import ibis
ibis.options.interactive = True
@ibis.udf.scalar.builtin
def mismatches(left: str, right: str) -> int:
...Reference built-in functions
Scalar functions
Functions that aren’t exposed in ibis directly can be accessed using the @ibis.udf.scalar.builtin decorator.
Ibis APIs may already exist for your function.
Builtin scalar UDFs are designed to be an escape hatch when Ibis doesn’t have a defined API for a built-in database function.
See the reference documentation for existing APIs.
DuckDB
Ibis doesn’t directly expose many of the DuckDB text similarity functions. Let’s expose the mismatches API.
The ... is a visual indicator that the function definition is unknown to Ibis.
Ibis does not do anything with the function body.
Ibis will not execute the function body or otherwise inspect it. Any code you write in the function body will be ignored.
We can now call this function on any ibis expression:
con = ibis.duckdb.connect()- 1
- Connect to an in-memory DuckDB database
expr = mismatches("duck", "luck")
con.execute(expr)1Like any other ibis expression you can inspect the SQL:
ibis.to_sql(expr, dialect="duckdb")- 1
-
The
dialectkeyword argument must be passed, because we constructed a literal expression which has no backend attached.
SELECT
MISMATCHES('duck', 'luck') AS "mismatches_0('duck', 'luck')"Similarly we can expose Duckdb’s jaro_winkler_similarity function. Let’s alias it to jw_sim to illustrate some more of the Ibis udf API:
@ibis.udf.scalar.builtin(name="jaro_winkler_similarity")
def jw_sim(a: str, b: str) -> float:
...Because built-in UDFs are ultimately Ibis expressions, they compose with the rest of the library:
pkgs = ibis.read_parquet(
"https://storage.googleapis.com/ibis-tutorial-data/pypi/packages.parquet"
)
pandas_ish = pkgs[jw_sim(pkgs.name, "pandas") >= 0.9]
pandas_ish┏━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━┳━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┓ ┃ name ┃ version ┃ requires_python ┃ yanked ┃ has_binary_wheel ┃ has_vulnerabilities ┃ first_uploaded_at ┃ last_uploaded_at ┃ recorded_at ┃ downloads ┃ scorecard_overall ┃ in_google_assured_oss ┃ ┡━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━╇━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━┩ │ string │ string │ string │ boolean │ boolean │ boolean │ timestamp(6) │ timestamp(6) │ timestamp(6) │ int32 │ float64 │ boolean │ ├──────────┼───────────────────┼─────────────────┼─────────┼──────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼─────────────────────┼───────────┼───────────────────┼───────────────────────┤ │ bcpandas │ 2.4.1 │ >=3.8.1 │ True │ False │ False │ 2023-07-12 06:14:22 │ 2023-07-12 06:14:23 │ 2023-07-12 14:31:41 │ 0 │ NULL │ False │ │ espandas │ 1.0.4 │ ~ │ False │ False │ False │ 2018-12-22 20:52:30 │ 2018-12-22 20:52:30 │ 2023-07-12 14:58:47 │ 0 │ 3.6 │ False │ │ fpandas │ 0.5 │ ~ │ False │ False │ False │ 2020-03-09 02:35:31 │ 2020-03-09 02:35:31 │ 2023-07-12 15:04:23 │ 0 │ NULL │ False │ │ h3pandas │ 0.2.4 │ >=3.6 │ False │ False │ False │ 2023-03-19 17:58:16 │ 2023-03-19 17:58:16 │ 2023-07-12 15:10:06 │ 0 │ NULL │ False │ │ ipandas │ 0.0.1 │ ~ │ False │ False │ False │ 2019-05-29 18:46:12 │ 2019-05-29 18:46:12 │ 2023-07-12 15:15:34 │ 0 │ 3.6 │ False │ │ kpandas │ 0.0.1 │ >=3.6,<4.0 │ False │ False │ False │ 2019-05-02 18:00:29 │ 2019-05-02 18:00:31 │ 2023-07-12 15:20:21 │ 0 │ NULL │ False │ │ mpandas │ 0.0.2.1 │ ~ │ False │ False │ False │ 2022-07-03 16:21:21 │ 2022-07-03 16:21:23 │ 2023-07-12 15:30:35 │ 0 │ NULL │ False │ │ mtpandas │ 1.14.202306141807 │ >=3.6 │ False │ False │ False │ 2023-06-14 18:08:01 │ 2023-06-14 18:08:01 │ 2023-07-12 15:31:04 │ 0 │ 4.6 │ False │ │ mypandas │ 0.1.6 │ >=3.10 │ False │ False │ False │ 2022-10-24 21:01:10 │ 2022-10-24 21:01:12 │ 2023-07-12 15:32:04 │ 0 │ NULL │ False │ │ paandas │ 0.0.3 │ ~ │ False │ False │ False │ 2022-11-24 06:11:15 │ 2022-11-24 06:11:17 │ 2023-07-12 15:43:31 │ 0 │ NULL │ False │ │ … │ … │ … │ … │ … │ … │ … │ … │ … │ … │ … │ … │ └──────────┴───────────────────┴─────────────────┴─────────┴──────────────────┴─────────────────────┴─────────────────────┴─────────────────────┴─────────────────────┴───────────┴───────────────────┴───────────────────────┘
Defining Signatures
Sometimes the signatures of builtin functions are difficult to spell.
Consider a function that computes the length of any array: the elements in the array can be floats, integers, strings and even other arrays. Spelling that type is difficult.
