arrow_array/array/byte_view_array.rs
1// Licensed to the Apache Software Foundation (ASF) under one
2// or more contributor license agreements. See the NOTICE file
3// distributed with this work for additional information
4// regarding copyright ownership. The ASF licenses this file
5// to you under the Apache License, Version 2.0 (the
6// "License"); you may not use this file except in compliance
7// with the License. You may obtain a copy of the License at
8//
9// http://www.apache.org/licenses/LICENSE-2.0
10//
11// Unless required by applicable law or agreed to in writing,
12// software distributed under the License is distributed on an
13// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14// KIND, either express or implied. See the License for the
15// specific language governing permissions and limitations
16// under the License.
17
18use crate::array::print_long_array;
19use crate::builder::{ArrayBuilder, GenericByteViewBuilder};
20use crate::iterator::ArrayIter;
21use crate::types::bytes::ByteArrayNativeType;
22use crate::types::{BinaryViewType, ByteViewType, StringViewType};
23use crate::{Array, ArrayAccessor, ArrayRef, GenericByteArray, OffsetSizeTrait, Scalar};
24use arrow_buffer::{ArrowNativeType, Buffer, NullBuffer, ScalarBuffer};
25use arrow_data::{ArrayData, ArrayDataBuilder, ByteView, MAX_INLINE_VIEW_LEN};
26use arrow_schema::{ArrowError, DataType};
27use core::str;
28use num::ToPrimitive;
29use std::any::Any;
30use std::cmp::Ordering;
31use std::fmt::Debug;
32use std::marker::PhantomData;
33use std::sync::Arc;
34
35use super::ByteArrayType;
36
37/// [Variable-size Binary View Layout]: An array of variable length bytes views.
38///
39/// This array type is used to store variable length byte data (e.g. Strings, Binary)
40/// and has efficient operations such as `take`, `filter`, and comparison.
41///
42/// [Variable-size Binary View Layout]: https://arrow.apache.org/docs/format/Columnar.html#variable-size-binary-view-layout
43///
44/// This is different from [`GenericByteArray`], which also stores variable
45/// length byte data, as it represents strings with an offset and length. `take`
46/// and `filter` like operations are implemented by manipulating the "views"
47/// (`u128`) without modifying the bytes. Each view also stores an inlined
48/// prefix which speed up comparisons.
49///
50/// # See Also
51///
52/// * [`StringViewArray`] for storing utf8 encoded string data
53/// * [`BinaryViewArray`] for storing bytes
54/// * [`ByteView`] to interpret `u128`s layout of the views.
55///
56/// [`ByteView`]: arrow_data::ByteView
57///
58/// # Layout: "views" and buffers
59///
60/// A `GenericByteViewArray` stores variable length byte strings. An array of
61/// `N` elements is stored as `N` fixed length "views" and a variable number
62/// of variable length "buffers".
63///
64/// Each view is a `u128` value whose layout is different depending on the
65/// length of the string stored at that location:
66///
67/// ```text
68/// ┌──────┬────────────────────────┐
69/// │length│ string value │
70/// Strings (len <= 12) │ │ (padded with 0) │
71/// └──────┴────────────────────────┘
72/// 0 31 127
73///
74/// ┌───────┬───────┬───────┬───────┐
75/// │length │prefix │ buf │offset │
76/// Strings (len > 12) │ │ │ index │ │
77/// └───────┴───────┴───────┴───────┘
78/// 0 31 63 95 127
79/// ```
80///
81/// * Strings with length <= 12 ([`MAX_INLINE_VIEW_LEN`]) are stored directly in
82/// the view. See [`Self::inline_value`] to access the inlined prefix from a
83/// short view.
84///
85/// * Strings with length > 12: The first four bytes are stored inline in the
86/// view and the entire string is stored in one of the buffers. See [`ByteView`]
87/// to access the fields of the these views.
88///
89/// As with other arrays, the optimized kernels in [`arrow_compute`] are likely
90/// the easiest and fastest way to work with this data. However, it is possible
91/// to access the views and buffers directly for more control.
92///
93/// For example
94///
95/// ```rust
96/// # use arrow_array::StringViewArray;
97/// # use arrow_array::Array;
98/// use arrow_data::ByteView;
99/// let array = StringViewArray::from(vec![
100/// "hello",
101/// "this string is longer than 12 bytes",
102/// "this string is also longer than 12 bytes"
103/// ]);
104///
105/// // ** Examine the first view (short string) **
106/// assert!(array.is_valid(0)); // Check for nulls
107/// let short_view: u128 = array.views()[0]; // "hello"
108/// // get length of the string
109/// let len = short_view as u32;
110/// assert_eq!(len, 5); // strings less than 12 bytes are stored in the view
111/// // SAFETY: `view` is a valid view
112/// let value = unsafe {
113/// StringViewArray::inline_value(&short_view, len as usize)
114/// };
115/// assert_eq!(value, b"hello");
116///
117/// // ** Examine the third view (long string) **
118/// assert!(array.is_valid(12)); // Check for nulls
119/// let long_view: u128 = array.views()[2]; // "this string is also longer than 12 bytes"
120/// let len = long_view as u32;
121/// assert_eq!(len, 40); // strings longer than 12 bytes are stored in the buffer
122/// let view = ByteView::from(long_view); // use ByteView to access the fields
123/// assert_eq!(view.length, 40);
124/// assert_eq!(view.buffer_index, 0);
125/// assert_eq!(view.offset, 35); // data starts after the first long string
126/// // Views for long strings store a 4 byte prefix
127/// let prefix = view.prefix.to_le_bytes();
128/// assert_eq!(&prefix, b"this");
129/// let value = array.value(2); // get the string value (see `value` implementation for how to access the bytes directly)
130/// assert_eq!(value, "this string is also longer than 12 bytes");
131/// ```
132///
133/// [`MAX_INLINE_VIEW_LEN`]: arrow_data::MAX_INLINE_VIEW_LEN
134/// [`arrow_compute`]: https://docs.rs/arrow/latest/arrow/compute/index.html
135///
136/// Unlike [`GenericByteArray`], there are no constraints on the offsets other
137/// than they must point into a valid buffer. However, they can be out of order,
138/// non continuous and overlapping.
139///
140/// For example, in the following diagram, the strings "FishWasInTownToday" and
141/// "CrumpleFacedFish" are both longer than 12 bytes and thus are stored in a
142/// separate buffer while the string "LavaMonster" is stored inlined in the
143/// view. In this case, the same bytes for "Fish" are used to store both strings.
