polars_core/chunked_array/
cast.rs

1//! Implementations of the ChunkCast Trait.
2
3use std::borrow::Cow;
4
5use polars_compute::cast::CastOptionsImpl;
6#[cfg(feature = "serde-lazy")]
7use serde::{Deserialize, Serialize};
8
9use super::flags::StatisticsFlags;
10#[cfg(feature = "dtype-datetime")]
11use crate::prelude::DataType::Datetime;
12use crate::prelude::*;
13use crate::utils::handle_casting_failures;
14
15#[derive(Copy, Clone, Debug, Default, PartialEq, Hash, Eq)]
16#[cfg_attr(feature = "serde-lazy", derive(Serialize, Deserialize))]
17#[cfg_attr(feature = "dsl-schema", derive(schemars::JsonSchema))]
18#[repr(u8)]
19pub enum CastOptions {
20    /// Raises on overflow
21    #[default]
22    Strict,
23    /// Overflow is replaced with null
24    NonStrict,
25    /// Allows wrapping overflow
26    Overflowing,
27}
28
29impl CastOptions {
30    pub fn is_strict(&self) -> bool {
31        matches!(self, CastOptions::Strict)
32    }
33}
34
35impl From<CastOptions> for CastOptionsImpl {
36    fn from(value: CastOptions) -> Self {
37        let wrapped = match value {
38            CastOptions::Strict | CastOptions::NonStrict => false,
39            CastOptions::Overflowing => true,
40        };
41        CastOptionsImpl {
42            wrapped,
43            partial: false,
44        }
45    }
46}
47
48pub(crate) fn cast_chunks(
49    chunks: &[ArrayRef],
50    dtype: &DataType,
51    options: CastOptions,
52) -> PolarsResult<Vec<ArrayRef>> {
53    let check_nulls = matches!(options, CastOptions::Strict);
54    let options = options.into();
55
56    let arrow_dtype = dtype.try_to_arrow(CompatLevel::newest())?;
57    chunks
58        .iter()
59        .map(|arr| {
60            let out = polars_compute::cast::cast(arr.as_ref(), &arrow_dtype, options);
61            if check_nulls {
62                out.and_then(|new| {
63                    polars_ensure!(arr.null_count() == new.null_count(), ComputeError: "strict cast failed");
64                    Ok(new)
65                })
66
67            } else {
68                out
69            }
70        })
71        .collect::<PolarsResult<Vec<_>>>()
72}
73
74fn cast_impl_inner(
75    name: PlSmallStr,
76    chunks: &[ArrayRef],
77    dtype: &DataType,
78    options: CastOptions,
79) -> PolarsResult<Series> {
80    let chunks = match dtype {
81        #[cfg(feature = "dtype-decimal")]
82        DataType::Decimal(_, _) => {
83            let mut chunks = cast_chunks(chunks, dtype, options)?;
84            // @NOTE: We cannot cast here as that will lower the scale.
85            for chunk in chunks.iter_mut() {
86                *chunk = std::mem::take(
87                    chunk
88                        .as_any_mut()
89                        .downcast_mut::<PrimitiveArray<i128>>()
90                        .unwrap(),
91                )
92                .to(ArrowDataType::Int128)
93                .to_boxed();
94            }
95            chunks
96        },
97        _ => cast_chunks(chunks, &dtype.to_physical(), options)?,
98    };
99
100    let out = Series::try_from((name, chunks))?;
101    use DataType::*;
102    let out = match dtype {
103        Date => out.into_date(),
104        Datetime(tu, tz) => match tz {
105            #[cfg(feature = "timezones")]
106            Some(tz) => {
107                TimeZone::validate_time_zone(tz)?;
108                out.into_datetime(*tu, Some(tz.clone()))
109            },
110            _ => out.into_datetime(*tu, None),
111        },
112        Duration(tu) => out.into_duration(*tu),
113        #[cfg(feature = "dtype-time")]
114        Time => out.into_time(),
115        #[cfg(feature = "dtype-decimal")]
116        Decimal(precision, scale) => out.into_decimal(*precision, scale.unwrap_or(0))?,
117        _ => out,
118    };
119
120    Ok(out)
121}
122
123fn cast_impl(
124    name: PlSmallStr,
125    chunks: &[ArrayRef],
126    dtype: &DataType,
127    options: CastOptions,
128) -> PolarsResult<Series> {
129    cast_impl_inner(name, chunks, dtype, options)
130}
131
132#[cfg(feature = "dtype-struct")]
133fn cast_single_to_struct(
134    name: PlSmallStr,
135    chunks: &[ArrayRef],
136    fields: &[Field],
137    options: CastOptions,
138) -> PolarsResult<Series> {
139    polars_ensure!(fields.len() == 1, InvalidOperation: "must specify one field in the struct");
140    let mut new_fields = Vec::with_capacity(fields.len());
