一、現象
1. 報警：我們的redis私有云有幾百個redis實例，對於每個實例的存活，都有定時任務會每隔5分鐘，重試多次ping命令去檢查它的存活性，比如多次都ping不通，這時候就會收到報警短信： xxx系統-實例(10.10.xx.xx:xxxx)-由運行中變爲心跳停止。這時候我就會去線上看一下出現了哪些問題。
2. 檢測代碼：比較簡易，但是比較有效。
public boolean isRun(final String ip, final int port) {
boolean isRun = new IdempotentConfirmer() {
private int timeOutFactor = 1;
@Override
public boolean execute() {
Jedis jedis = new Jedis(ip, port);
try {
jedis.getClient().setConnectionTimeout(Protocol.DEFAULT_TIMEOUT * (timeOutFactor++));
jedis.getClient().setSoTimeout(Protocol.DEFAULT_TIMEOUT * (timeOutFactor++));
String pong = jedis.ping();
return pong != null && pong.equalsIgnoreCase("PONG");
} catch (JedisDataException e) {
String message = e.getMessage();
logger.warn(e.getMessage());
if (StringUtils.isNotBlank(message) && message.startsWith("LOADING")) {
return true;
}
return false;
} catch (Exception e) {
logger.warn("{}:{} error message is {} ", ip, port, e.getMessage());
return false;
} finally {
jedis.close();
}
}
}.run();
return isRun;
}package com.sohu.cache.util;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
/**
* 冪等操作器
*/
public abstract class IdempotentConfirmer {
private final Logger logger = LoggerFactory.getLogger(this.getClass());
private int retry = 3;
protected IdempotentConfirmer(int retry) {
this.retry = retry;
}
public IdempotentConfirmer() {
}
public abstract boolean execute();
public boolean run() {
while (retry-- > 0) {
try {
boolean isOk = execute();
if (isOk){
return true;
}
} catch (Exception e) {
logger.error(e.getMessage(), e);
continue;
}
}
return false;
}
}
二、Redis-Cluster的拓撲結構
1. 拓撲圖：
2. 拓撲表
| nodeid | ip:port | replication status |
| b6073bedf256d45e1dce97cd9242bb4789d52343 | 10.10.81.94:7497 | master |
| efc1dfb19e120a2f41e0cec3d03d02f43e168238 | 10.10.79.150:6389 |
slave |
| 792a356d3cd1474b8f1374aba1e5ceced1c6868d | 10.10.81.95:7497 |
master |
| 472823a5f04ad10027b3cde2672cc33f5accdf99 | 10.10.79.157:6390 | slave |
| 19544b078897d8999b68a41f2538a20cbf84288d | 10.10.83.180:6382 | master |
| 74d5cda02571651bba185cbc1b10d5810d048ac3 | 10.10.81.96:6391 | slave |
| b95759d39b6714544917e4aefa383b4de80f871c | 10.10.81.97:7499 | master |
| 3379f2da275825c9ce4e476820254ccaf7b997d1 | 10.10.78.52:6398 | slave |
| 5711268a4881b723273e618894ba9276b27a297f | 10.10.81.98:7496 | master |
| 5b2aa80014f168fd26c246104b5f6adf6426275e | 10.10.78.53:6396 | slave |
三、日誌分析：
1. [10:35:09] 10.10.81.97:7499（b95759d39b6714544917e4aefa383b4de80f871c）發現10.10.81.94:7497（b6073bedf256d45e1dce97cd9242bb4789d52343）不可達：
4051:M 26 Nov 10:35:09.768 * Marking node b6073bedf256d45e1dce97cd9242bb4789d52343 as failing (quorum reached). 
2.10.10.81.97:7499（b95759d39b6714544917e4aefa383b4de80f871c）通過gossip協議廣播10.10.81.94:7497（b6073bedf256d45e1dce97cd9242bb4789d52343）失敗的信息，其他實例收到了這個流言：
29177:S 26 Nov 10:35:10.939 * FAIL message received from b95759d39b6714544917e4aefa383b4de80f871c about b6073bedf256d45e1dce97cd9242bb4789d52343
3.投票：（在redisCluster中，只有Master會參與投票，因爲從節點只是負責備份用），進入新的epoch
9611:M 26 Nov 10:35:12.068 # Failover auth granted to efc1dfb19e120a2f41e0cec3d03d02f43e168238 for epoch 18
redis源碼：
// 主機已經是一個出錯節點了，自己作爲從機可以升級爲主機
void clusterHandleSlaveFailover(void) {
......
