Redis, another step on the road

Redis, another step on the road
2015-05-15
Ant
yftzeng@gmail.com

2015
2/123
Agenda
✔ Redis history
✔ Redis 3.0
✔ Redis features
✔ Redis and Memcached
✔ Redis and Aerospike
✔ Insight on the pit
議程
✔ Redis 歷史
✔ Redis 3.0
✔ Redis 特性
✔ Redis 與 Memcached
✔ Redis 與 Aerospike
✔ 坑裡的洞見

2015
3/123
Agenda
✔ Redis history
✔ Redis 3.0
✔ Redis features
議程
✔ Redis 歷史
✔ Redis 3.0
✔ Redis 特性
✔ 坑裡的洞見

4/123
2015
Redis history
Redis 歷史
Redis 2.0
✔ Key-value store policy, key in memory, value in disk.
Redis 2.0
✔ Key-value 儲存策略， key 存記憶體， value 存硬碟。

5/123
2015
Redis history
Redis 歷史
Redis 2.4
✔ Key-value store policy, both key & vlaue are in memory.
✔ Add 2 background threads,
✔ fsync file descriptor.
✔ close file descriptor.
Redis 2.4
✔ Key-value 儲存策略， key 及 value 存記憶體。
✔ 除了 main thread 外，引入了 2 個 background threads,
✔ fsync file descriptor 。
✔ close file descriptor 。

6/123
2015
Redis history
Redis 歷史
Redis 2.6
✔ Server side Lua scripting.
✔ Virtual Memory removed.
✔ Milliseconds resolution expires.
✔ Removed hardcoded number of clients.
Redis 2.6
✔ 伺服器端支援 Lua 。
✔ 移除 Virtual Memory 。
✔ 「 Expires 」的毫秒精準度。
✔ 移除寫死的客戶端數量限制。

7/123
2015
Redis history
Redis 歷史
Redis 2.8
✔ Saving synchronization resource.
✔ Before 2.8, slave use SYNC with the master.
✔ After 2.8, slave use PSYNC with the master.
Redis 2.8
✔ 節省同步資源。
✔ 2.8 以前， slave 使用 SYNC 與 master 同步。
✔ 2.8 以後，改用 PSYNC ( 偏移量同步 ) 與 master 同步。

2015
8/123
Agenda
✔ Redis history
✔ Redis 3.0
✔ Redis features
議程
✔ Redis 歷史
✔ Redis 3.0
✔ Redis 特性
✔ 坑裡的洞見

9/123
2015
Redis 3.0
Redis 3.0
Release date: 1 Apr 2015.
✔ Redis Cluster.
✔ New "embedded string" object.
✔ Improved LRU approximation algorithm.
2015 年 4 月 1 日正式釋出。
✔ Redis Cluster 。
✔ “新的 embedded string” 。
✔ LRU 演算法的改進。

2015
10/123
Agenda
✔ Redis history
✔ Redis 3.0
✔ Redis features
議程
✔ Redis 歷史
✔ Redis 3.0
✔ Redis 特性
✔ 坑裡的洞見

11/123
2015
Redis features
Redis 特性
✔ RDBMS
✔ Oracle, DB2, PostgreSQL, MySQL, SQL Server, ...
✔ NoSQL
✔ Cassandra, HBase, Memcached, MongoDB, Redis, ...
✔ NewSQL
✔ Aerospike, FoundationDB, RethinkDB, ...
✔ RDBMS
✔ Oracle, DB2, PostgreSQL, MySQL, SQL Server, ...
✔ NoSQL
✔ Cassandra, HBase, Memcached, MongoDB, Redis, ...
✔ NewSQL
✔ Aerospike, FoundationDB, RethinkDB, ...

12/123
2015
Redis features
Redis 特性
✔ Key-value NoSQL
✔ Memcached, Redis, ...
✔ Column family NoSQL
✔ Cassandra, HBase, ...
✔ Documen NoSQL
✔ MongoDB, ...
✔ Graph NoSQL
✔ Neo4j, ...
✔ Key-value NoSQL
✔ Memcached, Redis, ...
✔ Column family NoSQL
✔ Cassandra, HBase, ...
✔ Documen NoSQL
✔ MongoDB, ...
✔ Graph NoSQL
✔ Neo4j, ...

13/123
2015
Redis features
Redis 特性
Pure
✔ ANSI C.
✔ Lesser 3rd
-party libraries.
✔ Memcached depends on libevent.
✔ Redis implement its own epoll event loop.
✔ KISS principle.
✔ Data structure do what it should do.
簡純
✔ ANSI C 撰寫。
✔ 幾乎不依賴第三方函式庫。
✔ memcached 使用 libevent ，程式碼龐大。
✔ Redis 參考 libevent 實現了自己的 epoll event loop 。
✔ KISS 原則。
✔ 每個數據結構只負責自己應當做的。

14/123
2015
Redis features
Redis 特性
Simple
✔ No map-reduce.
✔ No indexes.
✔ No vector clocks.
簡單
✔ No map-reduce.
✔ No indexes.
✔ No vector clocks.

15/123
2015
Redis features
Ref: http://oldblog.antirez.com/post/redis-manifesto.html
5 - We're against complexity. We believe designing
systems is a fight against complexity. We'll accept to
fight the complexity when it's worthwhile but we'll try
hard to recognize when a small feature is not worth
1000s of lines of code. Most of the time the best way
to fight complexity is by not creating it at all.

