KEMBAR78
Scaling with Symfony - PHP UK | PDF
SCALING
WITH SYMFONY

+

=

PHP UK CONFERENCE 2014
HELLO PHPUK!
•

Ricard Clau, born and grown up in Barcelona, Spain

•

Software engineer at Hailo London

•

Symfony2 lover and PHP believer

•

Open-source contributor, sometimes I give talks

•

Twitter @ricardclau / Gmail ricard.clau@gmail.com
WHY THIS TALK?
•

Lots of PHP haters these days. I just don´t get it.

•

Some misconceptions about full-stack PHP frameworks

•

PHP is absolutely fine for most of your applications

•

But at some point you may need to replace some bits

•

DragonCity - Social game - 7.5M DAU - Symfony
AGENDA
•

Performance and scalability

•

Profiling and optimization

•

Storage: SQL vs NoSQL

•

Symfony2 is fast enough! Seriously!

•

Some war stories!

•

Almost everything applies to all PHP projects
Millions of
Users
Worldwide
24x7x365

CONCEPTS
Scalability, Availability, Load Balancing, Sharding
SCALABILITY
Which model would you choose?
AVAILABLE 24X7X365
We need to recover fast from any failure
LOAD BALANCING
It allows you to make changes without stopping your app
SHARDING
Luckily, you will get to a point when you need to do it...
And it can hurt if you have never considered it!
DON´T IF NOT NEEDED
•

Avoid premature optimizations

•

Microoptimizations are completely useless

•

Focus on impacting things

•

Incremental improvements

•

If the business is not working, it will be useless
BUT ACTIVELY PREVENT RISKS
•

Monitor servers: Ram, CPUs, Disk, network, idle
processes...

•

Try to engage stakeholders

•

When something surpasses half capacity, start
creating a B plan

•

Death of success is terrible
HANDS ON!
Where do we start from?
LOVE YOUR FRONTEND
Reduce # of requests and their size
SOME QUICK ADVICES
•

Fastest request -> No actual request

•

Icons in base64 inside your CSS

•

Minify CSS and JS, image size (jpegoptim, pngcrush, ...)

•

Use HTTP headers to cache at several layers

•

Network latency matters a lot

•

User experience takes benefit from all of these
SCALING PHP
Facebook, Wikipedia, Yahoo, Wordpress, Etsy, Zynga, YouPorn, Hailo,
Softonic, Emagister, Privalia, SocialPoint, Tuenti, ...
IS SCALING ACTUALLY EASY?
•

PHP bootstraps on every Request, so scaling is as “easy” as adding machines
behind a Load Balancer

•

But it makes it much slower than many other languages

•

Some things are shared among requests (sessions, storage, ...) but frameworks like
Symfony make it easier

•

PHP has been proof-tested in several BIG projects

•

Most of the times, problems are caused by the DB
LOOK BEYOND LAMP
•

Light-weight HTTP servers / CDNs for assets

•

Introduce reverse proxy cache like Varnish to cache pages (or ESI blocks)

•

Cache all the things (APC, Memcached, Redis, ...). 10 seconds cache -> Only 6
hits per minute!

•

RDBMS can deal with A LOT of data. But some problems are much better
solved with NoSQL databases

•

Sometimes other languages are more suitable for our problem
PHP PERFORMANCE
Opcode caches, composer, HHVM, quick wins
Upgrading from PHP5.3 will boost ~20-40%
APC: OUR OLD BESTIE
•

If you have ever worried about performance... you know it!