144///
145/// [`ByteView`]: arrow_data::ByteView
146///
147/// ```text
148/// ┌───┐
149/// ┌──────┬──────┬──────┬──────┐ offset │...│
150/// "FishWasInTownTodayYay" │ 21 │ Fish │ 0 │ 115 │─ ─ 103 │Mr.│
151/// └──────┴──────┴──────┴──────┘ │ ┌ ─ ─ ─ ─ ▶ │Cru│
152/// ┌──────┬──────┬──────┬──────┐ │mpl│
153/// "CrumpleFacedFish" │ 16 │ Crum │ 0 │ 103 │─ ─│─ ─ ─ ┘ │eFa│
154/// └──────┴──────┴──────┴──────┘ │ced│
155/// ┌──────┬────────────────────┐ └ ─ ─ ─ ─ ─ ─ ─ ─ ▶│Fis│
156/// "LavaMonster" │ 11 │ LavaMonster │ │hWa│
157/// └──────┴────────────────────┘ offset │sIn│
158/// 115 │Tow│
159/// │nTo│
160/// │day│
161/// u128 "views" │Yay│
162/// buffer 0 │...│
163/// └───┘
164/// ```
165pub struct GenericByteViewArray<T: ByteViewType + ?Sized> {
166 data_type: DataType,
167 views: ScalarBuffer<u128>,
168 buffers: Vec<Buffer>,
169 phantom: PhantomData<T>,
170 nulls: Option<NullBuffer>,
171}
172
173impl<T: ByteViewType + ?Sized> Clone for GenericByteViewArray<T> {
174 fn clone(&self) -> Self {
175 Self {
176 data_type: T::DATA_TYPE,
177 views: self.views.clone(),
178 buffers: self.buffers.clone(),
179 nulls: self.nulls.clone(),
180 phantom: Default::default(),
181 }
182 }
183}
184
185impl<T: ByteViewType + ?Sized> GenericByteViewArray<T> {
186 /// Create a new [`GenericByteViewArray`] from the provided parts, panicking on failure
187 ///
188 /// # Panics
189 ///
190 /// Panics if [`GenericByteViewArray::try_new`] returns an error
191 pub fn new(views: ScalarBuffer<u128>, buffers: Vec<Buffer>, nulls: Option<NullBuffer>) -> Self {
192 Self::try_new(views, buffers, nulls).unwrap()
193 }
194
195 /// Create a new [`GenericByteViewArray`] from the provided parts, returning an error on failure
196 ///
197 /// # Errors
198 ///
199 /// * `views.len() != nulls.len()`
200 /// * [ByteViewType::validate] fails
201 pub fn try_new(
202 views: ScalarBuffer<u128>,
203 buffers: Vec<Buffer>,
204 nulls: Option<NullBuffer>,
205 ) -> Result<Self, ArrowError> {
206 T::validate(&views, &buffers)?;
207
208 if let Some(n) = nulls.as_ref() {
209 if n.len() != views.len() {
210 return Err(ArrowError::InvalidArgumentError(format!(
211 "Incorrect length of null buffer for {}ViewArray, expected {} got {}",
212 T::PREFIX,
213 views.len(),
214 n.len(),
215 )));
216 }
217 }
218
219 Ok(Self {
220 data_type: T::DATA_TYPE,
221 views,
222 buffers,
223 nulls,
224 phantom: Default::default(),
225 })
226 }
227
228 /// Create a new [`GenericByteViewArray`] from the provided parts, without validation
229 ///
230 /// # Safety
231 ///
232 /// Safe if [`Self::try_new`] would not error
233 pub unsafe fn new_unchecked(
234 views: ScalarBuffer<u128>,
235 buffers: Vec<Buffer>,
236 nulls: Option<NullBuffer>,
237 ) -> Self {
238 if cfg!(feature = "force_validate") {
239 return Self::new(views, buffers, nulls);
240 }
241
242 Self {
243 data_type: T::DATA_TYPE,
244 phantom: Default::default(),
245 views,
246 buffers,
247 nulls,
248 }
249 }
250
251 /// Create a new [`GenericByteViewArray`] of length `len` where all values are null
252 pub fn new_null(len: usize) -> Self {
253 Self {
254 data_type: T::DATA_TYPE,
255 views: vec![0; len].into(),
256 buffers: vec![],
257 nulls: Some(NullBuffer::new_null(len)),
258 phantom: Default::default(),
259 }
260 }
261
262 /// Create a new [`Scalar`] from `value`
263 pub fn new_scalar(value: impl AsRef<T::Native>) -> Scalar<Self> {
264 Scalar::new(Self::from_iter_values(std::iter::once(value)))
265 }
266
267 /// Creates a [`GenericByteViewArray`] based on an iterator of values without nulls
268 pub fn from_iter_values<Ptr, I>(iter: I) -> Self
269 where
270 Ptr: AsRef<T::Native>,
271 I: IntoIterator<Item = Ptr>,
272 {
273 let iter = iter.into_iter();
274 let mut builder = GenericByteViewBuilder::<T>::with_capacity(iter.size_hint().0);
275 for v in iter {
276 builder.append_value(v);
277 }
278 builder.finish()
279 }
280
281 /// Deconstruct this array into its constituent parts
282 pub fn into_parts(self) -> (ScalarBuffer<u128>, Vec<Buffer>, Option<NullBuffer>) {
283 (self.views, self.buffers, self.nulls)
284 }
285
286 /// Returns the views buffer
287 #[inline]
288 pub fn views(&self) -> &ScalarBuffer<u128> {
289 &self.views
290 }
291
292 /// Returns the buffers storing string data
293 #[inline]
294 pub fn data_buffers(&self) -> &[Buffer] {
295 &self.buffers
296 }
297
298 /// Returns the element at index `i`
299 /// # Panics
300 /// Panics if index `i` is out of bounds.
301 pub fn value(&self, i: usize) -> &T::Native {
302 assert!(
303 i < self.len(),
304 "Trying to access an element at index {} from a {}ViewArray of length {}",
305 i,
306 T::PREFIX,
307 self.len()
308 );
309
310 unsafe { self.value_unchecked(i) }
311 }
312
313 /// Returns the element at index `i` without bounds checking
314 ///
315 /// # Safety
316 ///
317 /// Caller is responsible for ensuring that the index is within the bounds
318 /// of the array
319 pub unsafe fn value_unchecked(&self, idx: usize) -> &T::Native {
320 let v = self.views.get_unchecked(idx);
321 let len = *v as u32;
322 let b = if len <= MAX_INLINE_VIEW_LEN {
323 Self::inline_value(v, len as usize)
324 } else {
325 let view = ByteView::from(*v);
326 let data = self.buffers.get_unchecked(view.buffer_index as usize);
327 let offset = view.offset as usize;
328 data.get_unchecked(offset..offset + len as usize)
329 };
330 T::Native::from_bytes_unchecked(b)
331 }
332
333 /// Returns the first `len` bytes the inline value of the view.