// 故障修復超時，重新啓動故障修復
if (auth_age > auth_retry_time) { // 兩次故障修復間隔不能過短
// 更新一些時間
......
redisLog(REDIS_WARNING,
"Start of election delayed for %lld milliseconds "
"(rank #%d, offset %lld).",
server.cluster->failover_auth_time - mstime(),
server.cluster->failover_auth_rank,
replicationGetSlaveOffset());
// 告知其他從機
/* Now that we have a scheduled election, broadcast our offset
* to all the other slaves so that they'll updated their offsets
* if our offset is better. */
clusterBroadcastPong(CLUSTER_BROADCAST_LOCAL_SLAVES);
return;
}
......
// 開頭投票
/* Ask for votes if needed. */
if (server.cluster->failover_auth_sent == 0) {
server.cluster->currentEpoch++;
server.cluster->failover_auth_epoch = server.cluster->currentEpoch;
redisLog(REDIS_WARNING,"Starting a failover election for epoch %llu.",
(unsigned long long) server.cluster->currentEpoch);
clusterRequestFailoverAuth();
server.cluster->failover_auth_sent = 1;
clusterDoBeforeSleep(CLUSTER_TODO_SAVE_CONFIG|
CLUSTER_TODO_UPDATE_STATE|
CLUSTER_TODO_FSYNC_CONFIG);
return; /* Wait for replies. */
}
}
4. 投票結果:
(1) 10.10.79.150:6389是新的master, 成功執行了failover
3173:S 26 Nov 10:35:11.003 # Start of election delayed for 515 milliseconds (rank #0, offset 21516181352).
3173:S 26 Nov 10:35:11.606 # Starting a failover election for epoch 18.
3173:S 26 Nov 10:35:11.660 # Failover election won: I'm the new master.
3173:S 26 Nov 10:35:11.660 # configEpoch set to 18 after successful failover
(2) 原來的master:10.10.81.94:7497(主) ，收到了判決書：節點失敗(fail), 授權10.10.79.150:6389爲它的master, 修改自己的cluster nodes, 沒有了slave, cluster status ok, 開始主從複製...
3022:M 26 Nov 10:35:15.560 # Cluster state changed: fail
3022:M 26 Nov 10:35:15.562 # Failover auth denied to efc1dfb19e120a2f41e0cec3d03d02f43e168238: its master is up
3022:M 26 Nov 10:35:15.595 # Configuration change detected. Reconfiguring myself as a replica of efc1dfb19e120a2f41e0cec3d03d02f43e168238
3022:S 26 Nov 10:35:15.595 # Connection with slave 10.10.79.150:6389 lost.
3022:S 26 Nov 10:35:15.616 # Cluster state changed: ok
.................
(3)新的主從通信成功，主從關係確定
5. 其他機器開始，認定新的slave
7925:S 26 Nov 10:36:20.879 * Marking node b6073bedf256d45e1dce97cd9242bb4789d52343 as failing (quorum reached).
7925:S 26 Nov 10:35:33.603 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.
6. 完整日誌
(1). 10.10.81.94:7497(主)
3022:M 26 Nov 10:35:15.560 # Cluster state changed: fail
3022:M 26 Nov 10:35:15.562 # Failover auth denied to efc1dfb19e120a2f41e0cec3d03d02f43e168238: its master is up
3022:M 26 Nov 10:35:15.595 # Configuration change detected. Reconfiguring myself as a replica of efc1dfb19e120a2f41e0cec3d03d02f43e168238
3022:S 26 Nov 10:35:15.595 # Connection with slave 10.10.79.150:6389 lost.