•

apc.stat -> Off (Apache reload to see changes)

•

apc.serializer -> igbinary (compact_strings -> Off)

•

apc.shm_size -> Check if 32Mb is enough

•

apc.write_lock -> Avoids cold cache issues

•

Check http://www.php.net/manual/en/apc.configuration.php
APC STATUS
http://svn.php.net/viewvc/pecl/apc/trunk/apc.php?view=markup
ZEND OPCACHE
•

Included in PHP5.5 standard distribution (valid with 5.2) will avoid weird
random issues like PHP5.4+APC

•

Better performance than APC in all tests

•

README https://github.com/zendtech/ZendOptimizerPlus

•

APC Userland Cache + Upload Hooks now live in APCU
(https://github.com/krakjoe/apcu)

•

Lots of companies are already using it!
OPCACHE STATUS
1 full-stack file, Rasmus style :)
(https://github.com/rlerdorf/opcache-status)
COMPOSER AUTOLOAD
•

You all already use Composer, right?

•

Default PSR-0 implementation is slow because we access
the FileSystem a lot

•

composer.phar install --optimize-autoloader (-o)
generates ClassMap

•

Same performance apc.stat = Off than
ApcUniversalClassLoader and less keys stored in APC
APC VS ZEND OPCACHE

http://www.ricardclau.com/2013/03/apc-vs-zend-optimizer-benchmarks-withsymfony2/
APC VS ZEND OPCACHE

http://www.ricardclau.com/2013/03/apc-vs-zend-optimizer-benchmarks-withsymfony2/
HHVM: THE FUTURE ENGINE?
•

Impressive progress in HHVM these past months

•

Most frameworks and libraries ~100% compatibility

•

Fewer memory, interesting speed improvements

•

We might finally have a language specification for PHP!

•

Specific HHVM features? -> Dangerous!

•

It will still not be comparable to compiled languages
FINE TUNING SYMFONY2
•

Disable unused bundles

•

SwiftMailer is quite slow, delegate to queue processes

•

If you can, use SubRequests with ESI and Varnish

•

Also check Twig C extension

•

You can implement a Cache Warmer for special needs

•

If using Apache, try app/console router:dump-apache
PROFILING
Diagnosing bottlenecks
PROFILING TOOLS
•

How often do you profile your code?

•

xDebug (Derick Rethans) and XHProf (Facebook)

•

Very convenient to install them in live servers and
activate to grab real data

•

Many things can also be discovered locally

•

Some tragedies only happen with live traffic
XHPROF (DEV VS PROD)
https://github.com/jonaswouters/XhprofBundle
HELLO WORLD
PHP and its frameworks...
Benchmarks comparing completely different things...
¿FRAMEWORKS == SLOW?
•

Some people think that writing their own
framework from scratch is a good idea

•

Also, that big frameworks use a lot of memory. (ZF1
and Symfony2 work with memory_limit 32M)

•

If this is your problem, PHP is not the solution

•

Don´t trust Hello World Performance Tests!

•

Nobody is better than a big community
TECHEMPOWER BENCHMARK
•

Symfony2 appears at the bottom

•

Comparison with light-weight frameworks is unfair.
Some incoherences (frameworks PHP >> PHP)

•

PHP will never be faster than a compiled language

•

In real apps, caching and DB optimizations equal things

•

Symfony2 roadmap improving performance
STORAGE
Use the right tool for the job. Silver bullets do not exist
RELATIONAL DBS
•

Mature technologies, good performance

•

Transactions: Atomicity, Consistency, Isolation, Durability

•

Foreign Keys, Joins, complex reporting queries

•

Millions of records without issues

•

We can use a NOSQLs key-value approach using
BLOB fields with serialized objects
NOSQL SOLUTIONS
•

Solutions to some concrete problems

•

CAP theorem (Consistency, Availability, Partition tolerance).
You can only get 2 of them... theoretically

•

Inmaturity in some of them. Complex to scale.

•

¿Big Data? 100M of records is NOT BIG DATA.

•

Redis, Cassandra, Riak and Solr work really well
QUEUE SYSTEMS
Read the docs, they are all completely different!
REAL-TIME DATA?
•

Try to delegate as much as possible to batch processes

•

Sending mails, external API requests, image
resizing, non-critical stats, ...

•

99% of stats don´t need to be real-time

•

60 seconds delay is mostly real-time!

