Online alter (LOCK=NONE, which is default) works as follows:
1. Acquire an exclusive MDL lock for a short duration
2. Acquire TL_READ
3. Downgrade MDL to MDL_SHARED_UPGRADABLE to allow concurrent writes
4. Copy and replicate online changes
5. Unlock TL
6. Replicate the rest of online changes
5. Upgrade to MDL_EXCLUSIVE
6. Unlock all

This means that now MCS can end up with inconsistent data if some concurrent writes were executed during alter table:

con1: ALTER TABLE t ....

con1: # concider it reached stage 4

con2: INSERT y; COMMIT;

this may end up in the record y inserted twice in the resulting table.

Context: MDL – Metadata lock, multi-granular lock that protects table metadata. Represented by MDL_request class and TABLE_LIST::mdl_request field respectively. See enum_mdl_type for the granularity types and MDL_lock::MDL_object_lock::m_granted_incompatible for the compatibility matrix.
TL lock - a table data lock represented by thr_lock_type enum and stored in TABLE_LIST::lock_type and TABLE::reg_info.lock_type. Typically propagated to a stronger lock type then provided by a storage engine in handler::store_lock. Refer to mysql_lock_tables/mysql_unlock_tables for further info.

Reproduction:
1. Compile the sql server with -DCMAKE_BUILD_TYPE=Debug
2. Run the mtr test

create or replace table t1 (a int, b int, key(a)) engine=columnstore;

insert t1 values (1, 11);

insert t1 values (2, 22);

insert t1 values (3, 33);

insert t1 values (4, 44);

insert t1 values (5, 55);

set debug_sync= 'alter_table_online_downgraded SIGNAL downgraded WAIT_FOR ins';

send alter table t1 add c int algorithm= copy, lock= none;

--connect (con1, localhost, root,,)

set debug_sync= 'now WAIT_FOR downgraded';

set lock_wait_timeout= 0; # the correct behavior is to wait until alter table ends

--error 0,ER_LOCK_WAIT_TIMEOUT

insert t1 values (7, 77); # autocommit

set debug_sync= 'now SIGNAL ins';

--connection default

reap;

# row 7, 77 should not happen twice! (but can be 0 times in case of lock timeout)

select * from t1;