141    // cast to first field dtype
142    let mut fields = fields.iter();
143    let fld = fields.next().unwrap();
144    let s = cast_impl_inner(fld.name.clone(), chunks, &fld.dtype, options)?;
145    let length = s.len();
146    new_fields.push(s);
147
148    for fld in fields {
149        new_fields.push(Series::full_null(fld.name.clone(), length, &fld.dtype));
150    }
151
152    StructChunked::from_series(name, length, new_fields.iter()).map(|ca| ca.into_series())
153}
154
155impl<T> ChunkedArray<T>
156where
157    T: PolarsNumericType,
158{
159    fn cast_impl(&self, dtype: &DataType, options: CastOptions) -> PolarsResult<Series> {
160        if self.dtype() == dtype {
161            // SAFETY: chunks are correct dtype
162            let mut out = unsafe {
163                Series::from_chunks_and_dtype_unchecked(
164                    self.name().clone(),
165                    self.chunks.clone(),
166                    dtype,
167                )
168            };
169            out.set_sorted_flag(self.is_sorted_flag());
170            return Ok(out);
171        }
172        match dtype {
173            // LEGACY
174            // TODO @ cat-rework: remove after exposing to/from physical functions.
175            #[cfg(feature = "dtype-categorical")]
176            DataType::Categorical(cats, _mapping) => {
177                let s = self.cast_with_options(&cats.physical().dtype(), options)?;
178                with_match_categorical_physical_type!(cats.physical(), |$C| {
179                    // SAFETY: we are guarded by the type system.
180                    type PhysCa = ChunkedArray<<$C as PolarsCategoricalType>::PolarsPhysical>;
181                    let ca: &PhysCa = s.as_ref().as_ref();
182                    Ok(CategoricalChunked::<$C>::from_cats_and_dtype(ca.clone(), dtype.clone())
183                        .into_series())
184                })
185            },
186
187            // LEGACY
188            // TODO @ cat-rework: remove after exposing to/from physical functions.
189            #[cfg(feature = "dtype-categorical")]
190            DataType::Enum(fcats, _mapping) => {
191                let s = self.cast_with_options(&fcats.physical().dtype(), options)?;
192                with_match_categorical_physical_type!(fcats.physical(), |$C| {
193                    // SAFETY: we are guarded by the type system.
194                    type PhysCa = ChunkedArray<<$C as PolarsCategoricalType>::PolarsPhysical>;
195                    let ca: &PhysCa = s.as_ref().as_ref();
196                    Ok(CategoricalChunked::<$C>::from_cats_and_dtype(ca.clone(), dtype.clone()).into_series())
197                })
198            },
199
200            #[cfg(feature = "dtype-struct")]
201            DataType::Struct(fields) => {
202                cast_single_to_struct(self.name().clone(), &self.chunks, fields, options)
203            },
204            _ => cast_impl_inner(self.name().clone(), &self.chunks, dtype, options).map(|mut s| {
205                // maintain sorted if data types
206                // - remain signed
207                // - unsigned -> signed
208                // this may still fail with overflow?
209                let to_signed = dtype.is_signed_integer();
210                let unsigned2unsigned =
211                    self.dtype().is_unsigned_integer() && dtype.is_unsigned_integer();
212                let allowed = to_signed || unsigned2unsigned;
213
214                if (allowed)
215                    && (s.null_count() == self.null_count())
216                    // physical to logicals
217                    || (self.dtype().to_physical() == dtype.to_physical())
218                {
219                    let is_sorted = self.is_sorted_flag();
220                    s.set_sorted_flag(is_sorted)
221                }
222                s
223            }),
224        }
225    }
226}
227
228impl<T> ChunkCast for ChunkedArray<T>
229where
230    T: PolarsNumericType,
231{
232    fn cast_with_options(&self, dtype: &DataType, options: CastOptions) -> PolarsResult<Series> {
233        self.cast_impl(dtype, options)
234    }
235
236    unsafe fn cast_unchecked(&self, dtype: &DataType) -> PolarsResult<Series> {
237        match dtype {
238            // LEGACY
239            // TODO @ cat-rework: remove after exposing to/from physical functions.
