简体中文README

PhxSQL is a high-availability and strong-consistency MySQL cluster built on Paxos and Percona.

Authors: Junchao Chen, Haochuan Cui, Duokai Huang, Ming Chen and Sifan Liu

Contact us: phxteam@tencent.com

#PhxSQL features:

• high availability by automatic failovers: the cluster works well when more than half of cluster nodes work and are interconnected.
• guarantee of data consistency among cluster nodes: replacing loss-less semi-sync between MySQL master and MySQL slaves with Paxos, PhxSQL ensures zero-loss binlogs between master and slaves and supports linearizable consistency, which is as strong as that of Zookeeper.
• complete compliance with MySQL and MySQL client: PhxSQL supports up to serializable isolation level of transaction.
• easy deployment and easy maintenance: PhxSQL, powered by in-house implementation of Paxos, has only 4 components including MySQL and doesn't depend on zookeeper or etcd for anything. PhxSql supports automated cluster membership hot reconfiguration.

This project includes

• Source codes
• Third party submodules
• Pre-compiled binaries for Ubuntu 64bit system.

Projects on which this project depends are also published by Tencent( phxpaxos, phxrpc, libco ). You can download or clone them with --recurse-submodule.

phxpaxos： http://github.com/Tencent/phxpaxos

phxrpc： http://github.com/Tencent/phxrpc

libco： http://github.com/Tencent/libco

Compilation of PhxSQL

If you prefer pre-compiled binaries, just skip this part.

Structure of PhxSQL Directories

• PhxSQL
  • phxsqlproxy
  • phxbinlogsvr
  • percona
  • phx_percona
    • plugin
    • phxsync_phxrpc
    • semisync
  • third_party
    • glog
    • leveldb
    • protobuf
    • phxpaxos
    • colib
    • phxrpc
  • tools
  • phxrpc_package_config

Introduction of Directories.

Name	Introduction
phxsqlproxy	surrogate between MySQL client and PhxSql
phxbinlogsvr	server for global binlog synchronization and storage, as well as management of mastership and membership
percona	Source code of percona5.6.31-77.0
phx_percona/plugin/phxsync_phxrpc	A plugin running in MySql that intercepts MySQL binlogs and forwards them to phxbinlogsvr
phx_percona/plugin/semisync	A semisync compatible with our modified plugin APIs of MySQL
third_party/glog	GLOG library
third_party/leveldb	LevelDB library
third_party/protobuf	Google Protobuf 3.0+ library
third_party/phxpaxos	PhxPaxos library
third_party/colib	Libco library
third_party/phxrpc	Phxrpc library

Preparation

Installation of third party libs

PhxSQL needs 6 third party libs(glog, leveldb, protobuf, phxpaxos, colib, phxrpc). Please install them in phxsql/third_party directory or just link to third_party.

NOTE: Please make sure -fPIC is added while executing configure in GLOG and Protobuf as well as specifying --prefix=/the/current/absolute/path.

For example: ./configure CXXFLAGS=-fPIC --prefix=/home/root/phxsql/third_party/glog.

Then download percona-server-5.6.31-77.0.tar.gz

Move percona-server-5.6\_5.6.31-77.0 to PhxSQL directory, rename or link as 'percona' (NOTE: Only percona-server-5.6_5.6.31-77.0 is available)

Preparation of installation enviroment

1. Execute ./autoinstall.sh && make && make install
2. Execute 'make package' to generate a tar.gz package so you can transfer to your target hosts.

(NOTE: We put the binaries in install_package/sbin, configuration files in install_package/tools/etc_template, install scripts in install_package/tools. The 'make package' command will pack 'install_package' into 'phxsql-$version.tar.gz'. Please specify -prefix=/the/path/you/want/to/install while executing ./autoinstall.sh)

Deployment of PhxSQL

Host requirements.

PhxSQL needs to run on more than 2 hosts. We suggest N >= 3 and N is an odd number, where N means the number of hosts.