334 ///
335 /// # Safety
336 /// - The `view` must be a valid element from `Self::views()` that adheres to the view layout.
337 /// - The `len` must be the length of the inlined value. It should never be larger than [`MAX_INLINE_VIEW_LEN`].
338 #[inline(always)]
339 pub unsafe fn inline_value(view: &u128, len: usize) -> &[u8] {
340 debug_assert!(len <= MAX_INLINE_VIEW_LEN as usize);
341 std::slice::from_raw_parts((view as *const u128 as *const u8).wrapping_add(4), len)
342 }
343
344 /// Constructs a new iterator for iterating over the values of this array
345 pub fn iter(&self) -> ArrayIter<&Self> {
346 ArrayIter::new(self)
347 }
348
349 /// Returns an iterator over the bytes of this array, including null values
350 pub fn bytes_iter(&self) -> impl Iterator<Item = &[u8]> {
351 self.views.iter().map(move |v| {
352 let len = *v as u32;
353 if len <= MAX_INLINE_VIEW_LEN {
354 unsafe { Self::inline_value(v, len as usize) }
355 } else {
356 let view = ByteView::from(*v);
357 let data = &self.buffers[view.buffer_index as usize];
358 let offset = view.offset as usize;
359 unsafe { data.get_unchecked(offset..offset + len as usize) }
360 }
361 })
362 }
363
364 /// Returns an iterator over the first `prefix_len` bytes of each array
365 /// element, including null values.
366 ///
367 /// If `prefix_len` is larger than the element's length, the iterator will
368 /// return an empty slice (`&[]`).
369 pub fn prefix_bytes_iter(&self, prefix_len: usize) -> impl Iterator<Item = &[u8]> {
370 self.views().into_iter().map(move |v| {
371 let len = (*v as u32) as usize;
372
373 if len < prefix_len {
374 return &[] as &[u8];
375 }
376
377 if prefix_len <= 4 || len as u32 <= MAX_INLINE_VIEW_LEN {
378 unsafe { StringViewArray::inline_value(v, prefix_len) }
379 } else {
380 let view = ByteView::from(*v);
381 let data = unsafe {
382 self.data_buffers()
383 .get_unchecked(view.buffer_index as usize)
384 };
385 let offset = view.offset as usize;
386 unsafe { data.get_unchecked(offset..offset + prefix_len) }
387 }
388 })
389 }
390
391 /// Returns an iterator over the last `suffix_len` bytes of each array
392 /// element, including null values.
393 ///
394 /// Note that for [`StringViewArray`] the last bytes may start in the middle
395 /// of a UTF-8 codepoint, and thus may not be a valid `&str`.
396 ///
397 /// If `suffix_len` is larger than the element's length, the iterator will
398 /// return an empty slice (`&[]`).
399 pub fn suffix_bytes_iter(&self, suffix_len: usize) -> impl Iterator<Item = &[u8]> {
400 self.views().into_iter().map(move |v| {
401 let len = (*v as u32) as usize;
402
403 if len < suffix_len {
404 return &[] as &[u8];
405 }
406
407 if len as u32 <= MAX_INLINE_VIEW_LEN {
408 unsafe { &StringViewArray::inline_value(v, len)[len - suffix_len..] }
409 } else {
410 let view = ByteView::from(*v);
411 let data = unsafe {
412 self.data_buffers()
413 .get_unchecked(view.buffer_index as usize)
414 };
415 let offset = view.offset as usize;
416 unsafe { data.get_unchecked(offset + len - suffix_len..offset + len) }
417 }
418 })
419 }
420
421 /// Returns a zero-copy slice of this array with the indicated offset and length.
422 pub fn slice(&self, offset: usize, length: usize) -> Self {
423 Self {
424 data_type: T::DATA_TYPE,
425 views: self.views.slice(offset, length),
426 buffers: self.buffers.clone(),
427 nulls: self.nulls.as_ref().map(|n| n.slice(offset, length)),
428 phantom: Default::default(),
429 }
430 }
431
432 /// Returns a "compacted" version of this array
433 ///
434 /// The original array will *not* be modified
435 ///
436 /// # Garbage Collection
437 ///
438 /// Before GC:
439 /// ```text
440 /// ┌──────┐
441 /// │......│
442 /// │......│
443 /// ┌────────────────────┐ ┌ ─ ─ ─ ▶ │Data1 │ Large buffer
444 /// │ View 1 │─ ─ ─ ─ │......│ with data that
445 /// ├────────────────────┤ │......│ is not referred
446 /// │ View 2 │─ ─ ─ ─ ─ ─ ─ ─▶ │Data2 │ to by View 1 or
447 /// └────────────────────┘ │......│ View 2
448 /// │......│
449 /// 2 views, refer to │......│
450 /// small portions of a └──────┘
451 /// large buffer
452 /// ```
453 ///
454 /// After GC:
455 ///
456 /// ```text
457 /// ┌────────────────────┐ ┌─────┐ After gc, only
458 /// │ View 1 │─ ─ ─ ─ ─ ─ ─ ─▶ │Data1│ data that is
459 /// ├────────────────────┤ ┌ ─ ─ ─ ▶ │Data2│ pointed to by
460 /// │ View 2 │─ ─ ─ ─ └─────┘ the views is
461 /// └────────────────────┘ left
462 ///
463 ///
464 /// 2 views
465 /// ```
466 /// This method will compact the data buffers by recreating the view array and only include the data
467 /// that is pointed to by the views.
468 ///
469 /// Note that it will copy the array regardless of whether the original array is compact.
470 /// Use with caution as this can be an expensive operation, only use it when you are sure that the view
471 /// array is significantly smaller than when it is originally created, e.g., after filtering or slicing.
472 ///
473 /// Note: this function does not attempt to canonicalize / deduplicate values. For this
474 /// feature see [`GenericByteViewBuilder::with_deduplicate_strings`].
