Scaling PHP apps

9 January 2018
Scaling PHP apps

CTO @
website: madisoft.it
tech blog: labs.madisoft.it

It’s from experience.
There are no lessons.

Nuvola
scuoladigitale.info
● > 3M HTTP requests / day
● > 1000 databases
● ~ 0.5T mysql data
● ~ 180M query / day
● ~ 105M of media files
● ~ 18T of media files
● From ~5k to ~200k sessions in 5 minutes

Scalability
Your app is scalable if it can adapt to
support an increasing amount of data
or a growing number of users.

“But… I don’t have an increasing load”
(http://www.freepik.com/free-photos-vectors/smile - Smile vector designed by Freepik)

“Scalability doesn’t matter to you.”

“I do have an increasing load”

Scaling… what?
PHP code?
Database?
Sessions?
Storage?
Async tasks?
Logs?

Scaling is about
app architecture

App architecture
How can you scale your web server if
you put everything inside?
Database, user files, sessions, ...

App architecture / Decouple
● Decouple services
● Service: do one thing and do it well

6 main areas
1. web server
2. sessions
3. database
4. filesystem
5. async tasks
6. logging
There are some more (http caching, frontend, etc): next talk!

Web server
A single big web server
(scale up)

Web server
Many small web servers
(scale out)

Web server
behind a load balancer

Web server
behind a load balancer,
inside an auto-scaling group

Web server
NGINX + php-fpm
PHP 7
(Symfony 4)

Web server
Avoid micro-optimization
Single quotes vs double quotes?
It doesn’t matter

Web server / Cache
PHP CACHE
APPLICATION CACHE
DOCTRINE CACHE

Web server / PHP cache
OPcache
OPcache improves PHP performance by storing precompiled script bytecode
in shared memory, thereby removing the need for PHP to load and parse
scripts on each request.
http://php.net/manual/en/intro.opcache.php

Web server / PHP cache
OPcache
Bytecode caching
opcache.enable = On
opcache.validate_timestamps = 0
Need to manually reset OPcache on deploy!
https://tideways.io/profiler/blog/fine-tune-your-opcache-configuration-to-avoid-caching-suprises

PHP code / Application cache
● Put application cache in ram
● Use cache warmers during deploy
releaseN/var/cache -> /var/www/project/cache/releaseN
“/etc/fstab”
tmpfs /var/www/project/cache tmpfs size=512m

PHP code / Doctrine cache
● Configure Doctrine to use cache
● Disable Doctrine logging and profiling on prod
doctrine.orm.default_metadata_cache:
type: apcu
doctrine.orm.default_query_cache:
type: apcu
doctrine.orm.default_result_cache:
type: apcu

PHP code / Cache
DISK I/O ~ 0%

PHP code / Profiling
Blackfire
New Relic
Tideways

PHP code / Recap
● Easy
● No need to change your PHP code
● It’s most configuration and tuning
● You can do one by one and measure how it affects performance
● Need to monitor and profile: New Relic for PHP
● Don’t waste time on micro-optimization
Take away: use cache!

Sessions
● Think session management as a service
● Use centralized Memcached or Redis (Ec2
or ElasticCache on AWS)
● Avoid sticky sessions (load balancer set up)

Session / Memcached
No bundle required
https://labs.madisoft.it/scaling-symfony-sessions-with-memcached

Session / Redis
https://github.com/snc/SncRedisBundle
https://labs.madisoft.it/scaling-symfony-sessions-with-redis

Session / Redis
config.yml
framework:
session:
handler_id: snc_redis.session.handler

Session / Redis
Bundle config
snc_redis:
clients:
session_client:
dsn: '%redis_dsn_session%'
logging: false # https://github.com/snc/SncRedisBundle/issues/161
type: phpredis
session:
client: session_client
locking: false
prefix: session_prefix_
ttl: '%session_ttl%'

Session / Redis
parameters.yml
redis_db: 3
redis_dsn_session: 'redis://%redis_ip%/%redis_db%'
redis_ip: redis-cluster.my-lan.com
session_ttl: 86400

Session / Recap
● Very easy
● No need to change your PHP code
● Redis better than Memcached: it has persistence and many other features
● Let AWS scale for you and deal with failover and sysadmin stuff
Take away: use Redis

Database
Aka “The bottleneck”

Database
If you need data integrity do
not replace your SQL db with
NOSQL to scale

Database
If dbsize < 10 GB
dont_worry();

Database / Big db problems
● Very slow backup. High lock time
● If mysql crashes, restart takes time
● It takes time to download and restore in dev
● You need expensive hardware (mostly RAM)

Database / Short-term solutions
Use a managed db service like AWS RDS
● It scales for you
● It handles failover and backup for you
But:
● It’s expensive for big db
● Problems are only mitigated but they are still there

Database / Long-term solutions
Sharding

Database / Sharding
Split a single big db into
many small dbs
(multi-tenant)

Database / Sharding
● Very fast backup. Low lock time
● If mysql crashes, restart takes little time
● Fast to download and restore in dev
● No need of expensive hardware
● You arrange your dbs on many machines

Database / Sharding
● How can Symfony deal with them?
● How to execute a cli command on one of them?
● How to apply a migration (ie: add column) to 1000 dbs?
● …...