Initialization of PhxSQL

1. Transfer phxsql.tar.gz to all of the hosts you want to install. Then do as the following steps:

   1. Execute tar -xvf phxsq.tar.gz .

   2. Enter phxsql/tools, Execute python install.py --help to get the help of installation.

      (For example:python2.7 install.py -i"your_inner_ip" -p 54321 -g 6000 -y 11111 -P 17000 -a 8001 -f/tmp/data/)

2. After executing 'install.py' on all the hosts, Execute './phxbinlogsvr_tools_phxrpc -f InitBinlogSvrMaster -h"ip1,ip2,ip3" -p 17000' in any one hosts. 17000 should be replaced with the port on which phxbinlogsvr is listening.

3. The cluster is active while a message shows master initialization is finished.

4. You can execute some SQLs to check the status of cluster through mysql -uroot -h"your_inner_ip" -P$phxsqlproxy_port

Simple tests.

1. Enter phxsql/tools/
2. Execute test_phxsql.sh phxsqlproxy_port ip1 ip2 ip3

Description of Configuration Files

PhxSQL have 3 configuration files in total.

1. my.cnf： The configuration of MySQL. Please modify it accroding to your own needs.

NOTE:Modify tools/etc_temlate/my.cnf before installation, Modify etc/my.cnf after installation

2. phxbinlogsvr.conf

Section name	Key name	comment
AgentOption	AgentPort	Port for the connection of binlogsvr and MySQL
	EventDataDir	Directory where to store the binlogsvr data
	MaxFileSize	File size per data of phxbinlogsvr, the unit is B
	MasterLease	Lease length of master, the unit is second
	CheckPointTime	The data before CheckPointTime will be deleted by phxbinlogsvr, but it will not be deleted if some other PhxSQL nodes have not learned yet, the unit is minute
	MaxDeleteCheckPointFileNum	The maximum number of files deleted each time by phxbinlogsvr
	FollowIP	Enabled if it is a follower node and will learn binlog from this FollowIP, this node will not vote
PaxosOption	PaxosLogPath	Directory where to store paxos data
	PaxosPort	Port for paxos to connect each other
	PacketMode	The maximum size of paxos log for PhxPaxos,1 means 100M, but the network timeout will be 1 minute, 0 means 50M and network timeout is 2s(changed in dynamic).
	UDPMaxSize	Our default network use udp and tcp combination, a message we use udp or tcp to send decide by a threshold. Message size under UDPMaxSize we use udp to send.
Server	IP	IP for phxbinlogsvr to listen
	Port	Port for phxbinlogsvr to listen
	LogFilePath	Directory to store log
	LogLevel	Log level of phxbinlogsvr

3. phxsqlproxy.conf

Section name	Key name	comment
Server	IP	IP for phxsqlproxy to listen
	Port	Port for phxsqlproxy to listen
	LogFilePath	Directory to store log
	LogLevel	Log level of phxbinlogsvr
	MasterEnableReadPort	Enable readonly-port in master node. If set to 0, master will forwarding readonly-port requests to one of slaves.
	TryBestIfBinlogsvrDead	After the local phxbinlogsvr is dead, phxsqlproxy will try to get master information from phxbinlogsvr on other machine, if this option set to 1.

PhxSQL Usasge

phxsqlproxy is the surrogate of PhxSQL, all requests will be sent to phxsqlproxy and then be forwarded to MySQL.

phxsqlproxy provides 2 different types of port for user.

Master Port( also called Read-Write Port )

It is the port configured in phxsqlproxy.conf. Every requests sent to this port will be forwarded to the master node to excute.

Slave Port( also called Read-Only Port )

It is (MasterPort + 1). You can also specify it by setting SlavePort = xxxxx in phxsqlproxy.conf.
Every requests will be executed on the local MySQL. A master node will make a redirection to another slave nodes if MasterEnableReadPort = 0 (this will save the CPU/IO resource for write requests)

SQL Command Execution

1. Using mysql -u$user -h$phxsqlproxyip -P$phxsqlproxyport -p$pwd to connect to phxsqlproxy
2. Execute SQL command.

$phxsqlproxyip can be any one IP of hosts in a clusters and $phxsqlproxyport can be MasterPort or SlavePort.

PhxSQL Management

PhxSQL provides a tool phxbinlogsvr_tools_phxrpc to help the mangerment of PhxSQL.