475 pub fn gc(&self) -> Self {
476 // 1) Read basic properties once
477 let len = self.len(); // number of elements
478 let nulls = self.nulls().cloned(); // reuse & clone existing null bitmap
479
480 // 1.5) Fast path: if there are no buffers, just reuse original views and no data blocks
481 if self.data_buffers().is_empty() {
482 return unsafe {
483 GenericByteViewArray::new_unchecked(
484 self.views().clone(),
485 vec![], // empty data blocks
486 nulls,
487 )
488 };
489 }
490
491 // 2) Calculate total size of all non-inline data and detect if any exists
492 let total_large = self.total_buffer_bytes_used();
493
494 // 2.5) Fast path: if there is no non-inline data, avoid buffer allocation & processing
495 if total_large == 0 {
496 // Views are inline-only or all null; just reuse original views and no data blocks
497 return unsafe {
498 GenericByteViewArray::new_unchecked(
499 self.views().clone(),
500 vec![], // empty data blocks
501 nulls,
502 )
503 };
504 }
505
506 // 3) Allocate exactly capacity for all non-inline data
507 let mut data_buf = Vec::with_capacity(total_large);
508
509 // 4) Iterate over views and process each inline/non-inline view
510 let views_buf: Vec<u128> = (0..len)
511 .map(|i| unsafe { self.copy_view_to_buffer(i, &mut data_buf) })
512 .collect();
513
514 // 5) Wrap up buffers
515 let data_block = Buffer::from_vec(data_buf);
516 let views_scalar = ScalarBuffer::from(views_buf);
517 let data_blocks = vec![data_block];
518
519 // SAFETY: views_scalar, data_blocks, and nulls are correctly aligned and sized
520 unsafe { GenericByteViewArray::new_unchecked(views_scalar, data_blocks, nulls) }
521 }
522
523 /// Copy the i‑th view into `data_buf` if it refers to an out‑of‑line buffer.
524 ///
525 /// # Safety
526 ///
527 /// - `i < self.len()`.
528 /// - Every element in `self.views()` must currently refer to a valid slice
529 /// inside one of `self.buffers`.
530 /// - `data_buf` must be ready to have additional bytes appended.
531 /// - After this call, the returned view will have its
532 /// `buffer_index` reset to `0` and its `offset` updated so that it points
533 /// into the bytes just appended at the end of `data_buf`.
534 #[inline(always)]
535 unsafe fn copy_view_to_buffer(&self, i: usize, data_buf: &mut Vec<u8>) -> u128 {
536 // SAFETY: `i < self.len()` ensures this is in‑bounds.
537 let raw_view = *self.views().get_unchecked(i);
538 let mut bv = ByteView::from(raw_view);
539
540 // Inline‑small views stay as‑is.
541 if bv.length <= MAX_INLINE_VIEW_LEN {
542 raw_view
543 } else {
544 // SAFETY: `bv.buffer_index` and `bv.offset..bv.offset+bv.length`
545 // must both lie within valid ranges for `self.buffers`.
546 let buffer = self.buffers.get_unchecked(bv.buffer_index as usize);
547 let start = bv.offset as usize;
548 let end = start + bv.length as usize;
549 let slice = buffer.get_unchecked(start..end);
550
551 // Copy out‑of‑line data into our single “0” buffer.
552 let new_offset = data_buf.len() as u32;
553 data_buf.extend_from_slice(slice);
554
555 bv.buffer_index = 0;
556 bv.offset = new_offset;
557 bv.into()
558 }
559 }
560
561 /// Returns the total number of bytes used by all non inlined views in all
562 /// buffers.
563 ///
564 /// Note this does not account for views that point at the same underlying
565 /// data in buffers
566 ///
567 /// For example, if the array has three strings views:
568 /// * View with length = 9 (inlined)
569 /// * View with length = 32 (non inlined)
570 /// * View with length = 16 (non inlined)
571 ///
572 /// Then this method would report 48
573 pub fn total_buffer_bytes_used(&self) -> usize {
574 self.views()
575 .iter()
576 .map(|v| {
577 let len = *v as u32;
578 if len > MAX_INLINE_VIEW_LEN {
579 len as usize
580 } else {
581 0
582 }
583 })
584 .sum()
585 }
586
587 /// Compare two [`GenericByteViewArray`] at index `left_idx` and `right_idx`
588 ///
589 /// Comparing two ByteView types are non-trivial.
590 /// It takes a bit of patience to understand why we don't just compare two &[u8] directly.
591 ///
592 /// ByteView types give us the following two advantages, and we need to be careful not to lose them:
593 /// (1) For string/byte smaller than [`MAX_INLINE_VIEW_LEN`] bytes, the entire data is inlined in the view.
594 /// Meaning that reading one array element requires only one memory access
595 /// (two memory access required for StringArray, one for offset buffer, the other for value buffer).
596 ///
597 /// (2) For string/byte larger than [`MAX_INLINE_VIEW_LEN`] bytes, we can still be faster than (for certain operations) StringArray/ByteArray,
598 /// thanks to the inlined 4 bytes.
599 /// Consider equality check:
600 /// If the first four bytes of the two strings are different, we can return false immediately (with just one memory access).
601 ///
602 /// If we directly compare two &[u8], we materialize the entire string (i.e., make multiple memory accesses), which might be unnecessary.
603 /// - Most of the time (eq, ord), we only need to look at the first 4 bytes to know the answer,
604 /// e.g., if the inlined 4 bytes are different, we can directly return unequal without looking at the full string.
605 ///
606 /// # Order check flow
607 /// (1) if both string are smaller than [`MAX_INLINE_VIEW_LEN`] bytes, we can directly compare the data inlined to the view.
608 /// (2) if any of the string is larger than [`MAX_INLINE_VIEW_LEN`] bytes, we need to compare the full string.
609 /// (2.1) if the inlined 4 bytes are different, we can return the result immediately.
610 /// (2.2) o.w., we need to compare the full string.
611 ///
612 /// # Safety
613 /// The left/right_idx must within range of each array
614 pub unsafe fn compare_unchecked(
615 left: &GenericByteViewArray<T>,
616 left_idx: usize,
617 right: &GenericByteViewArray<T>,
618 right_idx: usize,
619 ) -> Ordering {
620 let l_view = left.views().get_unchecked(left_idx);
621 let l_byte_view = ByteView::from(*l_view);
622
623 let r_view = right.views().get_unchecked(right_idx);
624 let r_byte_view = ByteView::from(*r_view);
625
626 let l_len = l_byte_view.length;
627 let r_len = r_byte_view.length;
628
629 if l_len <= 12 && r_len <= 12 {
630 return Self::inline_key_fast(*l_view).cmp(&Self::inline_key_fast(*r_view));
631 }
632
633 // one of the string is larger than 12 bytes,
634 // we then try to compare the inlined data first
635
636 // Note: In theory, ByteView is only used for string which is larger than 12 bytes,
637 // but we can still use it to get the inlined prefix for shorter strings.
638 // The prefix is always the first 4 bytes of the view, for both short and long strings.
