
Shutting Down
Slide 63
Slide 63 text
Shutting Down
while($this->run)!
{!
!
$job = $pheanstalk->reserve(60 * 5);!
!
if($job)!
{!
try {!
$this->processJob($job); !
} catch (\Exception $ex)!
{!
$this->run = false;!
}!
!
}!
}!
Slide 64
Slide 64 text
Controlling Workers
Slide 65
Slide 65 text
• Create a queue for each worker based off ID.
(example: system.worker1, system.worker2, etc)
• Send jobs for that worker, examples: shutdown, wait
• Useful for after code-deploy to reload changes
• Send timeouts to do “rolling restarts”
• Send w/ lowest priority (0) to ensure these are ran first.
Controlling Workers - System Queue
Slide 66
Slide 66 text
Keeping Workers
Running
Slide 67
Slide 67 text
• Linux based tool for keeping processes running on a server.
• Very easy to install, setup, and use.
• Will restart workers when they exit
• Configurable restarts & failure conditions for restarts
• Run multiple instances of the same command.
Supervisor
Slide 68
Slide 68 text
Supervisor
[program:worker]!
command=php /path/to/worker.php %(process_num)d!
process_name=%(program_name)s_%(process_num)d!
stdout_logfile=/var/log/%(program_name)s.log!
redirect_stderr=true!
stdout_capture_maxbytes=512MB!
stdout_logfile_backups=3!
numprocs=10!
numprocs_start=0!
autostart=true!
autorestart=true!
file: /etc/supervisor/conf.d/worker.conf
Slide 69
Slide 69 text
Whew…
Slide 70
Slide 70 text
Handling the Hiccups
Slide 71
Slide 71 text
Permissions
Slide 72
Slide 72 text
• Please, please, please … do NOT run as root!
• My preference: run as same user as the web user (i.e. www-data)
• You can create a separate user for workers
• Caveats for separate user: shared cache permissions
Permissions
Slide 73
Slide 73 text
Threading
• Threading is Awesome
• If you like to have bugs that
take down servers
• And can be a total pain to
track down
• Personal Opinion: 99% of the
time overly complex vs
performance gains.
Slide 74
Slide 74 text
Worker
Pools
• Easy to Maintain
• Easy to Scale
• Idle workers should really be
idle and not use resources
• Predictable to scale
Slide 75
Slide 75 text
Separate Server(s)
Slide 76
Slide 76 text
Confirming Jobs
Executed
Slide 77
Slide 77 text
• Create a record in Redis / DB
• Create Job
• Worker does job & updates
record
• Store Status Details in Record
Confirming Jobs Executed
• Details to Store:
• Job State
• Created Timestamp
• Last Update Timestamp
• Error Details
Slide 78
Slide 78 text
Job Order Dependencies
Slide 79
Slide 79 text
• Create Records to Track Job
• Pass “chain” of jobs with initial job
• Once first job finished, queue next job with rest of the chain
• Update Job Record along the way
Job Order Dependencies
Slide 80
Slide 80 text
Job Order Dependencies
{!
"job": "process_image",!
"data": {!
"source_url": "http://example.com/image.jpg"!
},!
"nextJobs": [!
{!
"queue": "email.regular",!
"job": {!
"job": "send_email",!
"data": {!
!
}!
}!
},!
{!
"queue": "email.regular",!
"job": {!
"job": "send_email",!
"data": {!
!
}!
}!
}!
]!
}!
{!
"job": "send_email",!
"data": {!
!
},!
"nextJobs": [!
{!
"queue": "email.regular",!
"job": {!
"job": "send_email",!
"data": {!
!
}!
}!
}!
]!
}!
Slide 81
Slide 81 text
Monitoring
Slide 82
Slide 82 text
• StatsD / Graphite - Health
• Nagios - Alerts
Monitoring
Monitoring Tools
• # of Workers Running
• # of Jobs Executed
• Alert if Jobs are failing to
start
• Timings on how long Jobs
take to run
What to Monitor
Slide 83
Slide 83 text
Final Thoughts
Slide 84
Slide 84 text
Keep It Simple
Slide 85
Slide 85 text
Use Best Practices
Slide 86
Slide 86 text
Well Written Workers
Scale Well
Slide 87
Slide 87 text
Questions?
Slide 88
Slide 88 text
Thank You
Twitter: @JustinCarmony
Email: [email protected]
Web: justincarmony.com
Please Leave Feedback: https://joind.in/10809