max_parallel_replicas delivers incorrect result for tables without sampling.

[den-crane]

19.8.3.8

I have a cluster configured as 1 shard with 6 replicas.

<replicated>
    <shard>
        <internal_replication>true</internal_replication>
        <replica><host>host1</host><port>9000</port></replica>
        <replica><host>host2</host><port>9000</port></replica>
        <replica><host>host3</host><port>9000</port></replica>
        <replica><host>host4</host><port>9000</port></replica>
        <replica><host>host5</host><port>9000</port></replica>
        <replica><host>host6</host><port>9000</port></replica>
    </shard>
</replicated>

At each node

create table max_p_test(a Int64) Engine=MergeTree order by a;
insert into max_p_test values(1);

At the first node

CREATE TABLE max_p_test_d as max_p_test
ENGINE = Distributed (replicated,currentDatabase(),'max_p_test');

SELECT *
FROM max_p_test_d

┌─a─┐
│ 1 │
└───┘

SET max_parallel_replicas = 10, prefer_localhost_replica = 0

SELECT *
FROM max_p_test_d

┌─a─┐
│ 1 │
└───┘
┌─a─┐
│ 1 │
└───┘
┌─a─┐
│ 1 │
└───┘
┌─a─┐
│ 1 │
└───┘
┌─a─┐
│ 1 │
└───┘
┌─a─┐
│ 1 │
└───┘

6 rows in set. Elapsed: 0.009 sec.

[alexey-milovidov]

It should not work without sampling at all.

[den-crane]

the same behavior 19.13.7.57