639 let l_inlined_be = l_byte_view.prefix.swap_bytes();
640 let r_inlined_be = r_byte_view.prefix.swap_bytes();
641 if l_inlined_be != r_inlined_be {
642 return l_inlined_be.cmp(&r_inlined_be);
643 }
644
645 // unfortunately, we need to compare the full data
646 let l_full_data: &[u8] = unsafe { left.value_unchecked(left_idx).as_ref() };
647 let r_full_data: &[u8] = unsafe { right.value_unchecked(right_idx).as_ref() };
648
649 l_full_data.cmp(r_full_data)
650 }
651
652 /// Builds a 128-bit composite key for an inline value:
653 ///
654 /// - High 96 bits: the inline data in big-endian byte order (for correct lexicographical sorting).
655 /// - Low 32 bits: the length in big-endian byte order, acting as a tiebreaker so shorter strings
656 /// (or those with fewer meaningful bytes) always numerically sort before longer ones.
657 ///
658 /// This function extracts the length and the 12-byte inline string data from the raw
659 /// little-endian `u128` representation, converts them to big-endian ordering, and packs them
660 /// into a single `u128` value suitable for fast, branchless comparisons.
661 ///
662 /// # Why include length?
663 ///
664 /// A pure 96-bit content comparison can’t distinguish between two values whose inline bytes
665 /// compare equal—either because one is a true prefix of the other or because zero-padding
666 /// hides extra bytes. By tucking the 32-bit length into the lower bits, a single `u128` compare
667 /// handles both content and length in one go.
668 ///
669 /// Example: comparing "bar" (3 bytes) vs "bar\0" (4 bytes)
670 ///
671 /// | String | Bytes 0–4 (length LE) | Bytes 4–16 (data + padding) |
672 /// |------------|-----------------------|---------------------------------|
673 /// | `"bar"` | `03 00 00 00` | `62 61 72` + 9 × `00` |
674 /// | `"bar\0"`| `04 00 00 00` | `62 61 72 00` + 8 × `00` |
675 ///
676 /// Both inline parts become `62 61 72 00…00`, so they tie on content. The length field
677 /// then differentiates:
678 ///
679 /// ```text
680 /// key("bar") = 0x0000000000000000000062617200000003
681 /// key("bar\0") = 0x0000000000000000000062617200000004
682 /// ⇒ key("bar") < key("bar\0")
683 /// ```
684 /// # Inlining and Endianness
685 ///
686 /// - We start by calling `.to_le_bytes()` on the `raw` `u128`, because Rust’s native in‑memory
687 /// representation is little‑endian on x86/ARM.
688 /// - We extract the low 32 bits numerically (`raw as u32`)—this step is endianness‑free.
689 /// - We copy the 12 bytes of inline data (original order) into `buf[0..12]`.
690 /// - We serialize `length` as big‑endian into `buf[12..16]`.
691 /// - Finally, `u128::from_be_bytes(buf)` treats `buf[0]` as the most significant byte
692 /// and `buf[15]` as the least significant, producing a `u128` whose integer value
693 /// directly encodes “inline data then length” in big‑endian form.
694 ///
695 /// This ensures that a simple `u128` comparison is equivalent to the desired
696 /// lexicographical comparison of the inline bytes followed by length.
697 #[inline(always)]
698 pub fn inline_key_fast(raw: u128) -> u128 {
699 // 1. Decompose `raw` into little‑endian bytes:
700 // - raw_bytes[0..4] = length in LE
701 // - raw_bytes[4..16] = inline string data
702 let raw_bytes = raw.to_le_bytes();
703
704 // 2. Numerically truncate to get the low 32‑bit length (endianness‑free).
705 let length = raw as u32;
706
707 // 3. Build a 16‑byte buffer in big‑endian order:
708 // - buf[0..12] = inline string bytes (in original order)
709 // - buf[12..16] = length.to_be_bytes() (BE)
710 let mut buf = [0u8; 16];
711 buf[0..12].copy_from_slice(&raw_bytes[4..16]); // inline data
712
713 // Why convert length to big-endian for comparison?
714 //
715 // Rust (on most platforms) stores integers in little-endian format,
716 // meaning the least significant byte is at the lowest memory address.
717 // For example, an u32 value like 0x22345677 is stored in memory as:
718 //
719 // [0x77, 0x56, 0x34, 0x22] // little-endian layout
720 // ^ ^ ^ ^
721 // LSB ↑↑↑ MSB
722 //
723 // This layout is efficient for arithmetic but *not* suitable for
724 // lexicographic (dictionary-style) comparison of byte arrays.
725 //
726 // To compare values by byte order—e.g., for sorted keys or binary trees—
727 // we must convert them to **big-endian**, where:
728 //
729 // - The most significant byte (MSB) comes first (index 0)
730 // - The least significant byte (LSB) comes last (index N-1)
731 //
732 // In big-endian, the same u32 = 0x22345677 would be represented as:
733 //
734 // [0x22, 0x34, 0x56, 0x77]
735 //
736 // This ordering aligns with natural string/byte sorting, so calling
737 // `.to_be_bytes()` allows us to construct
738 // keys where standard numeric comparison (e.g., `<`, `>`) behaves
739 // like lexicographic byte comparison.
740 buf[12..16].copy_from_slice(&length.to_be_bytes()); // length in BE
741
742 // 4. Deserialize the buffer as a big‑endian u128:
743 // buf[0] is MSB, buf[15] is LSB.
744 // Details:
745 // Note on endianness and layout:
746 //
747 // Although `buf[0]` is stored at the lowest memory address,
748 // calling `u128::from_be_bytes(buf)` interprets it as the **most significant byte (MSB)**,
749 // and `buf[15]` as the **least significant byte (LSB)**.
750 //
751 // This is the core principle of **big-endian decoding**:
752 // - Byte at index 0 maps to bits 127..120 (highest)
753 // - Byte at index 1 maps to bits 119..112
754 // - ...
755 // - Byte at index 15 maps to bits 7..0 (lowest)
756 //
757 // So even though memory layout goes from low to high (left to right),
758 // big-endian treats the **first byte** as highest in value.
759 //
760 // This guarantees that comparing two `u128` keys is equivalent to lexicographically
761 // comparing the original inline bytes, followed by length.
