From 545869eb7d9365bf8e41683f6d2c711882b071aa Mon Sep 17 00:00:00 2001
From: Su Ko <rhtn1128@gmail.com>
Date: Sun, 23 Nov 2025 16:21:39 +0900
Subject: [PATCH] Add replica sync checks before rebalancing in cluster manager

Signed-off-by: Su Ko <rhtn1128@gmail.com>
---
 src/valkey-cli.c | 102 ++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 101 insertions(+), 1 deletion(-)

diff --git a/src/valkey-cli.c b/src/valkey-cli.c
index c902d1255f..fd919f3a38 100644
--- a/src/valkey-cli.c
+++ b/src/valkey-cli.c
@@ -3714,7 +3714,8 @@ static void clusterManagerNodeArrayInit(clusterManagerNodeArray *array, int len)
 static void clusterManagerNodeArrayReset(clusterManagerNodeArray *array);
 static void clusterManagerNodeArrayShift(clusterManagerNodeArray *array, clusterManagerNode **nodeptr);
 static void clusterManagerNodeArrayAdd(clusterManagerNodeArray *array, clusterManagerNode *node);
-
+static int clusterManagerIsReplicaSynced(clusterManagerNode *node, char **err);
+static int clusterManagerWaitReplicasFullSync(long long timeout_ms);
 /* Cluster Manager commands. */
 
 static int clusterManagerCommandCreate(int argc, char **argv);
@@ -6786,6 +6787,95 @@ static void clusterManagerMode(clusterManagerCommandProc *proc) {
     exit(success ? 0 : 1);
 }
 
+static int clusterManagerIsReplicaSynced(clusterManagerNode *node, char **err) {
+    if (err) *err = NULL;
+
+    if (node == NULL || node->replicate == NULL) {
+        if (err) *err = sdsnew("Node is not a replica");
+        return 0;
+    }
+
+    valkeyReply *r = CLUSTER_MANAGER_COMMAND(node, "INFO replication");
+    if (!clusterManagerCheckValkeyReply(node, r, err)) {
+        if (r) freeReplyObject(r);
+        return 0;
+    }
+
+    if (r->type != VALKEY_REPLY_STRING) {
+        if (err) *err = sdsnew("Unexpected INFO replication reply type");
+        freeReplyObject(r);
+        return 0;
+    }
+
+    int link_up = 0;
+    int sync_in_progress = 0;
+
+    char *str = r->str;
+    char *p = strstr(str, "master_link_status:");
+    if (p) {
+        p += strlen("master_link_status:");
+        if (!strncmp(p, "up", 2)) link_up = 1;
+    }
+
+    p = strstr(str, "master_sync_in_progress:");
+    if (p) {
+        p += strlen("master_sync_in_progress:");
+        sync_in_progress = (int)strtol(p, NULL, 10);
+    }
+
+    freeReplyObject(r);
+
+    if (link_up && sync_in_progress == 0)
+        return 1;
+
+    if (err) *err = sdsnew("Replica still syncing");
+    return 0;
+}
+
+static int clusterManagerWaitReplicasFullSync(long long timeout_ms) {
+    long long start = mstime();
+    long long backoff = 100;
+    long long max_backoff = 2000;
+
+    while (mstime() - start < timeout_ms) {
+
+        int all_good = 1;
+        listIter li;
+        listNode *ln;
+
+        listRewind(cluster_manager.nodes, &li);
+        while ((ln = listNext(&li)) != NULL) {
+            clusterManagerNode *node = ln->value;
+
+            if (node->replicate == NULL) continue;
+
+            char *err = NULL;
+            if (!clusterManagerIsReplicaSynced(node, &err)) {
+                all_good = 0;
+
+                if (config.verbose) {
+                    if (err)
+                        clusterManagerLogInfo("[replica-sync] %s:%d not synced: %s\n",
+                                              node->ip, node->port, err);
+                    else
+                        clusterManagerLogInfo("[replica-sync] %s:%d not synced\n",
+                                              node->ip, node->port);
+                }
+                if (err) sdsfree(err);
+            }
+        }
+
+        if (all_good) return 1;
+
+        usleep(backoff * 1000);
+        backoff = backoff * 2;
+        if (backoff > max_backoff) backoff = max_backoff;
+    }
+
+    clusterManagerLogErr("[ERR] replica full-sync wait timed-out.\n");
+    return 0;
+}
+
 /* Cluster Manager Commands */
 
 static int clusterManagerCommandCreate(int argc, char **argv) {
@@ -7567,6 +7657,16 @@ static int clusterManagerCommandRebalance(int argc, char **argv) {
         result = 0;
         goto cleanup;
     }
+
+    clusterManagerLogInfo(">>> Checking replica sync state before rebalancing...\n");
+
+    if (!clusterManagerWaitReplicasFullSync(30000)) { /* 30 seconds timeout */
+        clusterManagerLogErr("[ERR] Replica sync incomplete. Aborting rebalance.\n");
+        return 0;
+    }
+
+    clusterManagerLogInfo(">>> All replicas are online. Safe to proceed with rebalancing.\n");
+
     /* Calculate the slots balance for each node. It's the number of
      * slots the node should lose (if positive) or gain (if negative)
      * in order to be balanced. */