Fix test failures in duckdb 0.10.0

[RobinL]

The root cause of the error is that we've been relying on implicit typecasting which is no longer possible

From duckdb 0.10.0 forwards, if dob is a date or timestamp, you cannot do this:
LEVENSHTEIN("dob_l", "dob_r") <= 1

Which means that you have to correctly cast all columns.

Error was: Binder Error: No function matches the given name and argument types 'levenshtein(TIMESTAMP_NS, TIMESTAMP_NS)'. You might need to add explicit type casts.
	Candidate functions:
	levenshtein(VARCHAR, VARCHAR) -> BIGINT

Note you can also no longer do try_strptime(dob_l, dob_r) either if dob is a datetime

Error was: Binder Error: No function matches the given name and argument types 'try_strptime(TIMESTAMP_NS, STRING_LITERAL)'. You might need to add explicit type casts.
	Candidate functions:
	try_strptime(VARCHAR, VARCHAR) -> TIMESTAMP
	try_strptime(VARCHAR, VARCHAR[]) -> TIMESTAMP

My proposed 'solution' is to have the function only work in duckdb 0.10.0 if the input is a string (cast_strings_to_date: bool = True). But NOT to error out if the user sets cast_strings_to_date: bool = False, because that would break things in duckdb < 0.10.0.

I think this makes sense because many of the function arguments only make sense if the input is a string e.g.:

• invalid_dates_as_null cannot be used if the input is a date
• separate_1st_january

When users have problems, we need to tell them to:

1. cast the date to a string before brining it into Splink, OR
2. downgrade back to duckdb 0.9.2

example

import pandas as pd

import splink.duckdb.comparison_library as cl
import splink.duckdb.comparison_template_library as ctl
from splink.datasets import splink_datasets
from splink.duckdb.blocking_rule_library import block_on
from splink.duckdb.linker import DuckDBLinker

df = splink_datasets.fake_1000
# convert dob column to datetime
df["dob"] = pd.to_datetime(df["dob"], format="%Y-%m-%d")

# get example value to check it's a date


settings = {
    "link_type": "dedupe_only",
    "blocking_rules_to_generate_predictions": [
        block_on("first_name"),

    ],
    "comparisons": [
        cl.exact_match("first_name"),
        cl.exact_match("surname"),
        ctl.date_comparison("dob", levenshtein_thresholds=[1]),
        cl.exact_match("city"),
        cl.exact_match("email"),
    ],
}

linker = DuckDBLinker(df, settings)
linker.predict()

[RobinL]

test no longer needed with later versions of pandas which seems to deal with typing better

[ADBond]

Cool, makes sense.
Not sure how the lockfile fits with the one in #1998 but presume you know which is more up-to-date