一、現(xiàn)象

實例名：r-bp1cxxxxxxxxxd04(主從)

問題：一分鐘內存上漲了2G，如下圖所示：

鍵值規(guī)模：6000萬左右

內存一分鐘增長2G.png

二、Redis內存分析

1. 內存組成

[[237264]]

上圖中的內存統(tǒng)計的是Redis的info memory命令中的used_memory屬性，例如：

redis>infomemory#Memoryused_memory:9195978072used_memory_human:8.56Gused_memory_rss:9358786560used_memory_peak:10190212744used_memory_peak_human:9.49Gused_memory_lua:38912mem_fragmentation_ratio:1.02mem_allocator:jemalloc-3.6.0 

每個屬性的詳細說明

計算公式如下：

used_memory = 自身內存+對象內存+緩沖內存+lua內存used_rss = used_memory + 內存碎片 

如下圖所示：

2. 內存分析

(1) 自身內存：一個空的Redis占用很小，可以忽略不計

(2) kv內存：key對象 + value對象

(3) 緩沖區(qū)：客戶端緩沖區(qū)(普通 + slave偽裝 + pubsub)以及aof緩沖區(qū)(比較固定，一般沒問題)

(4) Lua：Lua引擎所消耗的內存

3. 內存突增常見問題

(1) kv內存：bigkey、大量寫入

(2) 客戶端緩沖區(qū)：一般常見的有普通客戶端緩沖區(qū)(例如monitor命令)或者pubsub客戶端緩沖區(qū)

三、問題排查

(1) bigkey ? 經掃描未發(fā)現(xiàn)bigkey

(2) 鍵值個數(shù)增加？未發(fā)現(xiàn)鍵值有明顯變化

(3) 客戶端緩沖區(qū)

由于內存增上去后，長時間沒下落，如果是因為緩沖區(qū)問題，會從info clients找到明顯問題，執(zhí)行后發(fā)現(xiàn)：

執(zhí)行client中也沒有明顯的omem大于0的情況

id=80207addr=10.xx.0.4:63920fd=46name=age=624idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80215addr=10.xx.0.23:43489fd=36name=age=591idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80366addr=10.xx.0.8:59785fd=18name=age=84idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=delread=0write=0type=user 
id=80356addr=10.xx.0.33:32117fd=13name=age=114idle=0flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80064addr=10.xx.59.4:53446fd=38name=age=1070idle=1070flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=NULL read=0write=0type=admin 
id=80276addr=10.xx.0.23:48511fd=8name=age=387idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80188addr=10.xx.0.33:16265fd=42name=age=681idle=3flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80326addr=10.xx.0.32:59779fd=16name=age=209idle=0flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80065addr=10.xx.59.4:53447fd=45name=age=1070idle=1070flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=NULL read=0write=0type=admin 
id=79936addr=10.xx.0.22:10607fd=30name=age=1480idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80174addr=10.xx.0.5:60914fd=6name=age=722idle=2flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80300addr=10.xx.0.22:22757fd=48name=age=298idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80037addr=10.xx.0.5:55189fd=15name=age=1143idle=2flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80330addr=10.xx.0.8:48533fd=17name=age=199idle=10flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=79896addr=10.xx.0.30:26814fd=11name=age=1616idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80299addr=10.xx.0.24:11227fd=44name=age=303idle=3flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80086addr=10.xx.0.32:52526fd=40name=age=1002idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80202addr=10.xx.0.33:16658fd=26name=age=636idle=3flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80256addr=10.xx.0.24:60496fd=19name=age=448idle=2flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=79908addr=10.xx.0.29:18975fd=12name=age=1583idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80365addr=10.xx.0.29:46429fd=14name=age=85idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=79869addr=10.xx.27.4:48455fd=35name=age=1700idle=1700flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=NULL read=0write=0type=admin 
id=80334addr=10.xx.0.23:50012fd=39name=age=189idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80041addr=10.xx.0.32:51107fd=33name=age=1132idle=3flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=79992addr=10.xx.0.22:12068fd=28name=age=1289idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80251addr=10.xx.0.30:44213fd=23name=age=468idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80006addr=10.xx.0.2:45895fd=31name=age=1242idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80321addr=10.xx.0.30:48048fd=5name=age=224idle=3flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80381addr=10.xx.0.8:13360fd=22name=age=24idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=delread=0write=0type=user 
id=80200addr=10.xx.0.24:59183fd=24name=age=640idle=0flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80113addr=10.xx.0.2:52492fd=21name=age=915idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=174addr=11.216.117.242:53027fd=9name=age=281390idle=0flags=S db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=replconf read=0write=0type=admin 
id=79991addr=10.xx.0.4:48412fd=25name=age=1296idle=0flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80301addr=127.0.0.1:47869fd=49name=age=291idle=261flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=strlen read=0write=0type=admin 
id=80047addr=10.xx.59.4:53184fd=41name=age=1114idle=1114flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=NULL read=0write=0type=admin 
id=80236addr=10.xx.0.5:62546fd=47name=age=516idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80364addr=10.xx.0.4:18794fd=7name=age=85idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80175addr=10.xx.0.4:62245fd=29name=age=718idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80336addr=10.xx.0.29:45701fd=50name=age=180idle=1flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80050addr=10.xx.59.4:53188fd=43name=age=1114idle=1114flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=NULL read=0write=0type=admin 
id=79765addr=10.xx.0.2:33832fd=37name=age=2027idle=177flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=info read=0write=0type=user 
id=80170addr=10.xx.0.2:57853fd=20name=age=728idle=24flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=0obl=0oll=0omem=0events=r cmd=ping read=0write=0type=user 
id=80390addr=127.0.0.1:49449fd=27name=age=0idle=0flags=N db=0sub=0psub=0multi=-1qbuf=0qbuf-free=32768obl=0oll=0omem=0events=r cmd=client read=0write=0type=admin 

