Merge pull request ClickHouse#88906 from ClickHouse/backport/25.8/88671

Backport ClickHouse#88671 to 25.8: Don't try to mark non-existing entries as finished when recovering a Replicated database replica

Author: clickhouse-gh[bot]

src/Databases/DatabaseReplicated.cpp

@@ -1620,6 +1620,16 @@ void DatabaseReplicated::recoverLostReplica(const ZooKeeperPtr & current_zookeep
     /// It's a very rare case, and it's okay if some queries throw TIMEOUT_EXCEEDED when waiting for all replicas
     if (first_entry_to_mark_finished)
     {
+        /// Skip non-existing entries that were removed a long time ago (if the replica was offline for a long time)
+        Strings all_nodes = current_zookeeper->getChildren(fs::path(zookeeper_path) / "log");
+        std::erase_if(all_nodes, [] (const String & s) { return !startsWith(s, "query-"); });
+        auto oldest_node = std::min_element(all_nodes.begin(), all_nodes.end());
+        if (oldest_node != all_nodes.end())
+        {
+            UInt32 oldest_entry = DDLTaskBase::getLogEntryNumber(*oldest_node);
+            first_entry_to_mark_finished = std::max(oldest_entry, first_entry_to_mark_finished);
+        }
+
         /// If the replica is new and some of the queries applied during recovery
         /// where issued after the replica was created, then other nodes might be
         /// waiting for this node to notify them that the query was applied.