16/123
2015
Redis features
Redis 特性
Single thread
✔ No thread context switch.
✔ No thread race condition.
✔ No other complicated condition.
單執行緒
✔ No thread context switch.
✔ No thread race condition.
✔ No other complicated condition.

17/123
2015
Redis features
Redis 特性
In-memory but persistent on disk database
✔ Operation in memory.
✔ Persistent on disk.
記憶體資料庫，但可永久儲存於硬碟中
✔ 記憶體操作資料。
✔ 資料可永久儲存於硬碟。

18/123
2015
Redis features
Redis 特性
Remote dictionary server
✔ No only a cache server.
Remote dictionary server
✔ 不只是快取伺服器。
Ref: http://redis.io/topics/faq

2015
19/123
Agenda
✔ Redis history
✔ Redis 3.0
✔ Redis features
議程
✔ Redis 歷史
✔ Redis 3.0
✔ Redis 特性
✔ 坑裡的洞見

20/123
2015
Redis and Memcached
Redis 與 Memcached
✔ Redis is single thread IO multiplexing model.
✔ Simple operations to archive high throughput.
✔ Complicated (heavy) operations may block others.
✔ One instance usually only use one CPU.
✔ Redis 是單執行緒 IO 多路複用模式。
✔ 簡單的操作可以達到高吞吐。
✔ 複雜的操作容易阻塞其它的操作。
✔ 一個 Redis 實例通常只會用到一顆 CPU 。

21/123
2015
Redis and Memcached
Redis 與 Memcached
✔ Memcached is multi-threaded, non-blocking IO
✔ multiplexing network model.
✔ Multi-core architecture.
✔ But got cache coherency & lock issues.
✔ Memcached 是多執行緒非阻塞 IO 多路複用模式。
✔ 多執行緒可善用多顆 CPU 。
✔ 但會引入 cache coherency 及 lock 問題。

22/123
2015
Redis and Memcached
Redis 與 Memcached
✔ Redis can use jemalloc or tcmalloc to reduce
✔ memory fragmentation.
✔ But it depends on the allocation patterns.
✔ Rarely use the Free-list and other ways to optimize
✔ memory allocation.
✔ Redis is simple / pure / efficiency design.
✔ Redis 使用 jemalloc 或 tcmalloc 降低記憶體碎片。
✔ 但記憶體碎片的情形仍依賴於分配模式。
✔ 幾乎不用 Free-list 及其它方法來最佳化記憶體分配。
✔ 符合 Redis 簡單 / 單純 / 效率的設計原則。
Ref: http://www.databaseskill.com/1256096/
Ref: http://stackoverflow.com/questions/18097670/why-the-memory-fragmentation-is-less-than-1-in-redis

23/123
2015
Redis and Memcached
Redis 與 Memcached
✔ Memcached use pre-allocated / slot memory pool.
✔ slot and pool can reduce memory fragmentation.
✔ But bring some wasted space. (memory overhead)
✔ Memcached 使用預分配 slot 記憶體池。
✔ slot 及池能有效降低某種程度的記憶體碎片。
✔ 但會帶來一些空間浪費的問題。 (memory overhead)

24/123
2015
Redis and Memcached
Redis 與 Memcached
✔ Garbage Collection behavior: approximate LRU.
✔ Redis 2.6
✔ Random pick 3 samples, removed the oldest one,
✔ repeatedly until memory used less than
✔ 'maxmemory' limit.
✔ 垃圾回收行為：近似 LRU 演算法。
✔ Redis 2.6
✔ 預設隨機取 3 個樣本，移除最舊的該筆，如此反覆，
✔ 直到記憶體用量小於 maxmemory 的設定。
Ref: https://github.com/antirez/redis/blob/2.6/src/redis.c#L2464

25/123
2015
Redis and Memcached

26/123
2015
Redis and Memcached
Redis 與 Memcached
✔ Garbage Collection behavior: approximate LRU.
✔ Redis 3.0
✔ Default random pick 5 samples, insert/sort into
✔ a pool, remove the best one, repeatedly until
✔ memory used less than 'maxmemory' limit.
✔ 5 (now) is more than 3 (before) samples ;
✔ The best one is more approximate global.
✔ 垃圾回收行為： approximate LRU 。
✔ Redis 3.0
✔ 預設隨機取 5 個樣本，插入並排序至一個 pool ，移除
✔ 最佳者，如此反覆，直到記憶體用量小於 maxmemory
✔ 的設定。
✔ 樣本 5 比先前的 3 多；
✔ 從局部最優趨向全局最優。

27/123
2015
Redis and Memcached

28/123
2015
Redis and Memcached
Ref: http://redis.io/images/redisdoc/lru_comparison.png

29/123
2015
Redis and Memcached
Redis 與 Memcached
✔ Strong sides of Redis.
✔ Rich (data type) operations.
✔ Hashs, Lists, Sets, Sorted Sets, HyperLogLog etc.
✔ Bulit-in replication & cluster.
✔ in-place update operations.
✔ Support persistent on disk.
✔ Avoid thundering herd.
✔ Redis 的長處。
✔ 豐富的 ( 資料型態 ) 操作。
✔ Hashs, Lists, Sets, Sorted Sets, HyperLogLog 等。
✔ 內建 replication 及 cluster 。
✔ 就地更新 (in-place update) 操作。
✔ 支援持久化 ( 硬碟 ) 。
✔ 避免雪崩效應。

