Anatomy of a Proxy Server

MaxScale Internals

Ivan Zoratti - Percona Live Europe - September 2015

Meet ScaleDB Inc.
• Database company based in
Menlo Park
• Developers of
IDC - Intelligent Data Cluster
and
UDE - Universal Data Engine
• Main focus on Streaming Data,
Time Series and Real Time
Analytics, all in one Engine
• Put it simply,
Big Data on MySQL
• Try the Free version: www.
2 scaledb.com
A Bit of “Externals”
What is a Database Proxy?
A Database Proxy is a specialised Proxy Server

4
What is a Database Proxy?
A Database Proxy is a specialised Proxy Server

• A Database Proxy is an intermediary for requests from client

applications or other databases
• A Database Proxy
• Receives a request
• Evaluates the request and acts accordingly
• Connects to a database, directs the request to a database server, discards the request, logs the
request etc.
• Follows up on future events associated to the request
• Receives data from a database server, prepares and send a response etc.

5
What is MaxScale?
MaxScale is an event-based, I/O dispatcher that acts as a
database proxy.
MaxScale can be used to capture, analyze, log, transform, reroute incoming requests from a frontend
(client) to a backend (server) and the outcoming results from the backend to the frontend

• Client / Frontend - One or more applications, web servers, database servers etc.
• Server / Backend - One or more database servers, proxy servers, message brokers etc.

6
MaxScale - Prehistory and History

MySQL
Proxy

7
MaxScale - Prehistory and History

MySQL
Proxy

8
MaxScale - Prehistory and History

Apache-based
MySQL prototype
Proxy

9
MaxScale - Prehistory and History

Apache-based
MySQL prototype
Proxy

10
MaxScale - Prehistory and History

Apache-based
MySQL prototype
Proxy

epoll architecture
11 & MaxScale 0.1
MaxScale Modern Days
• Active Project at MariaDB Corporation (renamed MariaDB MaxScale)
• https://github.com/mariadb-corporation/MaxScale
• Used in production in many companies (including Booking.com), MariaDB and
Percona customers
• More projects are on going at ScaleDB and other companies

Current version: 1.2

• Packaged for Ubuntu/Debian and CentOS/RH
• Running as a service
• Many modules already available

12
MaxScale Architecture and
Implementation
The MaxScale Pentagon

14
MaxScale Modules
• Routers: ReadConn Route, ReadWrite Split, Schema Router,
Master/Master
• Monitors: MySQL Replication, MySQL Cluster/NDB, Galera
• Filters: Firewall, Query Log, Query Routing, RabbitMQ, Query Re-
write, Tee, Query Top
• Authentication: MySQL Authentication (up to MySQL 5.6)
• Protocol: MySQL Client, MySQL Server

15
MaxScale and MySQL Replication

16
MaxScale and Galera

17
MaxScale Master/Master

18
MaxScale Cop :-)

19
MaxScale Binlog Server

20
MaxScale Internals
MaxScale Design

The Three Golden Rules

1. Never wait for a response
2. Network sockets must be non-
blocking
3. Do not use mutexes that cause
I/O threads queueing

22
MaxScale Design
• NOT a database
• “Object Orientated”
The Three Golden Rules • Pure multi-threading
1. Never wait for a response • Non Preemptive
2. Network sockets must be non- • Event-driven I/O model
blocking • Events are queued for
3. Do not use mutexes that cause processing by a thread
I/O threads queueing • Implementation of IOCP
• epoll abstraction
• Self contained, isolated
23
plugins
Event Model Implementation
• epoll Implementation of event abstraction
• poll_init > epoll_create
• poll_add_dcb
• poll_remove_dcb
• poll_waitevents > polling and event dispatching with timeout bias
• pollq > choose an event to be executed by a thread
• poll_add_event_to_dcb
• The event queue is implemented within the DCB structure (DCBEVENTQ)
• Next/Previous event list
• Set of pending events
• Set of processing event
• Processing flag
• Event queue lock
• Event process:
• Lock dcb->evq structure
• Copy pending events to processing events
• Set processing flag
24 • Unlock dcb->evq structure
Functions & Classes
• spinlock
• acquire - wait for spinlock to
be available
• acquire_nowait
• release
• Not to be used when
making system calls

25
Functions & Classes
• spinlock
• acquire - wait for spinlock to
be available
• acquire_nowait
• release
• Not to be used when
making system calls

26
Functions & Classes
• spinlock
• atomic_add
• Basis for spinlock
• Add signed value
and return the
previous value

27
Functions & Classes
• spinlock
• atomic_add
• GWBUF
• Sharing buffers between different
threads without copies
• Chopping buffers up without copy
• Creating lists of buffers for
read/write (scatter/gather approach)

28
DCB - Descriptor Control Block
• One DCB per connection
• DCB holds the connection state
• Registered with epoll
• Buffer queues
• DCB->writeq
• DCB->delayq
• DCB->dcb_readqueue
Operations
• connect - Create a connection to the specified server
using a specific protocol for the session
• read - Read data from a socket
• write - Data write entry point (generic)
• close - Remove the DCB from the poll set
• write_ready- EPOLLOUT handler for the socket,
indicates that the socket is ready to send more data
• error - EPOLLERR handler for the socket
• hangup - EPOLLHUP handler for the socket
• 29 accept - Accept handler for listener socket only
Plugins - An Example
Filter Plugin (QLA)

30
Plugins - An Example
Filter Plugin (QLA)

31
Plugins - An Example
Filter Plugin (QLA)

32
The Future of MaxScale
• Memory Allocation and optimisation
• Authentication Plugin
• Parser Plugin (and optimisation)
• Protocol Plugins
• Router Plugins

33
Thank You!
Images
● MySQL 2007 Logo - Courtesy of Oracle Corporation
● ScaleBase Logo - Courtesy of ScaleArc
● Jan Kneschke, Massimiliano Pinto, Massimo Brignoli, Mark Riddoch - Courtesy of Twitter and Percona
● Jean-Francois Gagne Presentation - http://www.slideshare.net/JeanFranoisGagn/binlog-servers-at-bookingcom?qid=1fdc1af3-5658-400e-
8a9c-964ad40319c1&v=qf1&b=&from_search=3
● Caffe Trieste - Christopher Michel - https://www.flickr.com/photos/cmichel67/17596540570/in/photostream/