•

Batch processes can be coded in different languages!
PHP NEEDS FRIENDS
Sometimes it is just not the best tool for our needs
PHP BAD SCENARIOS
•

24x7 running CLI daemon apps

•

Heavy math calculations, massive data processing,
programming contests

•

High concurrency apps with non-cachable requests

•

Threading, Forking, concurrent programming...

•

Writing DSLs
LOOK BEYOND PHP
•

JavaScript will be the king in the client side

•

In the server, Erlang and Go are growing adopters. And time will say about
Node.JS

•

Compiled languages like Go, C, C++ or Java will always be used in our stack.

•

JVM languages like Scala and Clojure are sexy now!

•

Learning other languages makes you a better engineer!
SYMFONY2
It is fast enough for most of what you want to do
WHY SYMFONY2?
•

Symfony2 is NOT the fastest “Hello world” framework...
but it is fast enough for most APIs and web applications

•

Big-community frameworks allow you to test any
technology in less than 5 minutes

•

A custom framework has usually higher costs and risks

•

LTS releases roadmap, stability, big projects are using it

•

Mature community, # of bundles
SOCIAL VIDEOGAME
https://apps.facebook.com/dragoncity/
(~7.5M DAU Facebook, iOS & Android Game)
DRAGONCITY: NUMBERS
•

~7.5 million daily users

•

~300 millions of registered users

•

Hundreds of millions of records for analytics generated daily

•

MySQL (32x2 + Analytics), Redis (~30), Cassandra (3)

•

HTTP requests are your progress -> No cacheable

•

Yes, most of the code is Symfony2! :)
POWERED BY SYMFONY2
@adanlobato showing happiness and pride!
6 DIGITS IN ANALYTICS
Yeah, still Symfony2! o/
WRAP-UP
•