30/123
2015
Redis and Memcached
Redis 與 Memcached
✔ Strong sides of Memcached.
✔ Multi-threaded.
✔ Use almost all CPUs.
✔ Fewer blocking operations.
( And center locks don't scaled up to 5 threads )
✔ Lower memory overhead.
✔ Lower memory allocation pressure.
✔ Maybe less memory fragmentation.
✔ Memcached 的長處。
✔ 多執行緒。
✔ 善用多核 CPU 。 ( 而 center locks 不隨 CPU 擴展 )
✔ 更少的阻塞操作。
✔ 更少的記憶體開銷。
✔ 更少的記憶體分配壓力。
✔ 可能有更少的記憶體碎片。
Ref: https://github.com/memcached/memcached/blob/master/thread.c#L747

31/123
2015
Redis and Memcached
Ref: https://github.com/memcached/memcached/blob/master/thread.c#L747

32/123
2015
Redis and Memcached
Redis 與 Memcached
✔ My testbed (NO WARRANTY)
✔ Get: Memcached is usually faster than Redis.
✔ Set: Redis is usually faster than Memcached.
✔ Size from 0 ~ 100KB is better for Redis.
✔ Size from 100KB ~ 10MB is better for Memcached.
✔ Size from 10M ~ is better for Redis.
✔ 我的使用經驗 ( 免責聲明 )
✔ Get 時， Memcached 比 Redis 快。
✔ Set 時， Redis 比 Memcached 快。
✔ 數據 0~100KB 時，適合 Redis 。
✔ 數據 100KB~10MB 時，適合 Memcached 。
✔ 數據 10M 以上時，適合 Redis 。

33/123
2015
Redis and Memcached

2015
34/123
Agenda
✔ Redis history
✔ Redis 3.0
✔ Redis features
議程
✔ Redis 歷史
✔ Redis 3.0
✔ Redis 特性
✔ 坑裡的洞見

35/123
2015
Redis and Aerospike
Ref: http://www.aerospike.com/
✔ Speed
✔ Scalable
✔ Flash-optimized
✔ In-memory NoSQL
✔ ACID Compliant

36/123
2015
Redis and Aerospike
Redis 與 Aerospike
✔ Strong sides of Aerospike.
✔ Auto node discovery (cluster).
✔ ACID Compliant
✔ Flash-optimized (Memory & Disk persistence).
✔ Intelligent Client (Optimistic row locking etc.).
✔ Cross data center replication.
✔ Multi-core optimization.
✔ No hotspots.
✔ Aerospike 的長處。
✔ 自管理集群。
✔ ACID 兼容。
✔ 數據存儲最佳化 (Flash/SSD) 。
✔ 智能客戶端 (Optimistic row locking 等 ) 。
✔ 跨數據中心集群。
✔ 多核最佳化。
✔ 無熱點瓶頸。

37/123
2015
Redis and Aerospike
Ref: http://lynnlangit.com/2015/01/28/lessons-learned-benchmarking-nosql-on-the-aws-cloud-aerospikedb-and-redis/

38/123
2015
Redis and Aerospike
Ref: http://lynnlangit.com/2015/01/28/lessons-learned-benchmarking-nosql-on-the-aws-cloud-aerospikedb-and-redis/

39/123
2015
Redis and Aerospike
Redis 與 Aerospike
✔ Itamar Haber (Redis Labs, Chief Developers Advocate)
✔ Why didn't … use … pipelining and multi-key
✔ operations?
✔ Missing piece is a 20%-80% read/write test and a
✔ 100% write test.
✔ Totally unexplained by the fact that she used AOF.
✔ Comparisons are as hard to do right as they are
✔ easy to do wrong.
✔ Itamar Haber (Redis Labs 公司的首席開發者推廣師 )
✔ 為什麼不用 Redis 推薦做法，如使用 piplining 和多鍵操作。
✔ 沒有測試工作負載： 20%-80% 讀寫和 100% 寫的情境。
✔ 對於 AOF ，一般都是建議非主 Redis 實例執行。
✔ 最後，比較是一件很難做對卻很容易做錯的事。
Ref: https://redislabs.com/blog/the-lessons-missing-from-benchmarking-nosql-on-the-aws-cloud-aerospikedb-and-redis

40/123
2015
Redis and Aerospike
Redis 與 Aerospike
✔ Salvatore Sanfilippo (antirez, the author of Redis)
✔ GET/SET Benchmarks are not a great way to
✔ compare different database systems.
✔ A better performance comparison is by use case.
✔ Test with instance types most people are going to
✔ actually use, huge instance types can mask
✔ inefficiencies of certain database systems, and is
✔ anyway not what most people are going to use.
✔ Salvatore Sanfilippo (antirez, Redis 作者 )
✔ GET/SET 不能比較出資料庫間的效能差異。
✔ 效能是需要依據業務場景而定。
✔ 測試應當依據大多數用戶的實際案例，太多的案例會掩蓋
✔ 某些資料庫的低效率，而且這樣的案例也不是大多數用戶
✔ 會遇到的。
Ref: http://antirez.com/news/85

41/123
2015
Redis and Aerospike
Ref: https://aphyr.com/posts/324-call-me-maybe-aerospike

