Interestingly, the GROUP BY splits work with (deprecated) Druid Datasources (e.g. this chart created using pageviews_daily )

cc-ing @JAllemandou, who said wanted to look at it, we'll triage with him Monday when he's back from vacation

TL;DR: This problem comes from how queries are translated from SQL to druid-query-plan. I don't have a solution for this :(

The query that times-out is (available using 'view query' in the chart top-right menu):

SELECT FLOOR("__time" TO MONTH) AS "__timestamp",
       "referer_class" AS "referer_class",
       SUM(view_count) AS "sum_view_count"
FROM "druid"."pageviews_daily"
WHERE "__time" >= '2020-06-01 00:00:00.000000'
  AND "__time" < '2021-06-30 00:00:00.000000'
  AND "agent_type" != 'spider'
GROUP BY "referer_class",
         FLOOR("__time" TO MONTH)
LIMIT 50000;

In superset SQL-lab I asked druid to shopw me its query-plan using the following syntax: EXPLAIN PLAN FOR <QUERY>, and got the following result for the query at stake:

DruidQueryRel(query=[{"queryType":"groupBy","dataSource":{"type":"table","name":"pageviews_daily"},"....

The query that doesn't time-out is:

{
  "queryType": "topN",
  "dataSource": "pageviews_daily",
 .....

The TopN queries in druid are a lot faster than GroupBy, and this is why the older chart gives result while the new one doesn't.

Things to keep in mind: The TopN queries are approximative, so results can be a little off (they usually are fairly correct when looking at first rows of top).
It is possible to make druid-sql generate topN queries, but I have not managed to have it working using a time-aggregation (see https://druid.apache.org/docs/latest/querying/sql.html#query-types for details).

Also: One way to get results is to set the time-grain to the value: original value. This makes calcite use the topN query (single field in group-by instead of two). You'll get daily values instead of monthly but at least you'll have values :)

And the feature request: https://github.com/apache/druid/issues/11264