PHP and Symfony2 work for most projects

•

Clever caching strategy is helpful

•

Profiling can make performance improve

•

The storage is almost always the bottleneck

•

Proper architecture design is crucial

•

PHP is not the best solution to some problems
QUESTIONS?
•

Twitter: @ricardclau

•

E-mail: ricard.clau@gmail.com

•

Github: https://github.com/ricardclau

•

Blog about PHP and Symfony2: http://www.ricardclau.com

•

Please, rate this talk at https://joind.in/talk/view/10697

Scaling with Symfony - PHP UK

  • 1.
  • 2.
    HELLO PHPUK! • Ricard Clau,born and grown up in Barcelona, Spain • Software engineer at Hailo London • Symfony2 lover and PHP believer • Open-source contributor, sometimes I give talks • Twitter @ricardclau / Gmail ricard.clau@gmail.com
  • 3.
    WHY THIS TALK? • Lotsof PHP haters these days. I just don´t get it. • Some misconceptions about full-stack PHP frameworks • PHP is absolutely fine for most of your applications • But at some point you may need to replace some bits • DragonCity - Social game - 7.5M DAU - Symfony
  • 4.
    AGENDA • Performance and scalability • Profilingand optimization • Storage: SQL vs NoSQL • Symfony2 is fast enough! Seriously! • Some war stories! • Almost everything applies to all PHP projects
  • 5.
  • 6.
  • 7.
    AVAILABLE 24X7X365 We needto recover fast from any failure
  • 8.
    LOAD BALANCING It allowsyou to make changes without stopping your app
  • 9.
    SHARDING Luckily, you willget to a point when you need to do it... And it can hurt if you have never considered it!
  • 10.
    DON´T IF NOTNEEDED • Avoid premature optimizations • Microoptimizations are completely useless • Focus on impacting things • Incremental improvements • If the business is not working, it will be useless
  • 11.
    BUT ACTIVELY PREVENTRISKS • Monitor servers: Ram, CPUs, Disk, network, idle processes... • Try to engage stakeholders • When something surpasses half capacity, start creating a B plan • Death of success is terrible
  • 12.
    HANDS ON! Where dowe start from?
  • 13.
    LOVE YOUR FRONTEND Reduce# of requests and their size
  • 14.
    SOME QUICK ADVICES • Fastestrequest -> No actual request • Icons in base64 inside your CSS • Minify CSS and JS, image size (jpegoptim, pngcrush, ...) • Use HTTP headers to cache at several layers • Network latency matters a lot • User experience takes benefit from all of these
  • 15.
    SCALING PHP Facebook, Wikipedia,Yahoo, Wordpress, Etsy, Zynga, YouPorn, Hailo, Softonic, Emagister, Privalia, SocialPoint, Tuenti, ...
  • 16.
    IS SCALING ACTUALLYEASY? • PHP bootstraps on every Request, so scaling is as “easy” as adding machines behind a Load Balancer • But it makes it much slower than many other languages • Some things are shared among requests (sessions, storage, ...) but frameworks like Symfony make it easier • PHP has been proof-tested in several BIG projects • Most of the times, problems are caused by the DB
  • 17.
    LOOK BEYOND LAMP • Light-weightHTTP servers / CDNs for assets • Introduce reverse proxy cache like Varnish to cache pages (or ESI blocks) • Cache all the things (APC, Memcached, Redis, ...). 10 seconds cache -> Only 6 hits per minute! • RDBMS can deal with A LOT of data. But some problems are much better solved with NoSQL databases • Sometimes other languages are more suitable for our problem
  • 18.
    PHP PERFORMANCE Opcode caches,composer, HHVM, quick wins Upgrading from PHP5.3 will boost ~20-40%
  • 19.
    APC: OUR OLDBESTIE • If you have ever worried about performance... you know it! • apc.stat -> Off (Apache reload to see changes) • apc.serializer -> igbinary (compact_strings -> Off) • apc.shm_size -> Check if 32Mb is enough • apc.write_lock -> Avoids cold cache issues • Check http://www.php.net/manual/en/apc.configuration.php
  • 20.
  • 21.
    ZEND OPCACHE • Included inPHP5.5 standard distribution (valid with 5.2) will avoid weird random issues like PHP5.4+APC • Better performance than APC in all tests • README https://github.com/zendtech/ZendOptimizerPlus • APC Userland Cache + Upload Hooks now live in APCU (https://github.com/krakjoe/apcu) • Lots of companies are already using it!
  • 22.
    OPCACHE STATUS 1 full-stackfile, Rasmus style :) (https://github.com/rlerdorf/opcache-status)
  • 23.
    COMPOSER AUTOLOAD • You allalready use Composer, right? • Default PSR-0 implementation is slow because we access the FileSystem a lot • composer.phar install --optimize-autoloader (-o) generates ClassMap • Same performance apc.stat = Off than ApcUniversalClassLoader and less keys stored in APC
  • 24.
    APC VS ZENDOPCACHE http://www.ricardclau.com/2013/03/apc-vs-zend-optimizer-benchmarks-withsymfony2/
  • 25.
    APC VS ZENDOPCACHE http://www.ricardclau.com/2013/03/apc-vs-zend-optimizer-benchmarks-withsymfony2/
  • 26.