240            #[cfg(feature = "dtype-categorical")]
241            DataType::Categorical(cats, _mapping) => {
242                polars_ensure!(self.dtype() == &cats.physical().dtype(), ComputeError: "cannot cast numeric types to 'Categorical'");
243                with_match_categorical_physical_type!(cats.physical(), |$C| {
244                    // SAFETY: we are guarded by the type system.
245                    type PhysCa = ChunkedArray<<$C as PolarsCategoricalType>::PolarsPhysical>;
246                    let ca = unsafe { &*(self as *const ChunkedArray<T> as *const PhysCa) };
247                    Ok(CategoricalChunked::<$C>::from_cats_and_dtype_unchecked(ca.clone(), dtype.clone())
248                        .into_series())
249                })
250            },
251
252            // LEGACY
253            // TODO @ cat-rework: remove after exposing to/from physical functions.
254            #[cfg(feature = "dtype-categorical")]
255            DataType::Enum(fcats, _mapping) => {
256                polars_ensure!(self.dtype() == &fcats.physical().dtype(), ComputeError: "cannot cast numeric types to 'Enum'");
257                with_match_categorical_physical_type!(fcats.physical(), |$C| {
258                    // SAFETY: we are guarded by the type system.
259                    type PhysCa = ChunkedArray<<$C as PolarsCategoricalType>::PolarsPhysical>;
260                    let ca = unsafe { &*(self as *const ChunkedArray<T> as *const PhysCa) };
261                    Ok(CategoricalChunked::<$C>::from_cats_and_dtype_unchecked(ca.clone(), dtype.clone()).into_series())
262                })
263            },
264
265            _ => self.cast_impl(dtype, CastOptions::Overflowing),
266        }
267    }
268}
269
270impl ChunkCast for StringChunked {
271    fn cast_with_options(&self, dtype: &DataType, options: CastOptions) -> PolarsResult<Series> {
272        match dtype {
273            #[cfg(feature = "dtype-categorical")]
274            DataType::Categorical(cats, _mapping) => {
275                with_match_categorical_physical_type!(cats.physical(), |$C| {
276                    Ok(CategoricalChunked::<$C>::from_str_iter(self.name().clone(), dtype.clone(), self.iter())?
277                        .into_series())
278                })
279            },
280            #[cfg(feature = "dtype-categorical")]
281            DataType::Enum(fcats, _mapping) => {
282                let ret = with_match_categorical_physical_type!(fcats.physical(), |$C| {