762 u128::from_be_bytes(buf)
763 }
764}
765
766impl<T: ByteViewType + ?Sized> Debug for GenericByteViewArray<T> {
767 fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
768 write!(f, "{}ViewArray\n[\n", T::PREFIX)?;
769 print_long_array(self, f, |array, index, f| {
770 std::fmt::Debug::fmt(&array.value(index), f)
771 })?;
772 write!(f, "]")
773 }
774}
775
776impl<T: ByteViewType + ?Sized> Array for GenericByteViewArray<T> {
777 fn as_any(&self) -> &dyn Any {
778 self
779 }
780
781 fn to_data(&self) -> ArrayData {
782 self.clone().into()
783 }
784
785 fn into_data(self) -> ArrayData {
786 self.into()
787 }
788
789 fn data_type(&self) -> &DataType {
790 &self.data_type
791 }
792
793 fn slice(&self, offset: usize, length: usize) -> ArrayRef {
794 Arc::new(self.slice(offset, length))
795 }
796
797 fn len(&self) -> usize {
798 self.views.len()
799 }
800
801 fn is_empty(&self) -> bool {
802 self.views.is_empty()
803 }
804
805 fn shrink_to_fit(&mut self) {
806 self.views.shrink_to_fit();
807 self.buffers.iter_mut().for_each(|b| b.shrink_to_fit());
808 self.buffers.shrink_to_fit();
809 if let Some(nulls) = &mut self.nulls {
810 nulls.shrink_to_fit();
811 }
812 }
813
814 fn offset(&self) -> usize {
815 0
816 }
817
818 fn nulls(&self) -> Option<&NullBuffer> {
819 self.nulls.as_ref()
820 }
821
822 fn logical_null_count(&self) -> usize {
823 // More efficient that the default implementation
824 self.null_count()
825 }
826
827 fn get_buffer_memory_size(&self) -> usize {
828 let mut sum = self.buffers.iter().map(|b| b.capacity()).sum::<usize>();
829 sum += self.views.inner().capacity();
830 if let Some(x) = &self.nulls {
831 sum += x.buffer().capacity()
832 }
833 sum
834 }
835
836 fn get_array_memory_size(&self) -> usize {
837 std::mem::size_of::<Self>() + self.get_buffer_memory_size()
838 }
839}
840
841impl<'a, T: ByteViewType + ?Sized> ArrayAccessor for &'a GenericByteViewArray<T> {
842 type Item = &'a T::Native;
843
844 fn value(&self, index: usize) -> Self::Item {
845 GenericByteViewArray::value(self, index)
846 }
847
848 unsafe fn value_unchecked(&self, index: usize) -> Self::Item {
849 GenericByteViewArray::value_unchecked(self, index)
850 }
851}
852
853impl<'a, T: ByteViewType + ?Sized> IntoIterator for &'a GenericByteViewArray<T> {
854 type Item = Option<&'a T::Native>;
855 type IntoIter = ArrayIter<Self>;
856
857 fn into_iter(self) -> Self::IntoIter {
858 ArrayIter::new(self)
859 }
860}
861
862impl<T: ByteViewType + ?Sized> From<ArrayData> for GenericByteViewArray<T> {
863 fn from(value: ArrayData) -> Self {
864 let views = value.buffers()[0].clone();
865 let views = ScalarBuffer::new(views, value.offset(), value.len());
866 let buffers = value.buffers()[1..].to_vec();
867 Self {
868 data_type: T::DATA_TYPE,
869 views,
870 buffers,
871 nulls: value.nulls().cloned(),
872 phantom: Default::default(),
873 }
874 }
875}
876
877/// Efficiently convert a [`GenericByteArray`] to a [`GenericByteViewArray`]
878///
879/// For example this method can convert a [`StringArray`] to a
880/// [`StringViewArray`].
881///
882/// If the offsets are all less than u32::MAX, the new [`GenericByteViewArray`]
883/// is built without copying the underlying string data (views are created
884/// directly into the existing buffer)
885///
886/// [`StringArray`]: crate::StringArray
887impl<FROM, V> From<&GenericByteArray<FROM>> for GenericByteViewArray<V>
888where
889 FROM: ByteArrayType,
890 FROM::Offset: OffsetSizeTrait + ToPrimitive,
891 V: ByteViewType<Native = FROM::Native>,
892{
893 fn from(byte_array: &GenericByteArray<FROM>) -> Self {
894 let offsets = byte_array.offsets();
895
896 let can_reuse_buffer = match offsets.last() {
897 Some(offset) => offset.as_usize() < u32::MAX as usize,
898 None => true,
899 };
900
901 if can_reuse_buffer {
902 // build views directly pointing to the existing buffer
903 let len = byte_array.len();
904 let mut views_builder = GenericByteViewBuilder::<V>::with_capacity(len);
905 let str_values_buf = byte_array.values().clone();
906 let block = views_builder.append_block(str_values_buf);
907 for (i, w) in offsets.windows(2).enumerate() {
908 let offset = w[0].as_usize();
909 let end = w[1].as_usize();
910 let length = end - offset;
911
912 if byte_array.is_null(i) {
913 views_builder.append_null();
914 } else {
915 // Safety: the input was a valid array so it valid UTF8 (if string). And
916 // all offsets were valid
917 unsafe {
918 views_builder.append_view_unchecked(block, offset as u32, length as u32)
919 }
920 }
921 }
922 assert_eq!(views_builder.len(), len);
923 views_builder.finish()
924 } else {
925 // Otherwise, create a new buffer for large strings
926 // TODO: the original buffer could still be used
927 // by making multiple slices of u32::MAX length
928 GenericByteViewArray::<V>::from_iter(byte_array.iter())
929 }
930 }
931}
932
933impl<T: ByteViewType + ?Sized> From<GenericByteViewArray<T>> for ArrayData {
934 fn from(mut array: GenericByteViewArray<T>) -> Self {
935 let len = array.len();
936 array.buffers.insert(0, array.views.into_inner());
937 let builder = ArrayDataBuilder::new(T::DATA_TYPE)
938 .len(len)
939 .buffers(array.buffers)
940 .nulls(array.nulls);
941
942 unsafe { builder.build_unchecked() }
943 }
944}
945
946impl<'a, Ptr, T> FromIterator<&'a Option<Ptr>> for GenericByteViewArray<T>
947where
948 Ptr: AsRef<T::Native> + 'a,
949 T: ByteViewType + ?Sized,
950{
951 fn from_iter<I: IntoIterator<Item = &'a Option<Ptr>>>(iter: I) -> Self {
952 iter.into_iter()
953 .map(|o| o.as_ref().map(|p| p.as_ref()))
954 .collect()
955 }
956}
957
958impl<Ptr, T: ByteViewType + ?Sized> FromIterator<Option<Ptr>> for GenericByteViewArray<T>
959where
960 Ptr: AsRef<T::Native>,
961{
962 fn from_iter<I: IntoIterator<Item = Option<Ptr>>>(iter: I) -> Self {
963 let iter = iter.into_iter();
964 let mut builder = GenericByteViewBuilder::<T>::with_capacity(iter.size_hint().0);
965 builder.extend(iter);
966 builder.finish()
967 }
968}
969
970/// A [`GenericByteViewArray`] of `[u8]`
971///
972/// See [`GenericByteViewArray`] for format and layout details.