42/123
2015
Redis and Aerospike
Redis 與 Aerospike
✔ However, as the network shifts, … .By the time of the
✔ final read, about 10% of the increment operations
✔ have been lost.
✔ Just like the CaS register test, increment and read
✔ latencies will jump from ~1 millisecond to ~500
✔ milliseconds when a partition occurs.
✔ Aerospike can service every request successfully,
✔ peaking at ~2 seconds.
✔ 當 Network partition 發生時， Aerospike 會在某個很短的
✔ 時間內丟失操作。以每秒 500 次的 increment operations
✔ 測試，丟失約 10% 的寫入。
✔ 在 partition 完成後，會有幾秒很明顯的 latency 高峰出現。
✔ Aerospike 即使在已經執行已久的 partition 中，也會出現
✔ 服務中斷的情形，中斷甚至長達 2 秒。
Ref: http://antirez.com/news/85

43/123
2015
Redis and Aerospike
Redis 與 Aerospike
✔ In the summer of 2013 we faced exactly this problem:
✔ big-memory (192 GB RAM) server nodes were running
✔ out of memory and crashing again … We were being
✔ bitten by fragmentation.
✔ 2013 年夏天， Aerospike 突然有一台 192 GB RAM 的伺服器
✔ 因記憶體用盡而當機， ASMalloc 工具未查出 memory leak ，
✔ 所以看來是因為記憶體碎片造成的。
Ref: http://highscalability.com/blog/2015/3/17/in-memory-computing-at-aerospike-scale-when-to-choose-and-ho.html

2015
44/123
Agenda
✔ Redis history
✔ Redis 3.0
✔ Redis features
議程
✔ Redis 歷史
✔ Redis 3.0
✔ Redis 特性
✔ 坑裡的洞見

45/123
2015
Insight on the pit
坑裡的洞見
Server-side sessions with Redis
使用 Redis 共享 Sessions

46/123
2015
小心
先驅變先烈
朕知道
了

47/123
2015
Insight on the pit 【 Server-side sessions with Redis 】
Ref: http://vc2tea.com/redis-session/

48/123
2015
坑裡的洞見【使用 Redis 共享 Sessions 】
✔ Redis has many eviction policies, but most of them
✔ are based on 'sampling'.
✔ This means eviction item is not global optimization,
✔ but local optimization.
✔ When reach 'maxmemory', it may evict items not
✔ old enough.
✔ Users get logged out early, and the worst is you
✔ won’t even notice it, until users start complaining.
✔ Redis 有很多種移除舊數據的策略，但大多基於「抽樣」。
✔ 這意謂移除舊數據不是全局最優而是局部最優。
✔ 當達到 'maxmemory' 上限時，可能造成移除的數據「不
✔ 夠舊」。
✔ 使得使用者提前被登出。最糟的是，你可能都不會知道，
✔ 直到使用者開始抱怨。
Ref: http://redis.io/topics/lru-cache

49/123
2015
✔ Alternative solutions.
✔ Use database as an another back-end.
✔ 1. When write session, set both in Redis and
✔ database.
✔ 2. When read session, Redis first, database
✔ second.
✔ Redis 3.0.
✔ More 'sampling'.
✔ 替代方案。
✔ 使用資料庫為另一儲存後台。
✔1. 寫入 Session 時，同時寫進 Redis 及資料庫。
✔2. 讀出 Session 時， Redis 優先，資料庫其次。
✔ 使用 Redis 3.0 。
✔ 選擇較大的 'sampling' ( 抽樣數 ) 。

50/123
2015
✔ A better way is … (thinking)
✔ 還有其它的解法 !!! ( 思考 )

51/123
2015
Insight on the pit
坑裡的洞見
Maximize CPUs usage
善用多核 CPU

52/123
2015
偶爾任性是可愛，
一天到晚任性是妖孽。
朕知道
了

53/123
2015
Insight on the pit 【 Maximize CPUs usage 】
坑裡的洞見【善用多核 CPU 】
✔ Redis is single thread.
✔ One instance usually only use one CPU.
✔ (background threads.)
✔ (background tasks, such as BGSAVE, AOF rewrite.)
✔ Redis 是單執行緒。
✔ 一個 Redis 實例通常只會用到一顆 CPU 。
✔ ( 背景執行緒 )
✔ ( 背景工作，例如 BGSAVE 及 AOF rewrite)

54/123
2015

55/123
2015
坑裡的洞見【善用多核 CPU 】
✔ Maximize CPUs usage.
✔ Redis instances is same as CPU cores.
✔ But,
✔ 1.Set 'maxmemory' for each instance carefully.
✔ 2.Instance should have different 'dbfilename'.
✔ 3.Instance should have different 'appendfilename'.
✔ 善用多核 CPU 。
✔ 啟動的 Redis 實例與 CPU 核心數一樣多。
✔ 但，
✔ 1. 每個實例的 'maxmemory' 需要小心配置。
✔ 2. 每個實例的 'dbfilename' 需要不一樣。
✔ 3. 每個實例的 'appendfilename' 需要不一樣。

56/123
2015
Insight on the pit
坑裡的洞見
Memory optimization
記憶體優化

57/123
2015
前程四緊：
　手頭緊、
　眉頭緊、
　衣服緊、
　時間緊。
朕知道
了

58/123
2015
Insight on the pit 【 Memory optimization 】
坑裡的洞見【記憶體優化】
✔ Memory fragmentation.
✔ SET.
✔ rehash.
✔ When hash table needs to switch to a bigger
✔ or smaller table this happens incrementally.
✔ 記憶體碎片。
✔ SET 。
✔ rehash 。
✔ 當 dict 鍵值持續增加時，為保持良好的效能， dict
✔ 需要執行 rehash 。