283                    CategoricalChunked::<$C>::from_str_iter(self.name().clone(), dtype.clone(), self.iter())?
284                        .into_series()
285                });
286
287                if options.is_strict() && self.null_count() != ret.null_count() {
288                    handle_casting_failures(&self.clone().into_series(), &ret)?;
289                }
290
291                Ok(ret)
292            },
293            #[cfg(feature = "dtype-struct")]
294            DataType::Struct(fields) => {
295                cast_single_to_struct(self.name().clone(), &self.chunks, fields, options)
296            },
297            #[cfg(feature = "dtype-decimal")]
298            DataType::Decimal(precision, scale) => match (precision, scale) {
299                (precision, Some(scale)) => {
300                    let chunks = self.downcast_iter().map(|arr| {
301                        polars_compute::cast::binview_to_decimal(
302                            &arr.to_binview(),
303                            *precision,
304                            *scale,
305                        )
306                        .to(ArrowDataType::Int128)
307                    });
308                    Ok(Int128Chunked::from_chunk_iter(self.name().clone(), chunks)
309                        .into_decimal_unchecked(*precision, *scale)
310                        .into_series())
311                },
312                (None, None) => self.to_decimal_infer(100),
313                _ => {
314                    polars_bail!(ComputeError: "expected 'precision' or 'scale' when casting to Decimal")
315                },
316            },
317            #[cfg(feature = "dtype-date")]
318            DataType::Date => {
319                let result = cast_chunks(&self.chunks, dtype, options)?;
320                let out = Series::try_from((self.name().clone(), result))?;
321                Ok(out)
322            },
323            #[cfg(feature = "dtype-datetime")]
324            DataType::Datetime(time_unit, time_zone) => match time_zone {
325                #[cfg(feature = "timezones")]
326                Some(time_zone) => {
327                    TimeZone::validate_time_zone(time_zone)?;
328                    let result = cast_chunks(
329                        &self.chunks,
330                        &Datetime(time_unit.to_owned(), Some(time_zone.clone())),
331                        options,
332                    )?;
333                    Series::try_from((self.name().clone(), result))
334                },
335                _ => {
336                    let result =
337                        cast_chunks(&self.chunks, &Datetime(time_unit.to_owned(), None), options)?;
338                    Series::try_from((self.name().clone(), result))
339                },
340            },
341            _ => cast_impl(self.name().clone(), &self.chunks, dtype, options),
342        }
343    }
344
345    unsafe fn cast_unchecked(&self, dtype: &DataType) -> PolarsResult<Series> {
346        self.cast_with_options(dtype, CastOptions::Overflowing)
347    }
348}
349
350impl BinaryChunked {
351    /// # Safety
352    /// String is not validated
353    pub unsafe fn to_string_unchecked(&self) -> StringChunked {
354        let chunks = self
355            .downcast_iter()
356            .map(|arr| unsafe { arr.to_utf8view_unchecked() }.boxed())
357            .collect();
358        let field = Arc::new(Field::new(self.name().clone(), DataType::String));
359
360        let mut ca = StringChunked::new_with_compute_len(field, chunks);
361
362        use StatisticsFlags as F;
363        ca.retain_flags_from(self, F::IS_SORTED_ANY | F::CAN_FAST_EXPLODE_LIST);
364        ca
365    }
366}
367
368impl StringChunked {
369    pub fn as_binary(&self) -> BinaryChunked {
370        let chunks = self
371            .downcast_iter()
372            .map(|arr| arr.to_binview().boxed())
373            .collect();
374        let field = Arc::new(Field::new(self.name().clone(), DataType::Binary));
375
376        let mut ca = BinaryChunked::new_with_compute_len(field, chunks);
377
378        use StatisticsFlags as F;
379        ca.retain_flags_from(self, F::IS_SORTED_ANY | F::CAN_FAST_EXPLODE_LIST);
380        ca
381    }
382}
383
384impl ChunkCast for BinaryChunked {
385    fn cast_with_options(&self, dtype: &DataType, options: CastOptions) -> PolarsResult<Series> {
386        match dtype {
387            #[cfg(feature = "dtype-struct")]
388            DataType::Struct(fields) => {
389                cast_single_to_struct(self.name().clone(), &self.chunks, fields, options)
390            },
391            _ => cast_impl(self.name().clone(), &self.chunks, dtype, options),
392        }
393    }
394
395    unsafe fn cast_unchecked(&self, dtype: &DataType) -> PolarsResult<Series> {
396        match dtype {
397            DataType::String => unsafe { Ok(self.to_string_unchecked().into_series()) },
398            _ => self.cast_with_options(dtype, CastOptions::Overflowing),
399        }
400    }
401}
