4.5 延迟考虑

1. 尽可能使用批量操作：

mget、hmget而不是get和hget，对于set也是如此。
lpush向一个list一次性导入多个元素，而不用lset一个个添加
LRANGE 一次取出一个范围的元素，也不用LINDEX一个个取出

2. 尽可能的把redis和APP SERVER部署在一个网段甚至一台机器。

3. 对于数据量较大的集合，不要轻易进行删除操作，这样会阻塞服务器，一般采用重命名+批量删除的策略：

排序集合：

# Rename the key
newkey = "gc:hashes:" + redis.INCR("gc:index")
redis.RENAME("my.zset.key", newkey)

# Delete members from the sorted set in batche of 100s
while redis.ZCARD(newkey) > 0
  redis.ZREMRANGEBYRANK(newkey, 0, 99)
end

集合：

# Rename the key
newkey = "gc:hashes:" + redis.INCR("gc:index")
redis.RENAME("my.set.key", newkey)

# Delete members from the set in batches of 100
cursor = 0
loop
  cursor, members = redis.SSCAN(newkey, cursor, "COUNT", 100)
  if size of members > 0
redis.SREM(newkey, members)
  end
  if cursor == 0
break
  end
end

列表：

# Rename the key
newkey = "gc:hashes:" + redis.INCR("gc:index")
redis.RENAME("my.list.key", newkey)

# Trim off elements in batche of 100s
while redis.LLEN(newkey) > 0
  redis.LTRIM(newkey, 0, -99)
end

Hash：

# Rename the key
newkey = "gc:hashes:" + redis.INCR( "gc:index" )
redis.RENAME("my.hash.key", newkey)

# Delete fields from the hash in batche of 100s
cursor = 0
loop
  cursor, hash_keys = redis.HSCAN(newkey, cursor, "COUNT", 100)
  if hash_keys count > 0
redis.HDEL(newkey, hash_keys)
  end
  if cursor == 0
break
  end
end

4. 尽可能使用不要超过1M大小的kv。

5. 减少对大数据集的高时间复杂度的操作：根据复杂度计算，如下命令可以优化:

6. 尽可能使用pipeline操作：一次性的发送命令比一个个发要减少网络延迟和单个处理开销。一个性能测试结果为（注意并不是pipeline越大效率越高，注意最后一个测试结果） :

logger@BIGD1TMP:~> redis-benchmark -q -r 100000 -n 1000000 -c 50 
PING_INLINE: 90155.07 requests per second
PING_BULK: 92302.02 requests per second
SET: 85070.18 requests per second
GET: 86184.61 requests per second

logger@BIGD1TMP:~> redis-benchmark -q -r 100000 -n 1000000 -c 50 -P 10
PING_INLINE: 558035.69 requests per second
PING_BULK: 668002.69 requests per second
SET: 275027.50 requests per second
GET: 376647.84 requests per second

logger@BIGD1TMP:~> redis-benchmark -q -r 100000 -n 1000000 -c 50 -P 20
PING_INLINE: 705716.25 requests per second
PING_BULK: 869565.25 requests per second
SET: 343406.59 requests per second
GET: 459347.72 requests per second

logger@BIGD1TMP:~> redis-benchmark -q -r 100000 -n 1000000 -c 50 -P 50
PING_INLINE: 940733.81 requests per second
PING_BULK: 1317523.00 requests per second
SET: 380807.31 requests per second
GET: 523834.47 requests per second

logger@BIGD1TMP:~> redis-benchmark -q -r 100000 -n 1000000 -c 50 -P 100
PING_INLINE: 999000.94 requests per second
PING_BULK: 1440922.12 requests per second
SET: 386996.88 requests per second
GET: 602046.94 requests per second

logger@BIGD1TMP:~> redis-benchmark -q -r 100000 -n 1000000 -c 50 -P 200
PING_INLINE: 1078748.62 requests per second
PING_BULK: 1381215.50 requests per second
SET: 379218.81 requests per second
GET: 537634.38 requests per second

一个场景是一个购物车的设计，一般的设计思路是：在获取购物车内部货品时，不使用pipeline会很低效：可以修改为：