feat: avoid double fetches on `ibis.duckdb.connect().read_csv("https://slow_url").cache()`

[NickCrews]

Is your feature request related to a problem?

When you call .read_csv() on the duckdb backend, this makes duckdb actually go fetch [some] of the data in order to sniff the schema. Then, when you call .cache() on the created view, it actually goes and fetches the full data.

This is related to https://github.com/ibis-project/ibis/issues/9931.

What is the motivation behind your request?

I am working on relatively large tables on a slow internet connection. Each fetch takes about 30 seconds. I would like to avoid this double fetch.

Describe the solution you'd like

Since the result of .read_csv() needs to be a Table with a known schema, it is going to be required to fetch some data during that function call. So, I think we need to add an optional argument to the function, or create entirely new function. I would vote for adding params if we can come up with something sane. Maybe cache: bool?

What version of ibis are you running?

main

What backend(s) are you using, if any?

duckdb

Code of Conduct

• [x] I agree to follow this project's Code of Conduct

[cpcloud]

Can you break down the timing of:

CREATE VIEW v AS FROM read_csv('slow_url')

and

CREATE TABLE t AS FROM read_csv('slow_url')

?

It's not clear that there's a "double fetch" (as in they two fetches are equivalently slow) so much as a small fetch followed by a large fetch.

[NickCrews]

Yeah good idea, I'll do that and report back.

[NickCrews]

By running this several times:

import time

import duckdb

url_388mb = "https://github.com/NickCrews/apoc-data/releases/download/20250514-111142/income.csv"
url_4mb = "https://github.com/NickCrews/apoc-data/releases/download/20250514-111142/campaign_form.csv"


def timeit(sql):
    con = duckdb.connect()
    con.execute("SET enable_progress_bar = false;")
    start = time.time()
    con.sql(sql)
    end = time.time()
    print(f"Execution time: {end - start:.4f} seconds")


def view_then_table(url):
    return (
        f"CREATE VIEW v AS FROM read_csv('{url}'); CREATE TABLE t AS SELECT * FROM v;"
    )


def table_immediate(url):
    return f"CREATE TABLE t AS FROM read_csv('{url}');"


timeit(view_then_table(url_388mb))
timeit(table_immediate(url_388mb))
timeit(view_then_table(url_4mb))
timeit(table_immediate(url_4mb))

Here are a few results, on a 460Mbps download connection, per https://speed.measurementlab.net/#/

Execution time: 14.4625 seconds
Execution time: 13.3351 seconds
Execution time: 2.2171 seconds
Execution time: 1.0658 seconds

Execution time: 14.3710 seconds
Execution time: 12.8988 seconds
Execution time: 2.9992 seconds
Execution time: 1.0783 seconds

Execution time: 14.4168 seconds
Execution time: 12.9135 seconds
Execution time: 2.2149 seconds
Execution time: 1.0874 seconds

It appears that duckdb takes about 1-2 seconds to do the initial fetch to get enough rows to infer the schema of the table. For "large" (~400Mb) files, this isn't that big of a deal. But for "small" (4Mb) files, it can double the time taken.

[cpcloud]

I'm sympathetic to problem here, but can we perhaps try to get this performance difference fixed upstream in DuckDB so that we don't have to alter the Ibis API?

[NickCrews]

I believe this is solved in duckdb 1.3.0 with the external file cache