一次關於Redis記憶體詭異增長的排查過程實戰記錄
一、現象
例項名：r-bp1cxxxxxxxxxd04(主從)
問題：一分鐘記憶體上漲了2G，如下圖所示：
鍵值規模：6000萬左右
記憶體一分鐘增長2G.png
二、Redis記憶體分析
1. 記憶體組成
上圖中的記憶體統計的是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
| 屬性名 | 屬性說明 |
|---|---|
| used_memory | Redis 分配器分配的記憶體量，也就是實際儲存資料的記憶體總量 |
| used_memory_human | 以可讀格式返回 Redis 使用的記憶體總量 |
| used_memory_rss | 從作業系統的角度，Redis程序佔用的總實體記憶體 |
| used_memory_peak | 記憶體分配器分配的最大記憶體，代表used_memory的歷史峰值 |
| used_memory_peak_human | 以可讀的格式顯示記憶體消耗峰值 |
| used_memory_lua | Lua引擎所消耗的記憶體 |
| mem_fragmentation_ratio | used_memory_rss /used_memory比值，表示記憶體碎片率 |
| mem_allocator | Redis 所使用的記憶體分配器。預設: jemalloc |
計算公式如下：
used_memory = 自身記憶體+物件記憶體+緩衝記憶體+lua記憶體used_rss = used_memory + 記憶體碎片
如下圖所示：
2. 記憶體分析
(1) 自身記憶體：一個空的Redis佔用很小，可以忽略不計
(2) kv記憶體：key物件 + value物件
(3) 緩衝區：客戶端緩衝區(普通 + slave偽裝 + pubsub)以及aof緩衝區(比較固定，一般沒問題)
(4) Lua：Lua引擎所消耗的記憶體
3. 記憶體突增常見問題
(1) kv記憶體：bigkey、大量寫入
(2) 客戶端緩衝區：一般常見的有普通客戶端緩衝區(例如monitor命令)或者pubsub客戶端緩衝區
三、問題排查
(1) bigkey ? 經掃描未發現bigkey
Sampled 67234427 keys in the keyspace! Total key length in bytes is 1574032382 (avg len 23.41) Biggest string found 'CCARD_DEVICE_CARD_REF_MAP_KEY_016817000004209' has 20862 bytes Biggest list found 'CCARD_VALID_DEVICE_TRAIN_QUEUE_KEY' has 51 items Biggest hash found 'CCARD_VALID_DEVICE_TRAIN_MAP_KEY' has 51 fields 67234359 strings with 71767890 bytes (100.00% of keys,avg size 1.07) 67 lists with 151 items (00.00% of keys,avg size 2.25) 0 sets with 0 members (00.00% of keys,avg size 0.00) 1 hashs with 51 fields (00.00% of keys,avg size 51.00) 0 zsets with 0 members (00.00% of keys,avg size 0.00)
(2) 鍵值個數增加？未發現鍵值有明顯變化
(3) 客戶端緩衝區
由於記憶體增上去後，長時間沒下落，如果是因為緩衝區問題，會從info clients找到明顯問題，執行後發現：
redis> info clients # Clients connected_clients:43 client_longest_output_list:0 client_biggest_input_buf:0 blocked_clients:0 admin_clients:6 rejected_vpc_conn_count:0 close_idle_unknown_conn_count:0
執行client中也沒有明顯的omem大於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、揪出元凶
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雜湊表做了 rehash。
1. Redis的kv儲存結構
如下圖所示，Redis的所有kv儲存在dict中，其中ht對應兩個雜湊表ht[0]和ht[1]，平時一個空閒，一個用於儲存資料，只有當需要rehash時，ht[1]才會用到。
2. Redis的字典rehash
為了保證雜湊表的負載，當雜湊表的元素個數等於雜湊表槽數時候，會進行rehash擴容。擴容後h[1]的容量等於第一個大於等於ht[0].size*2的2n，例如hash表的初始化容量是4，那麼下一次擴容就是8，以此類推。
3. 測試
(1) 測試方法
先批量寫入到rehash閾值附近，然後在逐條去寫，觀察記憶體變化
// 為每個鍵設定1天過期時間
int expireTime = 60 * 60 * 24;
// rehash閾值 - 50為了方便觀察rehash記憶體變化
int rehashThreshold = (int) Math.pow(2,25) - 50;
// 1.批量寫入：pipeline批量寫入，由於是本機測試，這裡用10000，實際生產不要這麼用
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),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的個數為32769時，記憶體漲了一些，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的個數為1048577時，記憶體漲了32M。因為rehash會擴容，所以新的雜湊表中的槽位變為了221 * 2（因為每個key都設定了過期時間，expires表），指標為8個位元組，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的個數為67108865時，記憶體漲了2GB。因為rehash會擴容，所以新的雜湊表中的槽位變為了227 * 2（因為每個key都設定了過期時間，expires表），指標為8個位元組，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和記憶體變化圖，可以發現上面的規則是正確的：
4. 後續觀察
17點時，rehash結束，記憶體降了增加的2G的一半。
五、總結
由於雜湊表的特性，Redis 中鍵值數量大，不會對存取造成效能影響，但是會出現本文提到的問題。控制鍵個數有幾個建議：無用的鍵值設定過期時間或者定期刪除。優化鍵值設計：例如可以使用 ziplist hash合併優化部分字串型別。未來改進：核心層面支援 rehash 的審計日誌以及增強 rehash 的速度。
好了，以上就是這篇文章的全部內容了，希望本文的內容對大家的學習或者工作具有一定的參考學習價值，如果有疑問大家可以留言交流，謝謝大家對我們的支援。