402
403impl ChunkCast for BinaryOffsetChunked {
404    fn cast_with_options(&self, dtype: &DataType, options: CastOptions) -> PolarsResult<Series> {
405        match dtype {
406            #[cfg(feature = "dtype-struct")]
407            DataType::Struct(fields) => {
408                cast_single_to_struct(self.name().clone(), &self.chunks, fields, options)
409            },
410            _ => cast_impl(self.name().clone(), &self.chunks, dtype, options),
411        }
412    }
413
414    unsafe fn cast_unchecked(&self, dtype: &DataType) -> PolarsResult<Series> {
415        self.cast_with_options(dtype, CastOptions::Overflowing)
416    }
417}
418
419impl ChunkCast for BooleanChunked {
420    fn cast_with_options(&self, dtype: &DataType, options: CastOptions) -> PolarsResult<Series> {
421        match dtype {
422            #[cfg(feature = "dtype-struct")]
423            DataType::Struct(fields) => {
424                cast_single_to_struct(self.name().clone(), &self.chunks, fields, options)
425            },
426            #[cfg(feature = "dtype-categorical")]
427            DataType::Categorical(_, _) | DataType::Enum(_, _) => {
428                polars_bail!(InvalidOperation: "cannot cast Boolean to Categorical");
429            },
430            _ => cast_impl(self.name().clone(), &self.chunks, dtype, options),
431        }
432    }
433
434    unsafe fn cast_unchecked(&self, dtype: &DataType) -> PolarsResult<Series> {
435        self.cast_with_options(dtype, CastOptions::Overflowing)
436    }
437}
438
439/// We cannot cast anything to or from List/LargeList
440/// So this implementation casts the inner type
441impl ChunkCast for ListChunked {
442    fn cast_with_options(&self, dtype: &DataType, options: CastOptions) -> PolarsResult<Series> {
443        let ca = self
444            .trim_lists_to_normalized_offsets()
445            .map_or(Cow::Borrowed(self), Cow::Owned);
446        let ca = ca.propagate_nulls().map_or(ca, Cow::Owned);
447
448        use DataType::*;
449        match dtype {
450            List(child_type) => {
451                match (ca.inner_dtype(), &**child_type) {
452                    (old, new) if old == new => Ok(ca.into_owned().into_series()),
453                    // TODO @ cat-rework: can we implement this now?
454                    #[cfg(feature = "dtype-categorical")]
455                    (dt, Categorical(_, _) | Enum(_, _))
456                        if !matches!(dt, Categorical(_, _) | Enum(_, _) | String | Null) =>
457                    {
458                        polars_bail!(InvalidOperation: "cannot cast List inner type: '{:?}' to Categorical", dt)
459                    },
460                    _ => {
461                        // ensure the inner logical type bubbles up
462                        let (arr, child_type) = cast_list(ca.as_ref(), child_type, options)?;
463                        // SAFETY: we just cast so the dtype matches.
464                        // we must take this path to correct for physical types.
465                        unsafe {
466                            Ok(Series::from_chunks_and_dtype_unchecked(
467                                ca.name().clone(),
468                                vec![arr],
469                                &List(Box::new(child_type)),
470                            ))
471                        }
472                    },
473                }
474            },
475            #[cfg(feature = "dtype-array")]
476            Array(child_type, width) => {
477                let physical_type = dtype.to_physical();
478
479                // TODO @ cat-rework: can we implement this now?
480                // TODO!: properly implement this recursively.
481                #[cfg(feature = "dtype-categorical")]
482                polars_ensure!(!matches!(&**child_type, Categorical(_, _)), InvalidOperation: "array of categorical is not yet supported");
483
484                // cast to the physical type to avoid logical chunks.
485                let chunks = cast_chunks(ca.chunks(), &physical_type, options)?;
486                // SAFETY: we just cast so the dtype matches.
487                // we must take this path to correct for physical types.
488                unsafe {
489                    Ok(Series::from_chunks_and_dtype_unchecked(
490                        ca.name().clone(),
491                        chunks,
492                        &Array(child_type.clone(), *width),
493                    ))
494                }
495            },
496            #[cfg(feature = "dtype-u8")]
497            Binary => {
498                polars_ensure!(
499                    matches!(self.inner_dtype(), UInt8),
500                    InvalidOperation: "cannot cast List type (inner: '{:?}', to: '{:?}')",
501                    self.inner_dtype(),
502                    dtype,
503                );
504                let chunks = cast_chunks(self.chunks(), &DataType::Binary, options)?;
505
506                // SAFETY: we just cast so the dtype matches.
