Why does a union kill performance, when its fine to run the queries one-by-one?

TimescaleDB and PostgreSQL
1

I need to run a query that selects data from three tables. The “first” of the three is a timescaledb hypertable.

I have “organized” my query into three parts. If I run any of the three parts separately, the results are fetched immidately. However, if I run the three parts toghether, it takes “forever” (I let it run for 5 minutes before I cancelled it).

So, these queries run super-fast:

with "sensor" as (
	select 11958 as "id"
) (
	-- part 1
	select "sensorId",
		"timestamp",
		"value"
	from "SensorDataPoint",
		"sensor"
	where "aggregationMethodId" = 0
		and "sensorId" = "sensor"."id"
		and "timestamp" < '2024-03-01Z'
	order by "timestamp" desc
	limit 1
);

-- or

with "sensor" as (
	select 11958 as "id"
) 
(
	-- part 2
	select "sensorId",
		"timestamp",
		"sa"."value"::float8
	from "SensorEvent"
		join "StateAlias" "sa" on "sa"."id" = "stateAliasId",
		"sensor"
	where "sensorId" = "sensor"."id"
		and "timestamp" < '2024-03-01Z'
	order by "timestamp" desc
	limit 1
)
union all
(
	-- part 3
	select "sensorId",
		"timestamp",
		"value"
	from "ComputedSample",
		"sensor"
	where "sensorId" = "sensor"."id"
		and "timestamp" < '2024-03-01Z'
	order by "timestamp" desc
	limit 1
);

but, this query (i.e. a combination of the two queries above), takes forever:

with "sensor" as (
	select 11958 as "id"
) (
	-- part 1
	select "sensorId",
		"timestamp",
		"value"
	from "SensorDataPoint",
		"sensor"
	where "aggregationMethodId" = 0
		and "sensorId" = "sensor"."id"
		and "timestamp" < '2024-03-01Z'
	order by "timestamp" desc
	limit 1
)
union all
(
	-- part 2
	select "sensorId",
		"timestamp",
		"sa"."value"::float8
	from "SensorEvent"
		join "StateAlias" "sa" on "sa"."id" = "stateAliasId",
		"sensor"
	where "sensorId" = "sensor"."id"
		and "timestamp" < '2024-03-01Z'
	order by "timestamp" desc
	limit 1
)
union all
(
	-- part 3
	select "sensorId",
		"timestamp",
		"value"
	from "ComputedSample",
		"sensor"
	where "sensorId" = "sensor"."id"
		and "timestamp" < '2024-03-01Z'
	order by "timestamp" desc
	limit 1
)

Why?

How can I run a query that combines the results of these three queries, and still have the performance as if they were run separately?