Sentinel: when leader in wait-start, sense another leader as race.

[redis.git] / src / sentinel.c

diff --git a/src/sentinel.c b/src/sentinel.c
index d1c6befe2b8de44894181c52dc6b76b89d8e7b23..0c314745e9e6f3cced8050ab93b60d6b9142c04f 100644 (file)
--- a/src/sentinel.c
+++ b/src/sentinel.c
@@ -1384,6 +1384,7 @@ void sentinelRefreshInstanceInfo(sentinelRedisInstance *ri, const char *info) {
     ri->info_refresh = mstime();
     sdsfreesplitres(lines,numlines);
 
+    /* ---------------------------- Acting half ----------------------------- */
     if (sentinel.tilt) return;
 
     /* Act if a master turned into a slave. */
@@ -1407,11 +1408,12 @@ void sentinelRefreshInstanceInfo(sentinelRedisInstance *ri, const char *info) {
         if (!(ri->master->flags & SRI_FAILOVER_IN_PROGRESS) &&
             (runid_changed || first_runid))
         {
-            /* If a slave turned into a master, but at the same time the
-             * runid has changed, or it is simply the first time we see and
-             * INFO output from this instance, this is a reboot with a wrong
-             * configuration.
+            /* If a slave turned into maser but:
              *
+             * 1) Failover not in progress.
+             * 2) RunID hs changed, or its the first time we see an INFO output.
+             * 
+             * We assume this is a reboot with a wrong configuration.
              * Log the event and remove the slave. */
             int retval;
 
@@ -1422,8 +1424,7 @@ void sentinelRefreshInstanceInfo(sentinelRedisInstance *ri, const char *info) {
         } else if (ri->flags & SRI_PROMOTED) {
             /* If this is a promoted slave we can change state to the
              * failover state machine. */
-            if (ri->master &&
-                (ri->master->flags & SRI_FAILOVER_IN_PROGRESS) &&
+            if ((ri->master->flags & SRI_FAILOVER_IN_PROGRESS) &&
                 (ri->master->flags & SRI_I_AM_THE_LEADER) &&
                 (ri->master->failover_state ==
                     SENTINEL_FAILOVER_STATE_WAIT_PROMOTION))
@@ -1434,24 +1435,36 @@ void sentinelRefreshInstanceInfo(sentinelRedisInstance *ri, const char *info) {
                 sentinelEvent(REDIS_WARNING,"+failover-state-reconf-slaves",
                     ri->master,"%@");
             }
-        } else {
-            /* Otherwise we interpret this as the start of the failover. */
-            if (ri->master &&
-                (ri->master->flags & SRI_FAILOVER_IN_PROGRESS) == 0)
-            {
-                ri->master->flags |= SRI_FAILOVER_IN_PROGRESS;
-                sentinelEvent(REDIS_WARNING,"failover-detected",ri->master,"%@");
-                ri->master->failover_state = SENTINEL_FAILOVER_STATE_DETECT_END;
-                ri->master->failover_state_change_time = mstime();
-                ri->master->promoted_slave = ri;
-                ri->flags |= SRI_PROMOTED;
-                /* We are an observer, so we can only assume that the leader
-                 * is reconfiguring the slave instances. For this reason we
-                 * set all the instances as RECONF_SENT waiting for progresses
-                 * on this side. */
-                sentinelAddFlagsToDictOfRedisInstances(ri->master->slaves,
-                    SRI_RECONF_SENT);
+        } else if (!(ri->master->flags & SRI_FAILOVER_IN_PROGRESS) ||
+                    ((ri->master->flags & SRI_FAILOVER_IN_PROGRESS) &&
+                     (ri->master->flags & SRI_I_AM_THE_LEADER) &&
+                     ri->master->failover_state ==
+                     SENTINEL_FAILOVER_STATE_WAIT_START))
+        {
+            /* No failover in progress? Then it is the start of a failover
+             * and we are an observer.
+             *
+             * We also do that if we are a leader doing a failover, in wait
+             * start, but well, somebody else started before us. */
+
+            if (ri->master->flags & SRI_FAILOVER_IN_PROGRESS) {
+                sentinelEvent(REDIS_WARNING,"-failover-abort-race",
+                                ri->master, "%@");
+                sentinelAbortFailover(ri->master);
             }
+
+            ri->master->flags |= SRI_FAILOVER_IN_PROGRESS;
+            sentinelEvent(REDIS_WARNING,"+failover-detected",ri->master,"%@");
+            ri->master->failover_state = SENTINEL_FAILOVER_STATE_DETECT_END;
+            ri->master->failover_state_change_time = mstime();
+            ri->master->promoted_slave = ri;
+            ri->flags |= SRI_PROMOTED;
+            /* We are an observer, so we can only assume that the leader
+             * is reconfiguring the slave instances. For this reason we
+             * set all the instances as RECONF_SENT waiting for progresses
+             * on this side. */
+            sentinelAddFlagsToDictOfRedisInstances(ri->master->slaves,
+                SRI_RECONF_SENT);
         }
     }