Massively Scaled High Performance Web Services with PHP

Massively Scaled High Performance Web Services
with PHP
Demin Yin
https://joind.in/talk/e81e7

• Staff Engineer and Technical Lead on Glu Mobile’s Design Home team.
• Playing with PHP since 2000.
• Worked for a diverse group of companies in China and US, including eBay, Visa and Glu.
• Focus on building high performance, secure web services.
• Born in China, worked in U.S., live in Canada.
• On twitter: @deminy
• On Github: https://github.com/deminy
About Me
2

● An Overview on Design Home and the Inbox Service
● The Inbox Service: Key Improvements Made
1. Web Server: Dockerized PHP 7 containers
2. HTTP Processing: Background processing, HTTP compression, and more
3. Data Storage: Doing NoSQL with Redis and Couchbase
4. Hardware and Network: Migrating to Amazon Web Services
● Next
● Q&A
Agenda
3

● Next
● Q&A
Agenda
4

• Launched in Q4 2016.
• #1 Top Free and #7 Top Grossing game in the U.S. App Store for iPhone.
• #1 Top Free and #23 Top Grossing game in the U.S. Google Play Store.
• 1M+ daily active users; 40M votes daily; 1 billion designs to date.
• 100,000 API calls per minute; 60 56 ms response time on average.
Design Home:
Top-ranked Mobile Game with PHP Backend
5

Inbox
Localization
Receipt Validation
Moderation
Email
E-commerce
Event Logging
Main
Service
……
Database
Client
Client
Client
Client
Client
Client
Client
Design Home: List of Backend Services
6

● REST APIs with different CRUD endpoints to manage inbox messages.
● PHP 7, Nginx, Composer, NoSQL, Docker, AWS, Development Tools, ……
● NoSQL: Redis + Couchbase.
● One of the busiest microservices.
Average Requests: 14k rpm
Peak Requests: 34k rpm
Response Time: less than 9ms
The Inbox Service: Technical Overview
7

The Inbox Service: Technical Overview (cont.)
● PHP 7.
● Composer.
● Opcache.
● Data cache: APCu, distributed cache and chained cache.
● Tests in place.
● Unit tests.
● Blackbox tests, or feature tests, or functional tests.
● PHP The Right Way: Follow good coding practices including design patterns.
8

● Next
● Q&A
Agenda
9

Hardware & Network
● Migrate to Amazon ECS.
● Hardware upgrade.
● Move everything into VPC.
The Inbox Service: Key Improvements Made
HTTP Processing
● Use background processing.
● Use HTTP compression (gzip).
● Dynamic content caching.
Web Server Instances
● Decouple microservices.
● Dockerize the microservices.
● Development Tools Integration.
Data Storage
● NoSQL solutions.
● Fight against IO and disk space.
● Faster PHP drivers.
10

● Next
● Q&A
Agenda
11

1.1. Decouple Microservices
Inbox
Localization
Receipt Validation
Moderation
Email
E-commerce
Event Logging
Main
Service
……
Database
Client
Client
Client
Client
Client
Client
Client
12

• Improved and simplified development environments.
• Easier to upgrade web server and PHP environments.
• Easier to troubleshoot and debug issues.
1.2. Dockerized Microservices
13

1.2. Dockerized Microservices (cont.)
14
Base Images
HTTP/REST microservices
asynchronous microservices
job worker instances

1.2. Dockerized Microservices (cont.)
Specification:
1. Tags match with PHP versions.
2. Tag latest is never used.
3. Freeze tagged images once
used.
4. Images are built and tagged
manually.
Benefits:
1. Less efforts to build images for
different microservices.
2. Easy and safe to upgrade to
new versions of PHP.
3. Easier to fix security
vulnerabilities.
4. Easier to make improvements
across different microservices.
15
Base Images

● Error Handling, Error Reporting and System Monitoring
○ New Relic
○ Bugsnag
○ Amazon CloudWatch
● Security Check
○ SonarQube
○ SensioLabs Security Checker
● Debugging and Profiling: Blackfire, Xdebug, etc.
1.3. Development Tools Integration
16

● Next
● Q&A
Agenda
17

● Use cases: error handling, data reporting, outgoing emails, push notifications, etc.
● Common approaches:
● Execute external program.
● Register shutdown functions.
● Job server (Gearman, Resque, RabbitMQ, etc.).
● Function fastcgi_finish_request() under PHP-FPM.
● …...
2.1. Background Processing
exec("nohup ./http.sh > /dev/null &");
18

2.1. Background Processing (cont.)
Output: 1234 Output: 1
19

Average Response Time
Without background processing: 13 ms
With background processing: 9 ms
20

