Query very slow when using psycopg2 with TimescaleDB

Question

I am using psygopg2 to connect to a TimescaleDB instance. I want to query the latest entries from a big table (35 million rows) containing price information of assets with the columns datetime, asset_id, price, created_at:

create table prices
(
    datetime   timestamptz not null,
    asset_id   integer references assets (asset_id),
    price      real        not null,
    created_at timestamptz not null default now(),

    unique (datetime, asset_id)
);

This table was turned into a TimescaleDB hypertable with

select create_hypertable('"prices"', 'datetime');

and from this article's suggestion created the following index:

create index latest_prices_idx on prices (asset_id, datetime desc);

I use "Option 3" from this article to query the most recent price of each asset:

select distinct on (asset_id) datetime, asset_id, price
from prices
order by asset_id, datetime desc;

which takes ~300 ms via my IDE's console. However, when doing the same query via psycopg2 it takes about 10 seconds:

cursor.execute("""
    select distinct on (asset_id) datetime, asset_id, price
    from prices
    order by asset_id, datetime desc;
""")
res = cursor.fetchall()

Why might this be the case?

Prepending explain in front of the query outputs the exact same result for both the IDE and psycopg.

Edit: Answers to questions/suggestions in the comments:

• The database is running in a local docker container.
• The IDE is PyCharm, I do the queries from its database console and from a Python script within the IDE.
• Using a server side cursor (like this) did not change anything.
• PyCharm gives the query time in its console output. For getting the time of the Python code, I measure the execution time of the cursor.execute(...) call with time.perf_counter().
• Using psql in the shell in the container has the bad performance as well (~10 s).
• When dropping the index the performance flips: psql and psycopg take about 6 seconds, PyCharm's console about 13 seconds. The original behavior is reproducible when re-creating the index.