59/123
2015
Ref: http://redisbook.readthedocs.org/en/latest/internal-datastruct/dict.html

60/123
2015
✔ Key name length.
✔ Shorter is better.
✔ But also meaningful ones.
✔ “product:user1:count” is better than “pu1c”.
✔ Key 命名長度。
✔ 長度愈短愈好。
✔ 但還是要有意義。
✔ “product:user1:count” ”優於 pu1c” 。

61/123
2015
✔ Ziplist.
✔ The ziplist is a specially encoded dually linked
✔ list that is designed to be very memory efficient.
✔ Ziplist is space efficient.
✔ Ziplist 。
✔ 符合某種設定下，資料結構會以 Ziplist 方式儲存。
✔ 類似一維線性儲存，省去大量的指針開銷。
Ref: http://redis.io/topics/memory-optimization

62/123
2015
✔ Ziplist.
✔ hash-max-ziplist-entries 64
✔ hash-max-ziplist-value 512
✔ list-max-ziplist-entries 512
✔ list-max-ziplist-value 64
✔ zset-max-ziplist-entries 128
✔ zset-max-ziplist-value 64
✔ set-max-intset-entries 512
✔ Ziplist 。
✔ list-max-ziplist-entries 512
✔ list-max-ziplist-value 64
✔ zset-max-ziplist-entries 128
✔ zset-max-ziplist-value 64
✔ set-max-intset-entries 512

63/123
2015
✔ Ziplist.
✔ Use ziplist if entries count ≦ 64 or
✔ every entry size ≦ 512.
✔ Ziplist 。
✔ 如果 Hash ≦的數量 64 ，或其中一個 Hash ≦的值
✔ 512 ，則使用 Ziplist 。

64/123
2015
✔ Ziplist.
✔ Twitter use case.
✔ A Redis ziplist threshold is set to the max size
✔ of a Timeline. Never store a bigger Timeline
✔ than can be stored in a ziplist.
✔ Ziplist 。
✔ Twitter 的案例。
✔ Ziplist 的數量設定與 Timelines 的最大數量一致；
✔ Timeline 的儲存大小也不會超過 Ziplist 的上限。
Ref: http://highscalability.com/blog/2014/9/8/how-twitter-uses-redis-to-scale-105tb-ram-39mm-qps-10000-ins.html

65/123
2015
✔ REDIS_SHARED_INTEGERS.
✔ Default is 10,000.
✔ Integers can be stored in a shared memory pool,
✔ and don't have any memory overheads.
✔ REDIS_SHARED_INTEGERS 。
✔ 預設是 10,000 。
✔ 整數 ( 包括 0) 可以預分配在共享池，避免重複分配而節省
✔ 記憶體。

66/123
2015
Ref: http://redisbook.readthedocs.org/en/latest/datatype/object.html
Flyweight
src/redis.h

67/123
2015
✔ Bitmaps.
✔ HyperLogLogs.
✔ Bitmaps.
✔ HyperLogLogs.

68/123
2015
Insight on the pit
坑裡的洞見
Availability
可用性

69/123
2015
開發都想好自在，
客戶都要靠得住。
朕知道
了

70/123
2015
Insight on the pit 【 Availability 】
坑裡的洞見【可用性】
✔ Twemproxy (Twitter)
✔ Codis ( 豌豆荚 )
✔ Redis Cluster (Official)
✔ Cerberus (HunanTV)
✔ Redis Cluster ( 官方 )
✔ Cerberus ( 芒果 TV)

71/123
2015
✔ Twemproxy is proxy-based solution.
✔ Good parts
✔ Stable, enterprise ready.
✔ 代理分片機制。
✔ 優點
✔ 非常穩定，企業級方案。

72/123
2015
✔ Bad parts
✔ SPOF (Single Point Of Failure)
✔ Keepalived etc.
✔ Smoothless on scale.
✔ No dashboard.
✔ Proxy-based, more route trip times, higher latency.
✔ Single-threaded proxy model.
✔ 缺點
✔ 單點故障。
✔ 需依賴第三方軟體，如 Keepalived 。
✔ 無法平滑地橫向擴展。
✔ 沒有後台介面。
✔ 代理分片機制引入更多的來回次數並提高延遲。
✔ 單核模式，無法充份利用多核，除非多實例。

73/123
2015
✔ Bad parts
✔ Twemproxy is not used by Twitter internally.
✔ 缺點
✔ Twitter 官方內部不再繼續使用 Twemproxy 。

74/123
2015
✔ Codis is proxy-based solution.
✔ 豌豆莢 open source on Jan 2014.
✔ Written in Go and C.
✔ 代理分片機制。
✔ 豌豆莢於 2014 年 11 月開放源碼。
✔ 基於 Go 與 C 開發。

75/123
2015
✔ Good parts
✔ Stable, enterprise ready.
✔ Auto Rebalance.
✔ High performance.
✔ Simple testbed is faster 100% than Twemproxy.
✔ Multi-threaded proxy model.
✔ 優點
✔ 非常穩定，企業級方案。
✔ 數據自動平衡。
✔ 高效能。
✔ 簡單的測試顯示較 Twemproxy 快一倍。
✔ 善用多核 CPU 。

76/123
2015
✔ Good parts
✔ Simple
✔ No paxos-like coordinators,
✔ No master-slave replication.
✔ Dashboard.
✔ 優點
✔ 簡單。
✔ 沒有 Paxos 類的協調機制。
✔ 沒有主從複製。
✔ 有後台介面。