507                unsafe {
508                    Ok(Series::from_chunks_and_dtype_unchecked(
509                        self.name().clone(),
510                        chunks,
511                        &DataType::Binary,
512                    ))
513                }
514            },
515            _ => {
516                polars_bail!(
517                    InvalidOperation: "cannot cast List type (inner: '{:?}', to: '{:?}')",
518                    ca.inner_dtype(),
519                    dtype,
520                )
521            },
522        }
523    }
524
525    unsafe fn cast_unchecked(&self, dtype: &DataType) -> PolarsResult<Series> {
526        use DataType::*;
527        match dtype {
528            List(child_type) => cast_list_unchecked(self, child_type),
529            _ => self.cast_with_options(dtype, CastOptions::Overflowing),
530        }
531    }
532}
533
534/// We cannot cast anything to or from List/LargeList
535/// So this implementation casts the inner type
536#[cfg(feature = "dtype-array")]
537impl ChunkCast for ArrayChunked {
538    fn cast_with_options(&self, dtype: &DataType, options: CastOptions) -> PolarsResult<Series> {
539        let ca = self
540            .trim_lists_to_normalized_offsets()
541            .map_or(Cow::Borrowed(self), Cow::Owned);
542        let ca = ca.propagate_nulls().map_or(ca, Cow::Owned);
543
544        use DataType::*;
545        match dtype {
546            Array(child_type, width) => {
547                polars_ensure!(
548                    *width == ca.width(),
549                    InvalidOperation: "cannot cast Array to a different width"
550                );
551
552                match (ca.inner_dtype(), &**child_type) {
553                    (old, new) if old == new => Ok(ca.into_owned().into_series()),
554                    // TODO @ cat-rework: can we implement this now?
555                    #[cfg(feature = "dtype-categorical")]
556                    (dt, Categorical(_, _) | Enum(_, _)) if !matches!(dt, String) => {
557                        polars_bail!(InvalidOperation: "cannot cast Array inner type: '{:?}' to dtype: {:?}", dt, child_type)
558                    },
559                    _ => {
560                        // ensure the inner logical type bubbles up
561                        let (arr, child_type) =
562                            cast_fixed_size_list(ca.as_ref(), child_type, options)?;
563                        // SAFETY: we just cast so the dtype matches.
564                        // we must take this path to correct for physical types.
565                        unsafe {
566                            Ok(Series::from_chunks_and_dtype_unchecked(
567                                ca.name().clone(),
568                                vec![arr],
569                                &Array(Box::new(child_type), *width),
570                            ))
571                        }
572                    },
573                }
574            },
575            List(child_type) => {
576                let physical_type = dtype.to_physical();
577                // cast to the physical type to avoid logical chunks.
578                let chunks = cast_chunks(ca.chunks(), &physical_type, options)?;
579                // SAFETY: we just cast so the dtype matches.
580                // we must take this path to correct for physical types.
581                unsafe {
582                    Ok(Series::from_chunks_and_dtype_unchecked(
583                        ca.name().clone(),
584                        chunks,
585                        &List(child_type.clone()),
586                    ))
587                }
588            },
589            _ => {
590                polars_bail!(
591                    InvalidOperation: "cannot cast Array type (inner: '{:?}', to: '{:?}')",
592                    ca.inner_dtype(),
593                    dtype,
594                )
595            },
596        }
597    }
598
599    unsafe fn cast_unchecked(&self, dtype: &DataType) -> PolarsResult<Series> {
600        self.cast_with_options(dtype, CastOptions::Overflowing)
601    }
602}
603
604// Returns inner data type. This is needed because a cast can instantiate the dtype inner
605// values for instance with categoricals
606fn cast_list(
607    ca: &ListChunked,
608    child_type: &DataType,
609    options: CastOptions,
610) -> PolarsResult<(ArrayRef, DataType)> {
611    // We still rechunk because we must bubble up a single data-type
612    // TODO!: consider a version that works on chunks and merges the data-types and arrays.
613    let ca = ca.rechunk();
614    let arr = ca.downcast_as_array();
615    // SAFETY: inner dtype is passed correctly
616    let s = unsafe {
617        Series::from_chunks_and_dtype_unchecked(
618            PlSmallStr::EMPTY,
619            vec![arr.values().clone()],
620            ca.inner_dtype(),
621        )
622    };
623    let new_inner = s.cast_with_options(child_type, options)?;
624
625    let inner_dtype = new_inner.dtype().clone();
626    debug_assert_eq!(&inner_dtype, child_type);
627
628    let new_values = new_inner.array_ref(0).clone();
629
630    let dtype = ListArray::<i64>::default_datatype(new_values.dtype().clone());
631    let new_arr = ListArray::<i64>::new(
632        dtype,
633        arr.offsets().clone(),
634        new_values,
635        arr.validity().cloned(),
636    );
637    Ok((new_arr.boxed(), inner_dtype))
638}
639
640unsafe fn cast_list_unchecked(ca: &ListChunked, child_type: &DataType) -> PolarsResult<Series> {
641    // TODO! add chunked, but this must correct for list offsets.
