In short:
Atomic:
non-blocking drop table / rename table
tables delete (&detach) async (wait for selects finish but invisible for new selects)
atomic drop table (all files / folders removed)
atomic table swap (table swap by "EXCHANGE TABLES t1 AND t2;")
rename dictionary / rename database
unique automatic UUID paths in FS and ZK for Replicated

Ok I see, thanks. It's more flexible for some actions on cluster.

Last question, it's possible to convert a Ordinary database to Atomic ? If not it will ne possible later ?

AFAIK you already can move tables to atomic database using rename table. But later will be implemented a special method to convert database;

create database atom Engine=Atomic;
create database ord Engine=Ordinary;
create table ord.A (A Int64) Engine=MergeTree order by A;
insert into ord.A select number from numbers(1000);

rename table ord.A to atom.A;

select count() from atom.A;
┌─count()─┐
│    1000 │
└─────────┘

Ho nice ! Thanks.

So in this case I can start the project with Ordinary and migrate later when backup will support it.

Thank you it's clear.

@den-crane is ALTER TABLE non-blocking in Atomic engine?

Yes, ALTER TABLE is non-blocking in Atomic databases, while in Ordinary databases it blocks DROP and RENAME queries.