77/123
2015
✔ Bad parts
✔ Need 3rd
-party coordinators
✔ Zookeeper or Etcd.
✔ No master-slave replication.
✔ 缺點
✔ 需要第三方軟體支持協調機制。
✔ 目前支援 Zookeeper 及 Etcd 。
✔ 不支援主從複製，需要另外實作。

78/123
2015
Ref: http://0xffff.me/blog/2014/11/11/codis-de-she-ji-yu-shi-xian-part-3/
Codis 的设计与实现 Part 3
✔ Codis 采用了 Proxy 的方案，所以必然会带来单机性能
✔ 的损失。
✔ 经测试，在不开 pipeline 的情况下，大概会损失 40%
✔ 左右的性能，但是 Redis 本身是一个快得吓人的东西，
✔ 即使单机损失了 40% 仍然是一个很大的数字。

79/123
2015
✔ Official supports.
✔ Requires Redis 3.0 or higher.
✔ 官方支援。
✔ 需要 Redis 3.0 或更高版本。

80/123
2015
✔ Good parts
✔ Official supports.
✔ Pear-to-pear Gossip distributed model.
✔ Less route trip times, lower latency.
✔ Automatically sharded across multiple Redis nodes.
✔ Do not need 3rd
-party coordinators
✔ 優點
✔ 官方支援。
✔ 無中心的 P2P Gossip 分散式模式。
✔ 更少的來回次數並降低延遲。
✔ 自動於多個 Redis 節點進行分片。
✔ 不需要第三方軟體支持協調機制。

81/123
2015
✔ Bad parts
✔ Need time to prove its stability.
✔ No dashboard.
✔ Need smart client.
✔ Redis client need to support for Redis Cluster.
✔ More maintenance cost than Codis.
✔ 缺點。
✔ 需要時間驗證其穩定性。
✔ 需要智能客戶端。
✔ Redis 客戶端必須支援 Redis Cluster 設計。
✔ 較 Codis 有更多的維護升級成本。

82/123
2015
✔ Good parts
✔ Auto Rebalance.
✔ Implement Redis's Smart Client.
✔ Read-write split.
✔ 優點
✔ 數據自動平衡。
✔ 本身實現了 Redis 的 Smart Client 。
✔ 支援讀寫分離。
Ref: https://github.com/HunanTV/redis-cerberus

83/123
2015
✔ Bad parts
✔ Need time to prove its stability.
✔ No dashboard.
✔ 缺點
✔ 需要時間驗證其穩定性。

84/123
2015
Insight on the pit
坑裡的洞見
Stabilization
穩定性

85/123
2015
每一個穩定服務背後，
都有一個齷齪的實現。
朕知道
了

86/123
2015
Insight on the pit 【 Stabilization 】
坑裡的洞見【穩定性】
✔ Performance fluctuation.
✔ Out of memory.
✔ Big Ziplist.
✔ Master-slave.
✔ 效能抖動。
✔ 記憶體不足。
✔ Big Ziplist 。
✔ 主從模式。

87/123
2015
✔ For production, stabilization is more important than
✔ average performance.
✔ Easy to estimated, reduce the chances of an
✔ important moment occurred at lower point.
✔ Redis is single thread.
✔ 效能抖動。
✔ 對於一個上線服務而言，穩定性遠大於平均效能。
✔ 效能防抖動，好預估，降低重要時刻發生在低點的機率。
✔ Redis 是單執行緒。

88/123
2015
✔ Tips: Split heavy commands.
✔ MGET
✔ redis> MGET 1 2 3 … 999
✔ ZRANGE
✔ redis> ZRANGE myset 0 -1
✔ SORT / LREM / SUNION / SDIFF / SINTER
✔ KEYS / SMEMBERS / HGETALL
✔ 效能抖動。
✔ 拆解「重」指令。
✔ MGET 。
✔ redis> MGET 1 2 3 … 999
✔ ZRANGE 。
✔ redis> ZRANGE myset 0 -1
✔ SORT / LREM / SUNION / SDIFF / SINTER 。
✔ KEYS / SMEMBERS / HGETALL 。

89/123
2015
✔ Tips: Rethink block commands.
✔ BLPOP
✔ BRPOPLPUSH
✔ BRPOP
✔ MULTI / EXEC
✔ 效能抖動。
✔ 「阻塞」指令。
✔ BLPOP 。
✔ BRPOPLPUSH 。
✔ BRPOP 。
✔ MULTI / EXEC 。

90/123
2015
✔ Out of memory (OOM).
✔ Be careful those commands will ask huge memory.
✔ Reduce the chances of Redis to be killed by OOM.
✔ SWAP, lose a little performance is better than crash.
✔ 記憶體不足 (Out of memory, OOM) 。
✔ 留意那些會大量耗用記憶體的指令。
✔ 降低 Redis 強制被 Out of memory 關閉的機率。
✔ 開啟 SWAP ，效能下降總比服務停用來得好。

91/123
2015
✔ maxmemory
✔ overcommit_memory
✔ SWAP
✔ zone_reclaim_mode
✔ oom_adj
✔ maxmemory
✔ SWAP
✔ oom_adj

92/123
2015
✔ maxmemory
✔ A rule of thumbs is 50% of total memory.
✔ BGSAVE.
✔ AOF rewrite.
✔ maxmemory
✔ 經驗法則是設定為總記憶體的 50% 。
✔ BGSAVE 。
✔ AOF rewrite 。
Ref: http://redis.io/topics/admin