Fortunately, the udf.scalar.builtin decorator only requires you to specify the type of the return value. The type of the function parameters are not required. Thus, this is adequate:
@ibis.udf.scalar.builtin(name="array_length")
def cardinality(arr) -> int:
...We can pass arrays with different element types to our cardinality function:
con.execute(cardinality([1, 2, 3]))3con.execute(cardinality(["a", "b"]))2When you do not specify input types, Ibis isn’t able to catch typing errors early, and they are only caught during execution. The errors you get back are backend dependent:
con.execute(cardinality("foo"))BinderException: Binder Error: No function matches the given name and argument types 'array_length(STRING_LITERAL)'. You might need to add explicit type casts.
Candidate functions:
array_length(ANY[]) -> BIGINT
array_length(ANY[], BIGINT) -> BIGINTHere, DuckDB is informing us that the ARRAY_LENGTH function does not accept strings as input.
Aggregate functions
Aggregate functions that aren’t exposed in ibis directly can be accessed using the @ibis.udf.agg.builtin decorator.
Ibis APIs may already exist for your function.
Builtin aggregate UDFs are designed to be an escape hatch when Ibis doesn’t have a defined API for a built-in database function.
See the reference documentation for existing APIs.
Let’s the use the DuckDB backend to demonstrate how to access an aggregate function that isn’t exposed in ibis: kurtosis.
DuckDB
First, define the builtin aggregate function:
@ibis.udf.agg.builtin
def kurtosis(x: float) -> float:
...- 1
- Both the input and return type annotations indicate the element type of the input, not the shape (column or scalar). Aggregations can only be called on column expressions.
One of the powerful features of this API is that you can define your UD(A)Fs at any point during your analysis. You don’t need to connect to the database to define your functions.
Let’s compute the kurtosis of the number of votes across all movies:
from ibis import _
expr = (
ibis.examples.imdb_title_ratings.fetch()
.rename("snake_case")
.agg(kurt=lambda t: kurtosis(t.num_votes))
)
expr┏━━━━━━━━━━━━┓ ┃ kurt ┃ ┡━━━━━━━━━━━━┩ │ float64 │ ├────────────┤ │ 4764.64335 │ └────────────┘
Since this is an aggregate function, it has the same capabilities as other, builtin aggregates like sum: it can be used in a group by as well as in a window function expression.
Let’s compute kurtosis for all the different types of productions (shorts, movies, TV, etc):
basics = (
ibis.examples.imdb_title_basics.fetch()
.rename("snake_case")
.filter(_.is_adult == 0)
)
ratings = ibis.examples.imdb_title_ratings.fetch().rename("snake_case")
basics_ratings = ratings.join(basics, "tconst")
expr = (
basics_ratings.group_by("title_type")
.agg(kurt=lambda t: kurtosis(t.num_votes))
.order_by(_.kurt.desc())
.head()
)
expr┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┓ ┃ title_type ┃ kurt ┃ ┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━━┩ │ string │ float64 │ ├──────────────┼─────────────┤ │ tvEpisode │ 7452.859695 │ │ tvSeries │ 4133.328876 │ │ short │ 3992.123752 │ │ tvMiniSeries │ 1968.324239 │ │ tvSpecial │ 1461.166556 │ └──────────────┴─────────────┘
Similarly for window functions:
expr = (
basics_ratings.mutate(
kurt=lambda t: kurtosis(t.num_votes).over(group_by="title_type")
)
.relocate("kurt", after="tconst")
.filter(
[
_.original_title.lower().contains("godfather"),
_.title_type == "movie",
_.genres.contains("Crime") & _.genres.contains("Drama"),
]
)
)
expr┏━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┓ ┃ tconst ┃ kurt ┃ average_rating ┃ num_votes ┃ title_type ┃ primary_title ┃ original_title ┃ is_adult ┃ start_year ┃ end_year ┃ runtime_minutes ┃ genres ┃ ┡━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━┩ │ string │ float64 │ float64 │ int64 │ string │ string │ string │ int64 │ int64 │ string │ int64 │ string │ ├────────────┼─────────────┼────────────────┼───────────┼────────────┼────────────────────────┼────────────────────────┼──────────┼────────────┼──────────┼─────────────────┼────────────────────┤ │ tt0458027 │ 1127.262502 │ 3.5 │ 26 │ movie │ Mumbai Godfather │ Mumbai Godfather │ 0 │ 2005 │ NULL │ 110 │ Action,Crime,Drama │ │ tt0068646 │ 1127.262502 │ 9.2 │ 2001286 │ movie │ The Godfather │ The Godfather │ 0 │ 1972 │ NULL │ 175 │ Crime,Drama │ │ tt0071562 │ 1127.262502 │ 9.0 │ 1357703 │ movie │ The Godfather Part II │ The Godfather Part II │ 0 │ 1974 │ NULL │ 202 │ Crime,Drama │ │ tt0074412 │ 1127.262502 │ 5.2 │ 1785 │ movie │ Disco Godfather │ Disco Godfather │ 0 │ 1979 │ NULL │ 98 │ Action,Crime,Drama │ │ tt0099674 │ 1127.262502 │ 7.6 │ 422466 │ movie │ The Godfather Part III │ The Godfather Part III │ 0 │ 1990 │ NULL │ 162 │ Crime,Drama │ │ tt0250404 │ 1127.262502 │ 6.5 │ 252 │ movie │ Godfather │ Godfather │ 0 │ 1992 │ NULL │ NULL │ Crime,Drama │ │ tt13130308 │ 1127.262502 │ 5.2 │ 7543 │ movie │ Godfather │ Godfather │ 0 │ 2022 │ NULL │ 157 │ Action,Crime,Drama │ └────────────┴─────────────┴────────────────┴───────────┴────────────┴────────────────────────┴────────────────────────┴──────────┴────────────┴──────────┴─────────────────┴────────────────────┘