973///
974/// # Example
975/// ```
976/// use arrow_array::BinaryViewArray;
977/// let array = BinaryViewArray::from_iter_values(vec![b"hello" as &[u8], b"world", b"lulu", b"large payload over 12 bytes"]);
978/// assert_eq!(array.value(0), b"hello");
979/// assert_eq!(array.value(3), b"large payload over 12 bytes");
980/// ```
981pub type BinaryViewArray = GenericByteViewArray<BinaryViewType>;
982
983impl BinaryViewArray {
984 /// Convert the [`BinaryViewArray`] to [`StringViewArray`]
985 /// If items not utf8 data, validate will fail and error returned.
986 pub fn to_string_view(self) -> Result<StringViewArray, ArrowError> {
987 StringViewType::validate(self.views(), self.data_buffers())?;
988 unsafe { Ok(self.to_string_view_unchecked()) }
989 }
990
991 /// Convert the [`BinaryViewArray`] to [`StringViewArray`]
992 /// # Safety
993 /// Caller is responsible for ensuring that items in array are utf8 data.
994 pub unsafe fn to_string_view_unchecked(self) -> StringViewArray {
995 StringViewArray::new_unchecked(self.views, self.buffers, self.nulls)
996 }
997}
998
999impl From<Vec<&[u8]>> for BinaryViewArray {
1000 fn from(v: Vec<&[u8]>) -> Self {
1001 Self::from_iter_values(v)
1002 }
1003}
1004
1005impl From<Vec<Option<&[u8]>>> for BinaryViewArray {
1006 fn from(v: Vec<Option<&[u8]>>) -> Self {
1007 v.into_iter().collect()
1008 }
1009}
1010
1011/// A [`GenericByteViewArray`] that stores utf8 data
1012///
1013/// See [`GenericByteViewArray`] for format and layout details.
1014///
1015/// # Example
1016/// ```
1017/// use arrow_array::StringViewArray;
1018/// let array = StringViewArray::from_iter_values(vec!["hello", "world", "lulu", "large payload over 12 bytes"]);
1019/// assert_eq!(array.value(0), "hello");
1020/// assert_eq!(array.value(3), "large payload over 12 bytes");
1021/// ```
1022pub type StringViewArray = GenericByteViewArray<StringViewType>;
1023
1024impl StringViewArray {
1025 /// Convert the [`StringViewArray`] to [`BinaryViewArray`]
1026 pub fn to_binary_view(self) -> BinaryViewArray {
1027 unsafe { BinaryViewArray::new_unchecked(self.views, self.buffers, self.nulls) }
1028 }
1029
1030 /// Returns true if all data within this array is ASCII
1031 pub fn is_ascii(&self) -> bool {
1032 // Alternative (but incorrect): directly check the underlying buffers
1033 // (1) Our string view might be sparse, i.e., a subset of the buffers,
1034 // so even if the buffer is not ascii, we can still be ascii.
1035 // (2) It is quite difficult to know the range of each buffer (unlike StringArray)
1036 // This means that this operation is quite expensive, shall we cache the result?
1037 // i.e. track `is_ascii` in the builder.
1038 self.iter().all(|v| match v {
1039 Some(v) => v.is_ascii(),
1040 None => true,
1041 })
1042 }
1043}
1044
1045impl From<Vec<&str>> for StringViewArray {
1046 fn from(v: Vec<&str>) -> Self {
1047 Self::from_iter_values(v)
1048 }
1049}
1050
1051impl From<Vec<Option<&str>>> for StringViewArray {
1052 fn from(v: Vec<Option<&str>>) -> Self {
1053 v.into_iter().collect()
1054 }
1055}
1056
1057impl From<Vec<String>> for StringViewArray {
1058 fn from(v: Vec<String>) -> Self {
1059 Self::from_iter_values(v)
1060 }
1061}
1062
1063impl From<Vec<Option<String>>> for StringViewArray {
1064 fn from(v: Vec<Option<String>>) -> Self {
1065 v.into_iter().collect()
1066 }
1067}
1068
1069#[cfg(test)]
1070mod tests {
1071 use crate::builder::{BinaryViewBuilder, StringViewBuilder};
1072 use crate::types::BinaryViewType;
1073 use crate::{
1074 Array, BinaryViewArray, GenericBinaryArray, GenericByteViewArray, StringViewArray,
1075 };
1076 use arrow_buffer::{Buffer, ScalarBuffer};
1077 use arrow_data::{ByteView, MAX_INLINE_VIEW_LEN};
1078 use rand::prelude::StdRng;
1079 use rand::{Rng, SeedableRng};
1080
1081 const BLOCK_SIZE: u32 = 8;
1082
1083 #[test]
1084 fn try_new_string() {
1085 let array = StringViewArray::from_iter_values(vec![
1086 "hello",
1087 "world",
1088 "lulu",
1089 "large payload over 12 bytes",
1090 ]);
1091 assert_eq!(array.value(0), "hello");
1092 assert_eq!(array.value(3), "large payload over 12 bytes");
1093 }
1094
1095 #[test]
1096 fn try_new_binary() {
1097 let array = BinaryViewArray::from_iter_values(vec![
1098 b"hello".as_slice(),
1099 b"world".as_slice(),
1100 b"lulu".as_slice(),
1101 b"large payload over 12 bytes".as_slice(),
1102 ]);
1103 assert_eq!(array.value(0), b"hello");
1104 assert_eq!(array.value(3), b"large payload over 12 bytes");
1105 }
1106
1107 #[test]
1108 fn try_new_empty_string() {
1109 // test empty array
1110 let array = {
1111 let mut builder = StringViewBuilder::new();
1112 builder.finish()
1113 };
1114 assert!(array.is_empty());
1115 }
1116
1117 #[test]
1118 fn try_new_empty_binary() {
1119 // test empty array
1120 let array = {
1121 let mut builder = BinaryViewBuilder::new();
1122 builder.finish()
1123 };
1124 assert!(array.is_empty());
1125 }
1126
1127 #[test]
1128 fn test_append_string() {
1129 // test builder append
1130 let array = {
1131 let mut builder = StringViewBuilder::new();
1132 builder.append_value("hello");
1133 builder.append_null();
1134 builder.append_option(Some("large payload over 12 bytes"));
1135 builder.finish()
1136 };
1137 assert_eq!(array.value(0), "hello");
1138 assert!(array.is_null(1));
1139 assert_eq!(array.value(2), "large payload over 12 bytes");
1140 }
1141
1142 #[test]
1143 fn test_append_binary() {
1144 // test builder append
1145 let array = {
1146 let mut builder = BinaryViewBuilder::new();
1147 builder.append_value(b"hello");
1148 builder.append_null();
1149 builder.append_option(Some(b"large payload over 12 bytes"));
1150 builder.finish()
1151 };
1152 assert_eq!(array.value(0), b"hello");
1153 assert!(array.is_null(1));
1154 assert_eq!(array.value(2), b"large payload over 12 bytes");
1155 }
1156
1157 #[test]
1158 fn test_in_progress_recreation() {
1159 let array = {
1160 // make a builder with small block size.