3022:S 26 Nov 10:35:15.616 # Cluster state changed: ok
3022:S 26 Nov 10:35:16.061 * Connecting to MASTER 10.10.79.150:6389
3022:S 26 Nov 10:35:16.061 * MASTER <-> SLAVE sync started
3022:S 26 Nov 10:35:16.062 * Non blocking connect for SYNC fired the event.
3022:S 26 Nov 10:35:16.062 * Master replied to PING, replication can continue...
3022:S 26 Nov 10:35:16.062 * Partial resynchronization not possible (no cached master)
3022:S 26 Nov 10:35:16.062 * Full resync from master: 80ac45a2b236fc44ebbef3d851e535bcbdd6367b:21516181353
3022:S 26 Nov 10:35:45.273 * MASTER <-> SLAVE sync: receiving 494097946 bytes from master
3022:S 26 Nov 10:35:51.052 * MASTER <-> SLAVE sync: Flushing old data
3022:S 26 Nov 10:36:22.165 * MASTER <-> SLAVE sync: Loading DB in memory
3022:S 26 Nov 10:37:44.355 * MASTER <-> SLAVE sync: Finished with success
3022:S 26 Nov 10:37:44.440 * Background append only file rewriting started by pid 4036
3022:S 26 Nov 10:38:34.988 * AOF rewrite child asks to stop sending diffs.
4036:C 26 Nov 10:38:34.988 * Parent agreed to stop sending diffs. Finalizing AOF...
4036:C 26 Nov 10:38:34.988 * Concatenating 0.02 MB of AOF diff received from parent.
4036:C 26 Nov 10:38:34.989 * SYNC append only file rewrite performed
4036:C 26 Nov 10:38:35.029 * AOF rewrite: 5 MB of memory used by copy-on-write
3022:S 26 Nov 10:38:35.127 * Background AOF rewrite terminated with success
3022:S 26 Nov 10:38:35.127 * Residual parent diff successfully flushed to the rewritten AOF (0.00 MB)
3022:S 26 Nov 10:38:35.128 * Background AOF rewrite finished successfully
(2) 10.10.79.150:6389(從)
3173:S 26 Nov 10:35:10.934 * FAIL message received from b95759d39b6714544917e4aefa383b4de80f871c about b6073bedf256d45e1dce97cd9242bb4789d52343
3173:S 26 Nov 10:35:11.003 # Start of election delayed for 515 milliseconds (rank #0, offset 21516181352).
3173:S 26 Nov 10:35:11.606 # Starting a failover election for epoch 18.
3173:S 26 Nov 10:35:11.660 # Failover election won: I'm the new master.
3173:S 26 Nov 10:35:11.660 # configEpoch set to 18 after successful failover
3173:M 26 Nov 10:35:11.660 # Connection with master lost.
3173:M 26 Nov 10:35:11.660 * Caching the disconnected master state.
3173:M 26 Nov 10:35:11.660 * Discarding previously cached master state.
3173:M 26 Nov 10:35:15.605 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: master without slots is reachable again.
3173:M 26 Nov 10:35:16.071 * Slave 10.10.81.94:7497 asks for synchronization
3173:M 26 Nov 10:35:16.071 * Full resync requested by slave 10.10.81.94:7497
3173:M 26 Nov 10:35:16.071 * Starting BGSAVE for SYNC with target: disk
3173:M 26 Nov 10:35:16.188 * Background saving started by pid 23531
3173:M 26 Nov 10:35:18.091 * Asynchronous AOF fsync is taking too long (disk is busy?). Writing the AOF buffer without waiting for fsync to complete, this may slow down Redis.
3173:M 26 Nov 10:35:31.056 * Asynchronous AOF fsync is taking too long (disk is busy?). Writing the AOF buffer without waiting for fsync to complete, this may slow down Redis.
3173:M 26 Nov 10:35:40.040 * Asynchronous AOF fsync is taking too long (disk is busy?). Writing the AOF buffer without waiting for fsync to complete, this may slow down Redis.
