Weird result in corner case with external aggregation

[CurtizJ]

Consider the following sql script:

drop table if exists data;
drop table if exists dist;

set max_bytes_before_external_group_by = 1;
set group_by_two_level_threshold_bytes = 1;

set max_untracked_memory = '1Mi';
set memory_profiler_step = '1Mi';

create table data (key String) Engine=Memory;
create table dist (key LowCardinality(String)) engine=Distributed(test_cluster_two_shards, currentDatabase(), data);
insert into data values ('foo');

select * from dist group by key;

The last query returns (non-deterministically) the following result which is wrong (key is duplicated):

┌─key─┐
│ foo │
└─────┘
┌─key─┐
│ foo │
└─────┘

In logs we can see that identical keys go to different buckets:

2023.12.13 14:18:23.348050 [ 2207426 ] {8c005767-91d3-44ae-8d67-c5b4ba610f79} <Trace> Aggregator: Merging partially aggregated blocks (bucket = 174).
2023.12.13 14:18:23.348077 [ 2207426 ] {8c005767-91d3-44ae-8d67-c5b4ba610f79} <Debug> Aggregator: Merged partially aggregated blocks for bucket #174. Got 1 rows, 1.00 B from 1 source rows in 1.7239e-05 sec. (58008.005 rows/sec., 56.65 KiB/sec.)

2023.12.13 14:18:23.348123 [ 2207477 ] {8c005767-91d3-44ae-8d67-c5b4ba610f79} <Trace> Aggregator: Merging partially aggregated blocks (bucket = 222).
2023.12.13 14:18:23.348153 [ 2207477 ] {8c005767-91d3-44ae-8d67-c5b4ba610f79} <Debug> Aggregator: Merged partially aggregated blocks for bucket #222. Got 1 rows, 1.00 B from 1 source rows in 1.9668e-05 sec. (50844.011 rows/sec., 49.65 KiB/sec.)

Unfortunately I couldn't reproduce it on more real example. Also note that if LowCardinality(String)) is changed to String the result is correct. So probably there is some issue with conversion. However there are the same aggregation methods selected on both shards: Aggregation method: key_string.

[nickitat]

difference in hashing methods can cause this: #42630