How to prevent and address failed operations properly?
● Basic data validation first.
● Exponential backoff.
● https://github.com/Crowdstar/exponential-backoff
● Proper error reporting.
21

For more details, please check
• Background Processing: https://github.com/Crowdstar/background-processing
• Technical Discussion: https://github.com/deminy/background-processing-in-php
22

2.2. HTTP Compression
gzip on;
24

2.2. HTTP Compression (cont.)
gzip on;
gzip_types text/plain text/css application/json text/xml application/xml application/xml+rss
text/javascript;
25

gzip on;
text/javascript;
gzip_proxied any;
26

gzip on;
text/javascript;
gzip_proxied any;
gzip_comp_level 5;
gzip_min_length 1280;
27

If HTTP header ‘Content-Length’ is not set in a PHP response, Nginx will always compress the
response even its length is less than the minimum length.
• Slim 3 always has HTTP header ‘Content-Length’ set by default.
• Laravel and Lumen don’t have HTTP header ‘Content-Length’ set.
28

A Lumen Middleware to Make HTTP Compression Work as Should
29

2.3. Dynamic Content Caching
Challenge: feed new data to the client.
Possible Solutions:
● Server push, but it’s not supported in HTTP 1 (It’s supported under HTTP 2 and the WebSocket API).
● HTTP Clients built on top of HTTP 1 have to hit servers again and again to check and fetch new data.
Our Approach:
● Data cache on the server side if OK.
● Use the Last-Modified and If-Modified-Since headers, and return HTTP 304 "Not Modified" if no need to
retransmit the requested resources.
30

2.3. Dynamic Content Caching (cont.)
Related HTTP headers:
● If-None-Match, If-Unmodified-Since
● Cache-Control
● Content-Location
● ETag
● Expires
● Vary
● …...
31

● Next
● Q&A
Agenda
32

● Lots of choices: Redis, Couchbase, Aerospike, MongoDB, …
● Our selection and implementation: Redis and Couchbase.
● Redis.
● Couchbase.
● Concerns with NoSQL
○ limitations on queries.
○ indexes.
○ serialization and compression.
3.1. NoSQL
33

• Data compression to reduce network IO and save disk space.
• Compress larger messages before saving.
• Couchbase encoder (serialization).
• Use message templates.
• Shorten field names.
• Add ttl (time to live) on certain messages (to expire messages).
3.2. Fight Against IO and Disk Space
34

• PhpRedis: the PHP extension for Redis.
• Predis: the Redis client written in PHP.
• Other drivers
• Package predis/predis-async: asynchronous (non-blocking) version of Predis built on
top of ReactPHP.
• Class SwooleCoroutineRedis: the asynchronous Redis Client from Swoole.
3.3. Faster PHP Drivers (Redis)
35

3.3. Faster PHP Drivers (Redis)
Before and after switching from Predis to PhpRedis
Average request time reduced from 16 ms to 11 ms
36

● Next
● Q&A
Agenda
37

4.1. Deployment with Amazon ECS
How did we do code deployment before?
38

Benefits:
• Zero downtime deployments.
• Quick rollback.
• ELB health checks.
• Autoscaling.
4.1. Deployment with Amazon ECS (cont.)
39

• Multiple deployment environments in AWS: test, dev, RC, staging, and production.
• Staging environment is helpful.
• to do some final tests before major production deployments.
• to reproduce and identify production issues.
4.1. Deployment with Amazon ECS (cont.)
40

4.2. Hardware Upgrade
Feb 2018: upgraded API servers with faster ones
(c5.xlarge: 4 vCPU, 8GB memory; Enhanced networking enabled)
41

• Place instances closer to each other.
• Move instances to same locations (Availability Zones).
• Use internal network instead of Internet to communicate.
4.3. Move Everything into VPC
42

4.3. Move Everything into VPC (cont.)
Jan 2019: upgraded API servers with faster ones that were closer to the database nodes (same zone)
In the main service, Couchbase operation time decreased to 5 ms
43

● Next
● Q&A
Agenda
44

There are always rooms to
push the limit of PHP
45

• PHP-FPM
• Asynchronous I/O
• Swoole
• Network Protocols
• HTTP/2
• Tars
Next
46

● Next
● Q&A
Agenda
47

Questions?
48

Thanks
49

Massively Scaled High Performance Web Services with PHP

Recommended

More Related Content

What's hot (20)

Similar to Massively Scaled High Performance Web Services with PHP (20)

Recently uploaded (19)

Massively Scaled High Performance Web Services with PHP