23531:C 26 Nov 10:35:45.048 * DB saved on disk
23531:C 26 Nov 10:35:45.113 * RDB: 6 MB of memory used by copy-on-write
3173:M 26 Nov 10:35:45.281 * Background saving terminated with success
3173:M 26 Nov 10:35:51.059 * Synchronization with slave 10.10.81.94:7497 succeeded
3173:M 26 Nov 10:36:11.107 * FAIL message received from 792a356d3cd1474b8f1374aba1e5ceced1c6868d about b6073bedf256d45e1dce97cd9242bb4789d52343
3173:M 26 Nov 10:36:22.277 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.
(3) 10.10.81.95:7497(主)
14795:M 26 Nov 10:35:10.309 * FAIL message received from b95759d39b6714544917e4aefa383b4de80f871c about b6073bedf256d45e1dce97cd9242bb4789d52343
14795:M 26 Nov 10:35:11.034 # Failover auth granted to efc1dfb19e120a2f41e0cec3d03d02f43e168238 for epoch 18
14795:M 26 Nov 10:35:14.979 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: master without slots is reachable again.
14795:M 26 Nov 10:36:20.480 * Marking node b6073bedf256d45e1dce97cd9242bb4789d52343 as failing (quorum reached).
14795:M 26 Nov 10:36:31.652 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.(4) 10.10.79.157:6390(從)
29177:S 26 Nov 10:35:10.939 * FAIL message received from b95759d39b6714544917e4aefa383b4de80f871c about b6073bedf256d45e1dce97cd9242bb4789d52343
29177:S 26 Nov 10:35:15.609 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: master without slots is reachable again.
29177:S 26 Nov 10:36:11.111 * FAIL message received from 792a356d3cd1474b8f1374aba1e5ceced1c6868d about b6073bedf256d45e1dce97cd9242bb4789d52343
29177:S 26 Nov 10:36:22.281 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.
(5) 10.10.83.180:6382(主)
9611:M 26 Nov 10:35:11.342 * FAIL message received from b95759d39b6714544917e4aefa383b4de80f871c about b6073bedf256d45e1dce97cd9242bb4789d52343
9611:M 26 Nov 10:35:12.068 # Failover auth granted to efc1dfb19e120a2f41e0cec3d03d02f43e168238 for epoch 18
9611:M 26 Nov 10:35:16.013 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: master without slots is reachable again.
9611:M 26 Nov 10:36:11.516 * FAIL message received from 792a356d3cd1474b8f1374aba1e5ceced1c6868d about b6073bedf256d45e1dce97cd9242bb4789d52343
9611:M 26 Nov 10:36:22.686 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.
(6) 10.10.81.96:6391(從)
7925:S 26 Nov 10:35:10.260 * FAIL message received from b95759d39b6714544917e4aefa383b4de80f871c about b6073bedf256d45e1dce97cd9242bb4789d52343
7925:S 26 Nov 10:35:14.930 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: master without slots is reachable again.
7925:S 26 Nov 10:36:20.879 * Marking node b6073bedf256d45e1dce97cd9242bb4789d52343 as failing (quorum reached).
7925:S 26 Nov 10:35:33.603 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.
(7) 10.10.81.97:7499(主)
4051:M 26 Nov 10:35:09.768 * Marking node b6073bedf256d45e1dce97cd9242bb4789d52343 as failing (quorum reached).
4051:M 26 Nov 10:35:10.507 # Failover auth granted to efc1dfb19e120a2f41e0cec3d03d02f43e168238 for epoch 18
4051:M 26 Nov 10:35:14.452 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: master without slots is reachable again.
4051:M 26 Nov 10:36:09.955 * FAIL message received from 792a356d3cd1474b8f1374aba1e5ceced1c6868d about b6073bedf256d45e1dce97cd9242bb4789d52343
4051:M 26 Nov 10:36:21.125 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.
(8) 10.10.78.52:6398(從)
22972:S 26 Nov 10:35:10.944 * FAIL message received from b95759d39b6714544917e4aefa383b4de80f871c about b6073bedf256d45e1dce97cd9242bb4789d52343
22972:S 26 Nov 10:35:15.635 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: master without slots is reachable again.
22972:S 26 Nov 10:36:11.116 * FAIL message received from 792a356d3cd1474b8f1374aba1e5ceced1c6868d about b6073bedf256d45e1dce97cd9242bb4789d52343
22972:S 26 Nov 10:36:22.286 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.