PhxSQL cluster needs 1 MySQL admin accounts and 1 synchronization account. The default admin account is (root, "" ), the default synchronization account is ( replica, replica123 ), They can be modified( and only be modifyed ) via phxbinlogsvr_tools_phxrpc. DON'T DO THIS MANUALLY.

Following is some commands you may used frequently.

phxbinlogsvr_tools -f GetMasterInfoFromGlobal -h <host> -p <port>

**Function:**Get the current master info from quorum nodes( IP and timeout ).

Arguments:

• Host: Any one IP of clusters nodes
• Port: Port which phxbinlogsvr is listening. like 17000

phxbinlogsvr_tools -f SetMySqlAdminInfo -h <host> -p <port> -u <admin username> -d <admin pwd> -U <new admin username> -D <new admin pwd>

Function: Set the user and password of admin account.

Arguments:

• Host: Any one IP of clusters nodes
• Port: Port which phxbinlogsvr is listening. like 17000
• Admin username: Current account user( default is root )
• Admin pwd: Current account password( default is "" )
• New admin username: New user
• New admin pwd: New password

phxbinlogsvr_tools -f SetMySqlReplicaInfo -h <host> -p <port> -u <admin username> -d <admin pwd> -U <new replica username> -D <new replica pwd>

Function: Set the user and password of synchronization account.

Arguments：

• Host: Any one IP of clusters nodes
• Port: Port which phxbinlogsvr is listening. like 17000
• Admin username: Current account user( default is root )
• Admin pwd: Current account password( default is "" )
• New replica username: New user
• New replica pwd: New password

phxbinlogsvr_tools_phxrpc -f GetMemberList -h <host> -p <port>

Function: Membership of this cluster, all IPs and Ports included.

Arguments:

• Host: Any one IP of clusters nodes
• Port: Port which phxbinlogsvr is listening. like 17000

Phxbinlogsvr Membership Managerment

Member Deletion

Execute phxbinlogsvr_tools_phxrpc -f RemoveMember -h<host> -p<port> -m <ip_of_nodeA> to delete node A. Once it is succesfully executed, A will not learn binlog after a small period.

Member Involvement

1. Execute phxbinlogsvr_tools -f AddMember -h<host> -p<port> -m <ip_of_nodeA> to add node A into the membership.
2. Install PhxSQL on A.
3. A will begin to learn data after installation is finished.
4. Copy a snapshot of MySQL from any other nodes to A.
5. Kill phxbinlogsvr and access MySQL through the local port( or socket ). then execute set global super_read_only = 0; set global read_only = 0;
6. Dump the snapshot into MySQL.
7. A will begin to work after a while.

Phxbinlogsvr fault Handling.

You can choose to reinstall PhxSQL if PhxSQL meets an unrecovery failure.

Phxbinlogsvr will pull the checkpoint in another node to reboot during reinstallation. It will self-kill after pulling is over(to make sure the consistency). You can reboot phxbinlogsvr after a message like "All sm load state ok, start to exit" appears.

phxbinlogsvr will stop working if a data problem arise in MySQL. We suggest you to check the status of MySQL.

You can observe logs with red "err" to check the abnormaly.

Performance Testing

Hosts Infomation

CPU : Intel(R) Xeon(R) CPU E5-2420 0 @ 1.90GHz * 24

Memory : 32G

Disk : SSD Raid10

Ping Costs

Master -> Slave : 3 ~ 4ms

Client -> Master : 4ms

Tools and Arguments

sysbench --oltp-tables-count=10 --oltp-table-size=1000000 --num-threads=500 --max-requests=100000 --report-interval=1 --max-time=200

Results

Client Threads	Clusters			Test sets
		insert.lua (100% write)		select.lua (0% write)		OLTP.lua (20% write)
		QPS	Response time(MS)	QPS	Response time(MS)	QPS	Response time(MS)
200	PhxSQL	5076	39.34/56.93	46334	4.21/5.12	25657	140.16/186.39
200	MySQL semi-sync	4055	49.27/66.64	47528	4.10/5.00	20391	176.39/226.76
500	PhxSQL	8260	60.41/83.14	105928	4.58/5.81	46543	192.93/242.85
500	MySQL semi-sync	7072	70.60/91.72	121535	4.17/5.08	33229	270.38/345.84

NOTE:The 2 Response times means average and 95% percentile