    HHVM: THE FUTUREENGINE? • Impressive progress in HHVM these past months • Most frameworks and libraries ~100% compatibility • Fewer memory, interesting speed improvements • We might finally have a language specification for PHP! • Specific HHVM features? -> Dangerous! • It will still not be comparable to compiled languages
  • 27.
    FINE TUNING SYMFONY2 • Disableunused bundles • SwiftMailer is quite slow, delegate to queue processes • If you can, use SubRequests with ESI and Varnish • Also check Twig C extension • You can implement a Cache Warmer for special needs • If using Apache, try app/console router:dump-apache
  • 28.
  • 29.
    PROFILING TOOLS • How oftendo you profile your code? • xDebug (Derick Rethans) and XHProf (Facebook) • Very convenient to install them in live servers and activate to grab real data • Many things can also be discovered locally • Some tragedies only happen with live traffic
  • 30.
    XHPROF (DEV VSPROD) https://github.com/jonaswouters/XhprofBundle
  • 31.
    HELLO WORLD PHP andits frameworks... Benchmarks comparing completely different things...
  • 32.
    ¿FRAMEWORKS == SLOW? • Somepeople think that writing their own framework from scratch is a good idea • Also, that big frameworks use a lot of memory. (ZF1 and Symfony2 work with memory_limit 32M) • If this is your problem, PHP is not the solution • Don´t trust Hello World Performance Tests! • Nobody is better than a big community
  • 33.
    TECHEMPOWER BENCHMARK • Symfony2 appearsat the bottom • Comparison with light-weight frameworks is unfair. Some incoherences (frameworks PHP >> PHP) • PHP will never be faster than a compiled language • In real apps, caching and DB optimizations equal things • Symfony2 roadmap improving performance
  • 34.
    STORAGE Use the righttool for the job. Silver bullets do not exist
  • 35.
    RELATIONAL DBS • Mature technologies,good performance • Transactions: Atomicity, Consistency, Isolation, Durability • Foreign Keys, Joins, complex reporting queries • Millions of records without issues • We can use a NOSQLs key-value approach using BLOB fields with serialized objects
  • 36.
    NOSQL SOLUTIONS • Solutions tosome concrete problems • CAP theorem (Consistency, Availability, Partition tolerance). You can only get 2 of them... theoretically • Inmaturity in some of them. Complex to scale. • ¿Big Data? 100M of records is NOT BIG DATA. • Redis, Cassandra, Riak and Solr work really well
  • 37.
    QUEUE SYSTEMS Read thedocs, they are all completely different!
  • 38.
    REAL-TIME DATA? • Try todelegate as much as possible to batch processes • Sending mails, external API requests, image resizing, non-critical stats, ... • 99% of stats don´t need to be real-time • 60 seconds delay is mostly real-time! • Batch processes can be coded in different languages!
  • 39.
    PHP NEEDS FRIENDS Sometimesit is just not the best tool for our needs
  • 40.
    PHP BAD SCENARIOS • 24x7running CLI daemon apps • Heavy math calculations, massive data processing, programming contests • High concurrency apps with non-cachable requests • Threading, Forking, concurrent programming... • Writing DSLs
  • 41.
    LOOK BEYOND PHP • JavaScriptwill be the king in the client side • In the server, Erlang and Go are growing adopters. And time will say about Node.JS • Compiled languages like Go, C, C++ or Java will always be used in our stack. • JVM languages like Scala and Clojure are sexy now! • Learning other languages makes you a better engineer!
  • 42.
    SYMFONY2 It is fastenough for most of what you want to do
  • 43.
    WHY SYMFONY2? • Symfony2 isNOT the fastest “Hello world” framework... but it is fast enough for most APIs and web applications • Big-community frameworks allow you to test any technology in less than 5 minutes • A custom framework has usually higher costs and risks • LTS releases roadmap, stability, big projects are using it • Mature community, # of bundles
  • 44.
  • 45.
    DRAGONCITY: NUMBERS • ~7.5 milliondaily users • ~300 millions of registered users • Hundreds of millions of records for analytics generated daily • MySQL (32x2 + Analytics), Redis (~30), Cassandra (3) • HTTP requests are your progress -> No cacheable • Yes, most of the code is Symfony2! :)
  • 46.
    POWERED BY SYMFONY2 @adanlobatoshowing happiness and pride!
  • 47.
    6 DIGITS INANALYTICS Yeah, still Symfony2! o/
  • 48.
    WRAP-UP • PHP and Symfony2work for most projects • Clever caching strategy is helpful • Profiling can make performance improve • The storage is almost always the bottleneck • Proper architecture design is crucial • PHP is not the best solution to some problems
  • 49.
    QUESTIONS? • Twitter: @ricardclau • E-mail: ricard.clau@gmail.com • Github:https://github.com/ricardclau • Blog about PHP and Symfony2: http://www.ricardclau.com • Please, rate this talk at https://joind.in/talk/view/10697