(9) 10.10.81.98:7496(主)
2062:M 26 Nov 10:35:10.908 * FAIL message received from b95759d39b6714544917e4aefa383b4de80f871c about b6073bedf256d45e1dce97cd9242bb4789d52343
2062:M 26 Nov 10:35:11.634 # Failover auth granted to efc1dfb19e120a2f41e0cec3d03d02f43e168238 for epoch 18
2062:M 26 Nov 10:35:15.599 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: master without slots is reachable again.
2062:M 26 Nov 10:36:11.080 * FAIL message received from 792a356d3cd1474b8f1374aba1e5ceced1c6868d about b6073bedf256d45e1dce97cd9242bb4789d52343
2062:M 26 Nov 10:36:22.250 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.
(10) 10.10.78.53:6396(從)
19527:S 26 Nov 10:35:10.944 * FAIL message received from b95759d39b6714544917e4aefa383b4de80f871c about b6073bedf256d45e1dce97cd9242bb4789d52343
19527:S 26 Nov 10:35:15.615 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: master without slots is reachable again.
19527:S 26 Nov 10:36:11.117 * FAIL message received from 792a356d3cd1474b8f1374aba1e5ceced1c6868d about b6073bedf256d45e1dce97cd9242bb4789d52343
19527:S 26 Nov 10:36:22.287 * Clear FAIL state for node b6073bedf256d45e1dce97cd9242bb4789d52343: slave is reachable again.
四、原因查找
redis節點被認爲失敗，可能有如下原因：
1. 機器本身問題：負載過高，網絡阻塞，cpu過高等等。
通過tsar命令，查詢機器當時各個指標，未發現有什麼問題：
Time ---cpu-- ---mem-- ---tcp-- -----traffic---- --sda--- --sdb--- --sdc--- --sdd--- --sde--- --sdf--- --sdg--- --sdh--- --sdi--- --sdj--- --sdk--- --sdl--- --sdm--- ---load-
Time util util retran bytin bytout util util util util util util util util util util util util util load1
26/11/15-10:20 4.07 67.89 0.00 1.7M 1.7M 0.05 0.00 1.75 0.00 8.07 0.00 0.00 0.00 0.00 0.00 0.38 0.25 0.00 1.77
26/11/15-10:25 4.11 67.89 0.00 1.7M 1.6M 0.01 0.00 1.95 0.00 7.45 0.00 0.00 0.00 0.00 0.00 3.47 0.27 0.00 0.94
26/11/15-10:30 4.37 67.89 0.00 1.7M 1.7M 0.18 0.00 2.00 0.00 7.79 0.00 0.00 0.00 0.00 0.00 5.31 0.15 0.00 0.94
26/11/15-10:35 4.86 68.38 0.00 1.8M 1.7M 0.47 0.00 2.40 0.00 7.99 0.00 0.00 0.00 0.00 0.00 3.06 0.15 0.00 1.22
26/11/15-10:40 7.29 67.66 0.00 2.9M 1.6M 1.55 0.00 2.92 0.00 8.46 0.00 0.00 0.00 0.00 0.00 0.87 0.11 0.00 1.14
26/11/15-10:45 4.08 67.68 0.00 1.2M 1.6M 0.09 0.00 0.34 0.00 9.12 0.00 0.00 0.00 0.00 0.00 0.25 1.35 0.00 0.46
2. cluster-node-timeout: 這個我們設置的是15秒，一直未發現問題，從別人的分享看，這個數字已經足夠了。
3. 被判斷失敗的redis本身的問題：阻塞操作(比較常見的有: aof阻塞、慢查詢）：從redis的日誌看沒有發生aof阻塞的情況，但是從慢查詢的日誌中，發現瞭如下，瞬間就明白了，許多慢查詢，redis的單線程什麼都堵住了。
五、注意：
本文的日誌時間做過糾正，實際有些機器沒有做時間同步，造成日誌實際對不上，不利排查，而且在分佈式系統中，時間的一致性通常非常重要，所以這個要注意。
參考文獻：
- Random and small clusterdown(github redis)
- redis集羣定時clusterdown分析