642    let ca = ca.rechunk();
643    let arr = ca.downcast_as_array();
644    // SAFETY: inner dtype is passed correctly
645    let s = unsafe {
646        Series::from_chunks_and_dtype_unchecked(
647            PlSmallStr::EMPTY,
648            vec![arr.values().clone()],
649            ca.inner_dtype(),
650        )
651    };
652    let new_inner = s.cast_unchecked(child_type)?;
653    let new_values = new_inner.array_ref(0).clone();
654
655    let dtype = ListArray::<i64>::default_datatype(new_values.dtype().clone());
656    let new_arr = ListArray::<i64>::new(
657        dtype,
658        arr.offsets().clone(),
659        new_values,
660        arr.validity().cloned(),
661    );
662    Ok(ListChunked::from_chunks_and_dtype_unchecked(
663        ca.name().clone(),
664        vec![Box::new(new_arr)],
665        DataType::List(Box::new(child_type.clone())),
666    )
667    .into_series())
668}
669
670// Returns inner data type. This is needed because a cast can instantiate the dtype inner
671// values for instance with categoricals
672#[cfg(feature = "dtype-array")]
673fn cast_fixed_size_list(
674    ca: &ArrayChunked,
675    child_type: &DataType,
676    options: CastOptions,
677) -> PolarsResult<(ArrayRef, DataType)> {
678    let ca = ca.rechunk();
679    let arr = ca.downcast_as_array();
680    // SAFETY: inner dtype is passed correctly
681    let s = unsafe {
682        Series::from_chunks_and_dtype_unchecked(
683            PlSmallStr::EMPTY,
684            vec![arr.values().clone()],
685            ca.inner_dtype(),
686        )
687    };
688    let new_inner = s.cast_with_options(child_type, options)?;
689
690    let inner_dtype = new_inner.dtype().clone();
691    debug_assert_eq!(&inner_dtype, child_type);
692
693    let new_values = new_inner.array_ref(0).clone();
694
695    let dtype = FixedSizeListArray::default_datatype(new_values.dtype().clone(), ca.width());
696    let new_arr = FixedSizeListArray::new(dtype, ca.len(), new_values, arr.validity().cloned());
697    Ok((Box::new(new_arr), inner_dtype))
698}
699
700#[cfg(test)]
701mod test {
702    use crate::chunked_array::cast::CastOptions;
703    use crate::prelude::*;
704
705    #[test]
706    fn test_cast_list() -> PolarsResult<()> {
707        let mut builder = ListPrimitiveChunkedBuilder::<Int32Type>::new(
708            PlSmallStr::from_static("a"),
709            10,
710            10,
711            DataType::Int32,
712        );
713        builder.append_opt_slice(Some(&[1i32, 2, 3]));
714        builder.append_opt_slice(Some(&[1i32, 2, 3]));
715        let ca = builder.finish();
716
717        let new = ca.cast_with_options(
718            &DataType::List(DataType::Float64.into()),
719            CastOptions::Strict,
720        )?;
721
722        assert_eq!(new.dtype(), &DataType::List(DataType::Float64.into()));
723        Ok(())
724    }
725
726    #[test]
727    #[cfg(feature = "dtype-categorical")]
728    fn test_cast_noop() {
729        // check if we can cast categorical twice without panic
730        let ca = StringChunked::new(PlSmallStr::from_static("foo"), &["bar", "ham"]);
731        let cats = Categories::global();
732        let out = ca
733            .cast_with_options(
734                &DataType::from_categories(cats.clone()),
735                CastOptions::Strict,
736            )
737            .unwrap();
738        let out = out.cast(&DataType::from_categories(cats)).unwrap();
739        assert!(matches!(out.dtype(), &DataType::Categorical(_, _)))
740    }
741}
polars_core/chunked_array/cast.rs

polars_core/chunked_array/
cast.rs