93/123
2015
✔ overcommit_memory = 1
✔ Do overcommit.
✔ maxmemory
✔overcommit_memory = 1
✔ 請求分配記憶體時，永遠假裝還有足夠的記憶體。

94/123
2015
✔ SWAP
✔ Use SWAP, and same size of memory.
✔ SWAP
✔ 使用 SWAP ，並且與記憶體一樣大。

95/123
2015
✔ zone_reclaim_mode = 0 (default)
✔ zone_reclaim_mode = 0 ( 預設 )

96/123
2015
✔ If RHEL/CentOS ≧ 6.4 or Kernel ≧ 3.5-rc1.
✔ (1) Prefer swap to OOM.
✔ vm.swappiness = 1
✔ (2) Prefer OOM to swap.
✔ Else
✔ 如果 RHEL/CentOS ≧ 6.4 或 Kernel ≧ 3.5-rc1 。
✔ (1) 寧願 swap 也不要 OOM 。
✔ (2) 寧願 OOM 也不要 swap 。
✔ 否則

97/123
2015
✔ If (1) then oom_adj.
✔ echo -15 > /proc/`pidof redis-server`/oom_adj
✔ Reduce the chances of redis to be killed.
✔ Tips
✔ for i in $(pidof redis-server);
✔ do echo -15 | sudo tee /proc/$i/oom_adj ; done
✔ 如果 (1) 則 oom_adj 。
✔ echo -15 > /proc/`pidof redis-server`/oom_adj
✔ 降低 Redis 強制被 Out of memory 關閉的機率。
✔ Tips
✔ for i in $(pidof redis-server);
✔ do echo -15 | sudo tee /proc/$i/oom_adj ; done

98/123
2015
✔ 過去 vm.swappiness 設定為 0 可以降低 swap 的發生率，
✔ 但非完全禁止。所以預期會發生 swap 而不會 OOM 。
✔ 在 Linux Kernel 3.5-RC1 及 RHEL/CentOS Kernel
✔ 2.6.32-303 (CentOS 6.4) 之後的版本，已經改變此行為。
✔ 設定為 0 時完全不會有任何 swap ，但非預期的記憶體壓力
✔ 可能會造成 OOM 而關閉 Redis 。

99/123
2015
Linux Kernel 3.4 (mm/vmscan.c)
Ref: https://www.kernel.org/pub/linux/kernel/v3.x/linux-3.4.tar.xz
Ref: https://www.kernel.org/pub/linux/kernel/v3.x/linux-3.5.1.tar.gz
Linux Kernel 3.5.1 (mm/vmscan.c)

100/123
2015
linux-2.6.32-504.12.2.el6 (CentOS 6.4, mm/vmscan.c)
Ref: http://rpm.pbone.net/index.php3/stat/3/srodzaj/2/search/kernel-2.6.32-504.12.2.el6.src.rpm

101/123
2015
✔ Redis have some background tasks.
✔ fsync file descriptor.
✔ close file descriptor.
✔ BGSAVE.
✔ AOF rewrite.
✔ Preserved CPU to do those tasks.
✔ Redis 會執行一些 background tasks 。
✔ fsync file descriptor 。
✔ close file descriptor 。
✔ BGSAVE 。
✔ AOF rewrite 。
✔ 預留一些 CPU 執行這些 tasks 。

102/123
2015
✔ Instance have its own synchronization.
✔ Disable automatic on BGSAVE / BGREWRITEAOF,
✔ and use manual control instead.
✔ Avoid execution at the same time.
✔ 每個實例都有自己的同步機制。
✔ 關閉自動 BGSAVE / BGREWRITEAOF ，改為手動。
✔ 避免各實例同時啟動，耗用大量資源。

103/123
2015
✔ Master-slave.
✔ Best practices.
✔ N Redis nodes.
✔ 1 master, 1 slave, N-2 slaves of slave.
✔ Never restart all or multiple slave instances.
✔ (Master) High CPU loading.
✔ (Master) May out of memory.
✔ 主從模式。
✔ 最佳實踐。
✔ N 台 Redis 。
✔ 1 台主服務， 1 台從服務， N-2 台從服務的從服務。
✔ 不要同時重啟所有或大量的 slave 實例。
✔ 造成主服務 CPU 負載過高。
✔ 造成主服務記憶體用量過高。

104/123
2015
✔ String value.
✔ String value can be at max 512 MB in length.
✔ A rule of thumbs is no more than 5KB.
✔ 字串值。
✔ 字串值最大可以儲存 512MB 的長度。
✔ 經驗上最好不要大於 5KB 。

105/123
2015
Insight on the pit
坑裡的洞見
Low latency
低延遲

106/123
2015
很多事都介於
「不說憋屈」
和
「說了矯情」
之間
朕知道
了

107/123
2015
Insight on the pit 【 Low latency 】
坑裡的洞見【低延遲】
✔ Durability vs latency tradeoffs, from higher to lower
latency.
✔ AOF + fsync always.
✔ AOF + fsync every second.
✔ AOF + fsync every second +
✔ No-appendfsync-on-rewrite set to yes.
✔ AOF + fsync nerver.
✔ RDB.
✔ 數據持久性 vs 延遲性的權衡，延遲性從高至低排列。
✔ AOF + fsync always 。
✔ AOF + fsync every second 。
✔ AOF + fsync every second +
✔ No-appendfsync-on-rewrite set to yes 。
✔ AOF + fsync nerver 。
✔ RDB 。
Ref: http://redis.io/topics/latency