四、揪出元兇

常用的幾招都用了，還是不行，同事@徑遠幫忙一起分析，懷疑是不是因為Redis的kv哈希表做了 rehash。

1. Redis的kv存儲結構

如下圖所示，Redis的所有kv保存在dict中，其中ht對應兩個哈希表ht[0]和ht[1]，平時一個空閑，一個用于存儲數(shù)據，只有當需要rehash時，ht[1]才會用到。

2. Redis的字典rehash

為了保證哈希表的負載，當哈希表的元素個數(shù)等于哈希表槽數(shù)時候，會進行rehash擴容。擴容后h[1]的容量等于第一個大于等于ht[0].size*2的2n，例如hash表的初始化容量是4，那么下一次擴容就是8，以此類推。

3. 測試

(1) 測試方法

先批量寫入到rehash閾值附近，然后在逐條去寫，觀察內存變化

​// 為每個鍵設置1天過期時間int expireTime = 60 * 60 * 24;// rehash閾值 - 50為了方便觀察rehash內存變化int rehashThreshold = (int) Math.pow(2, 25) - 50;// 1.批量寫入：pipeline批量寫入，由于是本機測試，這里用10000，實際生產不要這么用Pipeline pipeline = jedis.pipelined(); 
pipeline = jedis.pipelined();for (int i = 0; i < rehashThreshold; i++) { 
    pipeline.setex(String.valueOf(i), expireTime, String.valueOf(i));    if (i % 10000 == 0) { 
        pipeline.sync(); 
    } 
} 
pipeline.sync();// 2.等待寫增量TimeUnit.SECONDS.sleep(5);for (int i = rehashThreshold; i < rehashThreshold + 200; i++) { 
    jedis.setex(String.valueOf(i), expireTime, String.valueOf(i)); 
    TimeUnit.SECONDS.sleep(1); 
} 

(2) 開始測試

(a) 當閾值=215=32768，從下面可以看出到key的個數(shù)為32769時，內存漲了一些，但是還不明顯。

​keys       mem      clients blocked requests            connections32766      4.69M    3       0       32797 (+2)          4 
32767      4.69M    3       0       32799 (+2)          4 
32768      4.69M    3       0       32801 (+2)          4 
32769      5.44M    3       0       32803 (+2)          4 

(b) 當閾值=220=1048576，從下面可以看出到key的個數(shù)為1048577時，內存漲了32M。因為rehash會擴容，所以新的哈希表中的槽位變?yōu)榱?21 * 2（因為每個key都設置了過期時間，expires表），指針為8個字節(jié)，221 ? 2 ? 8 = 225 = 32MB。

​keys       mem      clients blocked requests            connections1048574    128.69M  3       0       3364129 (+2)        16 
1048575    128.69M  3       0       3364131 (+2)        16 
1048576    128.69M  3       0       3364133 (+2)        16 
1048577    160.69M  3       0       3364135 (+2)        16 
1048578    160.69M  3       0       3364137 (+2)        16 

(c) 當閾值=226=67108864，從下面可以看出到key的個數(shù)為67108865時，內存漲了2GB。因為rehash會擴容，所以新的哈希表中的槽位變?yōu)榱?27 * 2（因為每個key都設置了過期時間，expires表），指針為8個字節(jié)，227 ? 2 ? 8 = 231 = 2GB。

​keys       mem      clients blocked requests            connections67108862   9.70G    3       0       70473683 (+2)       18 
67108863   9.70G    3       0       70473685 (+2)       18 
67108864   9.70G    3       0       70473687 (+2)       18 
67108865   11.70G   3       0       70473689 (+2)       18 
67108866   11.70G   3       0       70473691 (+2)       18 
67108867   11.70G   3       0       70473693 (+2)       18 

回過來看r-bp1c15fd9b142d04的key和內存變化圖，可以發(fā)現(xiàn)上面的規(guī)則是正確的：

4. 后續(xù)觀察

17點時，rehash結束，內存降了增加的2G的一半。

五、總結

由于哈希表的特性，Redis 中鍵值數(shù)量大，不會對存取造成性能影響，但是會出現(xiàn)本文提到的問題。控制鍵個數(shù)有幾個建議：無用的鍵值設置過期時間或者定期刪除。優(yōu)化鍵值設計：例如可以使用 ziplist hash合并優(yōu)化部分字符串類型。未來改進：內核層面支持 rehash 的審計日志以及增強 rehash 的速度。