redis深入浅出

Agenda

Feature&Architecture Admin&capcity forecast Replication Presistence Table design

Feature

redis memcached

data structure string|list|hash|set|zset string

thread model third(main thread, two auxiliary thread)

many

demultiplexer ae.c libevent(3rd library)

persistence yes no

replication yes no

design weak no

String

Set hello world

Set hello 1234567 Set hello 1234

List

Lpush list aaaa bbb ccc

Double link list, so can lpush,rpush,lpop,rpop

Zipmap

Hset test hello world

Hset test aliyun dba

Zipmap defect:Zmlen one bytes, only 253 subkey

Architecture

Rehash

Cache forecast

string 类型的内存大小 = 键值个数 * (dictEntry大 小 + redisObject 大小 + 包含 key的 sds 大小 +

包含 value的 sds大小 ) + bucket 个数 * 4

zipmap 类型的内存大小 = hashkey 个数 * (dictEntry 大小 + redisObject 大小 + 包含 key的sds 大小 + subkey的总大小 ) + bucket 个数 * 4

Jemalloc size class

Admin

Redis.conf(static), config set(dynamic)

– maxmemory <bytes>

– Appendfsync everysec(fsync pagecache to disk)

– hash-max-zipmap-entries 512

– hash-max-zipmap-value 64

Replication

procedure map Master slave – slave – slave - slave

defects Without resume broken transfer Without lag(slave position)

replication

Persistence

Snapshot Fork process, loop hash table, save file dump.rdb Yes!!! sequential write Write dump.rdb need O_DIRECT

Aof Like binlog, for recover after crash Auto bgrewrteaof Auxiliary threads Postpone fsync

Auxiliary thread

Table desgin login user

Design a login user system Heap table

userid login_times last_login_time 1 5 2011-1-1 2 1 2011-1-2 3 2 2011-1-3

Last login man? Create index on last_login_time Max login man? Create index on login_times Index, explan plan, compute statistics is suck!!!

Login user

data

Set userid:1:login_times 5

Set userid:2:login_times 1

Set userid:3:login_times 2

Set userid:1:last_login 2011-1-1

Set userid:2:last_login 2011-1-2

Set userid:3:last_login 2011-1-3

Last login

lpush user_last_login 1

lpush user_last_login 2

lpush user_last_login 3

ltrim user_last_login 0 1

Login user

Max login man zadd user:login_times 5 1 zadd user:login_times 1 2 zadd user:login_times 2 3 zcard user:login_times zrangebyscore user:login_times 3 +inf withscores

Column store data?

Suitable Scene

轻量级的高性能消息队列服务 ,生产者消费者 跨机器的共享内存 Redis的主要缺点是数据库容量受到物理内存的限制，不能用作海量数据的高性能读写，并且它没有原生的可扩展机制，不具有scale（可扩展）能力，要依赖客户端来实现分布式读写，因此 Redis适合的场景主要局限在较小数据量的高性能操作和运算上。

QA

@hoterran

www.hoterran.info