108/123
2015
✔ Latency induced by network and communication.
✔ Reduce the numbers of commands.
✔ Pipelining.
✔ MSET / MGET.
✔ 網路造成的延遲性。
✔ 減少指令的使用次數。
✔ Pipelining 。
✔ MSET / MGET 。

109/123
2015
✔ Fork time in different systems.
✔ 不同系統間的 Fork 時間。
Linux on physical machine (Xeon@2.27Ghz)Linux on physical machine (Xeon@2.27Ghz)
Linux VM on EC2 (Xen)Linux VM on EC2 (Xen)
Linux beefy VM on VMwareLinux beefy VM on VMware
Linux on physical machine (Unknown HW)Linux on physical machine (Unknown HW)
Linux VM on 6sync (KVM)Linux VM on 6sync (KVM)
Linux VM on Linode (Xen)Linux VM on Linode (Xen)
9 ms/GB9 ms/GB
10 ms/GB10 ms/GB
12.8 ms/GB12.8 ms/GB
424 ms/GB424 ms/GB
Linux on physical machine (Xeon@2.27Ghz)Linux on physical machine (Xeon@2.27Ghz)
Linux VM on EC2 (Xen)Linux VM on EC2 (Xen)
Linux beefy VM on VMwareLinux beefy VM on VMware
Linux on physical machine (Unknown HW)Linux on physical machine (Unknown HW)
Linux VM on 6sync (KVM)Linux VM on 6sync (KVM)
Linux VM on Linode (Xen)Linux VM on Linode (Xen)
9 ms/GB9 ms/GB
10 ms/GB10 ms/GB
424 ms/GB424 ms/GB

110/123
2015
✔ Never use Huge page.
✔ echo never >
/sys/kernel/mm/transparent_hugepage/enabled
✔ 永不用 Huge page 。
✔ echo never >
/sys/kernel/mm/transparent_hugepage/enabled

111/123
2015
✔ Do you really need Proxy-based solution (Codis) ?
✔ 真的需要代理分片機制 ( 如 Codis) 嗎？

112/123
2015
简单的测试，单 redis+ 单 proxy ，默认参数
Ref: https://github.com/wandoulabs/codis/issues/63
~50%
~57%
~428%
~561%
~937%
~50%

113/123
2015
✔ Codis.
✔ Disable pipeline.
✔ Less CPU cores.
✔ Codis 。
✔ 停用 pipeline 。
✔ 少核 CPU 。

114/123
2015
Disable pipeline.
Ref: https://github.com/wandoulabs/codis/blob/master/doc/benchmark_zh.md

115/123
2015
Less CPU cores.
Ref: https://github.com/wandoulabs/codis/blob/master/doc/benchmark_zh.md
4 cores
12 cores 16 cores
8 cores

116/123
2015
✔ Big Ziplist.
✔ Adding to and deleting from a ziplist is inefficient,
✔ especially with a very large list.
✔ Deleting from a ziplist uses memmove to move
✔ data around, to make sure the list is still contiguous.
✔ Adding to a ziplist requires a memory realloc call to
✔ make enough space for the new entry.
✔ Big Ziplist 。
✔ 從 Ziplist 中新增或刪除都沒有效率，尤其是 Big Ziplist 。
✔ 從 Ziplist 刪除會利用 memmove 移動資料，以確保 list
✔ 還是連續的。
✔ 在 Ziplist 中新增需要 memory realloc 以產出足夠的空間
✔ 供新值儲存。

117/123
2015
✔ Big Ziplist.
✔ Potential high latency for write operations due to
✔ timeline size.
✔ Big Ziplist 。
✔ Ziplist 中的寫操作很可能會因 Big Ziplist 而帶來高延遲。

118/123
2015
✔ Redis client.
✔ Connection pool.
✔ Keep alive.
✔ Redis 客戶端。
✔ Connection pool 。
✔ Keep alive 。

119/123
2015
✔ Redis 3.0 embedded string.
✔ New "embedded string" object.
✔ Reduce memory operations.
✔ If string length ≦ 39 bytes.
✔ Redis 3.0 embedded string 。
✔ 新的 "embedded string" 物件。
✔ 減少記憶體操作次數。
✔ ≦如果字串長度 39 。

120/123
2015
Redis 2.8.20
robj (16 bytes)
sdshdr (8 bytes)
string (N bytes)
ptr

121/123
2015
Redis 3.0
robj (16 bytes)
sdshdr (8 bytes)
string (40 bytes)
ptr
In 64 bit system:
jemalloc arena may “64 byte-long”.
64 - 16 (robj) - 8 (sdshdr) = 40
40 - 1 (null term, 0) = 39

2015
122/123
Feedback 勘誤回報
yftzeng@gmail.com

2015
123/123
End 結語
Parvenu use Redis/Memcached (Much memory) ；
TRS use Aerospike (Memory / SSD mixed) ；
Mortals use SSDB (Disk only) 。
暴發戶用 Redis/Memcached ( 記憶體多 ) ；
高富帥用 Aerospike ( 記憶體與 SSD 混用 ) ；
平民級用 SSDB ( 記憶體缺 ) 。

Redis, another step on the road

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (20)

Similar to Redis, another step on the road

Similar to Redis, another step on the road (20)

More from Yi-Feng Tzeng

More from Yi-Feng Tzeng (16)

Recently uploaded

Recently uploaded (20)

Redis, another step on the road