Database / Sharding
Doctrine DBAL & ORM

Database / Sharding
Define a DBAL connection and a ORM
entity manager for each db
https://symfony.com/doc/current/doctrine/multiple_entity_managers.html

Database / Sharding
doctrine:
orm:
entity_managers:
global:
connection: global
shard1:
connection: shard1
shard2:
connection: shard2
doctrine:
dbal:
connections:
global:
…..
shard1:
……
shard2:
…...
default_connection: global

Database / Sharding
This works for few dbs
(~ <5)

Database / Sharding
Doctrine sharding
http://docs.doctrine-project.org/projects/doctrine-dbal/en/latest/reference/sharding.html

Database / Doctrine sharding
● Suited for multi-tenant applications
● Global database to store shared data (ie: user data)
● Need to use uuid
http://docs.doctrine-project.org/projects/doctrine-dbal/en/latest/reference/sharding.html

Database / Sharding
Configuration
doctrine:
dbal:
default_connection: global
connections:
default:
shard_choser_service: vendor.app.shard_choser
shards:
shard1:
id: 1
host / user / dbname
shard2:
id: 2
host / user / dbname

Database / Sharding
ShardManager Interface
$shardManager = new PoolingShardManager();
$currentCustomerId = 3;
$shardManager->selectShard($currentCustomerId);
// all queries after this call hit the shard where customer
// with id 3 is on
$shardManager->selectGlobal();
// the global db is selected

Database / Sharding
● It works but it’s complex to be managed
● No documentation everywhere
● Need to manage shard configuration: adding a new
shard?
● Need to parallelize shard migrations: Gearman?
● Deal with sharding in test environment

Database / Recap
● NOSQL is not used to scale SQL: they have different purposes. You can
use both.
● Sharding is difficult to implement
● Need to change your code
● Short-term solution is to use AWS to leverage some maintenance
● Doctrine ORM sharding works well but you need to write code and
wrappers. Best suited for multi-tenant apps
● When it’s done, you can scale without any limit
Take away: do sharding if your REALLY need it

Filesystem
Users upload files: documents, media, etc
How to handle them?

Filesystem
● Need of filesystem abstraction
● Use external object storage like S3
● Avoid using NAS: it’s tricky to be set-up correctly

Filesystem / Abstraction
● FlysystemBundle
● KnpGaufretteBundle
https://github.com/1up-lab/OneupFlysystemBundle

https://github.com/1up-lab/OneupFlysystemBundle
● AWS S3
● Dropbox
● FTP
● Local filesystem
● ...

Configuration
oneup_flysystem:
adapters:
s3_adapter:
awss3v3:
client: s3_client
bucket: "%s3_bucket%"
oneup_flysystem:
adapters:
local_adapater:
local:
directory: ‘myLocalDir’

Configuration
prod.yml
oneup_flysystem:
filesystems:
my_filesystem:
adapter: s3_adapter
dev.yml
oneup_flysystem:
filesystems:
my_filesystem:
adapter: local_adapter

Usage
// LeagueFlysystemFilesystemInterface
$filesystem = $container->get(‘oneup_flysystem.my_filesystem’);
$path = ‘myFilePath’;
$filesystem->has($path);
$filesystem->read($path);
$filesystem->write($path, $contents);

Filesystem / Recap
● Easy
● Need to change your PHP code
● Ready-made bundles
● Avoid local filesystem and NAS
Take away: use FlystemBundle with S3

Async tasks
Sync vs Async
Can synchronous systems scale?

Async tasks
We need a system able to manage some
queues, deliver, store and route messages.
And some PHP code able to consume those
messages...

RabbitMQ
Open source message broker
https://www.rabbitmq.com

RabbitMQ
How can PHP talk to RabbitMQ?
AMQP library for PHP
https://github.com/php-amqplib/php-amqplib

RabbitMQ
RabbitMQ lets you scale the queue

RabbitMQ
You can scale consumers: how?

RabbitMQ
Putting some machines (containers) inside an
auto-scaling group!
They can scale based on:
● Hardware parameters: cpu / memory
● Number of queue items
● Add your custom metrics!

Async tasks / Recap
● You need an external system and some new machines / containers
● Need to change your PHP code
● Ready-made bundles and libraries
● Avoid blocking sync tasks. Put the message on the queue and move on.
Take away: use RabbitMQ with auto-scaling
consumers

Logging
Logs have to be collected and managed in a
centralized way...

Logging
● You need an external system
● Take a look at managed ones: loggly.com, logz.io, scalyr.com
● Don’t need to change your PHP code
● You can’t avoid it in a distributed system
Take away: use a managed service

Scaling / Recap
● Sessions and filesystem: easy. Do it
● PHP code: not difficult. Think of it. Save money.
● Database: very hard. Think a lot
● Async tasks: think of it if you have many of them.
● Logging: necessary. Easy to implement if you choose a
managed service

We are always looking for
talented people...!

Scaling PHP apps

Related slideshows

More Related Content

What's hot

What's hot (20)

Similar to Scaling PHP apps

Similar to Scaling PHP apps (20)

Recently uploaded

Recently uploaded (20)

Scaling PHP apps