Ok, but could you please explain which problem are you tying to solve? If table is partially shutdown, then status.is_leader must be false and everything should work fine, because DDLWorker takes it into account and tries to execute query on other replica. If it's not possible for some reason, then query fails with Task ... was not executed by anyone ... after several retries.

@tavplubix If all replica is dropped, status.is_leader is false on all replicas, so that the number of tries will not be updated by any replica. And it will have to wait for timeout currently it is MAX_EXECUTION_TIMEOUT_SEC = 3600s. It is too long and will blocking other DDL tasks.

You're right, there is a bug in DDLWorker. However, DROP TABLE is not the only thing that partially shutdowns replicated table (and that's why test_ddl_worker_non_leader/test.py::test_non_leader_replica failed), it also happens on lost ZooKeeper connection and SYSTEM RESTART REPLICA. I'm not sure if we should finish DDL query with Table is shutdown ... in these cases. Maybe check IStorage::is_dropped flag instead of partial_shutdown_called? On the other hand, is_dropped is false for detached tables, but DETACH/ATTACH are usually used to restart replica (and that's how SYSTEM RESTART REPLICA works). To avoid waiting on permanently detached tables we can use something like this:

bool replica_dropped = storage->is_dropped;
bool all_replicas_likely_detached = status.active_replicas == 0 && !DatabaseCatalog::instance().isTableExist(table_id, context);
if (replica_dropped || all_replicas_likely_detached) /// Table is shutdown ...

What do you think?

Btw, could you please add a test? It can be simple stateless test with test_shard_localhost cluster with one replica.

@tavplubix You are right. I have another commit. ^-^
But I find it is hard to test it, because the table has to be dropped during the ddl task's running stage. It is very difficult to make such a test. Could you please provide some idea?

Take a look at *.sh tests (such as 01150_ddl_guard_rwr and 00993_system_parts_race_condition_drop_zookeeper). You can run two threads, one thread will execute CREATE and DROP and other thread will execute ALTER ... ON CLUSTER.