1161 let mut builder = StringViewBuilder::new().with_fixed_block_size(14);
1162 builder.append_value("large payload over 12 bytes");
1163 builder.append_option(Some("another large payload over 12 bytes that double than the first one, so that we can trigger the in_progress in builder re-created"));
1164 builder.finish()
1165 };
1166 assert_eq!(array.value(0), "large payload over 12 bytes");
1167 assert_eq!(array.value(1), "another large payload over 12 bytes that double than the first one, so that we can trigger the in_progress in builder re-created");
1168 assert_eq!(2, array.buffers.len());
1169 }
1170
1171 #[test]
1172 #[should_panic(expected = "Invalid buffer index at 0: got index 3 but only has 1 buffers")]
1173 fn new_with_invalid_view_data() {
1174 let v = "large payload over 12 bytes";
1175 let view = ByteView::new(13, &v.as_bytes()[0..4])
1176 .with_buffer_index(3)
1177 .with_offset(1);
1178 let views = ScalarBuffer::from(vec![view.into()]);
1179 let buffers = vec![Buffer::from_slice_ref(v)];
1180 StringViewArray::new(views, buffers, None);
1181 }
1182
1183 #[test]
1184 #[should_panic(
1185 expected = "Encountered non-UTF-8 data at index 0: invalid utf-8 sequence of 1 bytes from index 0"
1186 )]
1187 fn new_with_invalid_utf8_data() {
1188 let v: Vec<u8> = vec![
1189 // invalid UTF8
1190 0xf0, 0x80, 0x80, 0x80, // more bytes to make it larger than 12
1191 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
1192 ];
1193 let view = ByteView::new(v.len() as u32, &v[0..4]);
1194 let views = ScalarBuffer::from(vec![view.into()]);
1195 let buffers = vec![Buffer::from_slice_ref(v)];
1196 StringViewArray::new(views, buffers, None);
1197 }
1198
1199 #[test]
1200 #[should_panic(expected = "View at index 0 contained non-zero padding for string of length 1")]
1201 fn new_with_invalid_zero_padding() {
1202 let mut data = [0; 12];
1203 data[0] = b'H';
1204 data[11] = 1; // no zero padding
1205
1206 let mut view_buffer = [0; 16];
1207 view_buffer[0..4].copy_from_slice(&1u32.to_le_bytes());
1208 view_buffer[4..].copy_from_slice(&data);
1209
1210 let view = ByteView::from(u128::from_le_bytes(view_buffer));
1211 let views = ScalarBuffer::from(vec![view.into()]);
1212 let buffers = vec![];
1213 StringViewArray::new(views, buffers, None);
1214 }
1215
1216 #[test]
1217 #[should_panic(expected = "Mismatch between embedded prefix and data")]
1218 fn test_mismatch_between_embedded_prefix_and_data() {
1219 let input_str_1 = "Hello, Rustaceans!";
1220 let input_str_2 = "Hallo, Rustaceans!";
1221 let length = input_str_1.len() as u32;
1222 assert!(input_str_1.len() > 12);
1223
1224 let mut view_buffer = [0; 16];
1225 view_buffer[0..4].copy_from_slice(&length.to_le_bytes());
1226 view_buffer[4..8].copy_from_slice(&input_str_1.as_bytes()[0..4]);
1227 view_buffer[8..12].copy_from_slice(&0u32.to_le_bytes());
1228 view_buffer[12..].copy_from_slice(&0u32.to_le_bytes());
1229 let view = ByteView::from(u128::from_le_bytes(view_buffer));
1230 let views = ScalarBuffer::from(vec![view.into()]);
1231 let buffers = vec![Buffer::from_slice_ref(input_str_2.as_bytes())];
1232
1233 StringViewArray::new(views, buffers, None);
1234 }
1235
1236 #[test]
1237 fn test_gc() {
1238 let test_data = [
1239 Some("longer than 12 bytes"),
1240 Some("short"),
1241 Some("t"),
1242 Some("longer than 12 bytes"),
1243 None,
1244 Some("short"),
1245 ];
1246
1247 let array = {
1248 let mut builder = StringViewBuilder::new().with_fixed_block_size(8); // create multiple buffers
1249 test_data.into_iter().for_each(|v| builder.append_option(v));
1250 builder.finish()
1251 };
1252 assert!(array.buffers.len() > 1);
1253
1254 fn check_gc(to_test: &StringViewArray) {
1255 let gc = to_test.gc();
1256 assert_ne!(to_test.data_buffers().len(), gc.data_buffers().len());
1257
1258 to_test.iter().zip(gc.iter()).for_each(|(a, b)| {
1259 assert_eq!(a, b);
1260 });
1261 assert_eq!(to_test.len(), gc.len());
1262 }
1263
1264 check_gc(&array);
1265 check_gc(&array.slice(1, 3));
1266 check_gc(&array.slice(2, 1));
1267 check_gc(&array.slice(2, 2));
1268 check_gc(&array.slice(3, 1));
1269 }
1270
1271 /// 1) Empty array: no elements, expect gc to return empty with no data buffers
1272 #[test]
1273 fn test_gc_empty_array() {
1274 let array = StringViewBuilder::new()
1275 .with_fixed_block_size(BLOCK_SIZE)
1276 .finish();
1277 let gced = array.gc();
1278 // length and null count remain zero
1279 assert_eq!(gced.len(), 0);
1280 assert_eq!(gced.null_count(), 0);
1281 // no underlying data buffers should be allocated
1282 assert!(
1283 gced.data_buffers().is_empty(),
1284 "Expected no data buffers for empty array"
1285 );
1286 }
1287
1288 /// 2) All inline values (<= INLINE_LEN): capacity-only data buffer, same values
1289 #[test]
1290 fn test_gc_all_inline() {
1291 let mut builder = StringViewBuilder::new().with_fixed_block_size(BLOCK_SIZE);
1292 // append many short strings, each exactly INLINE_LEN long
1293 for _ in 0..100 {
1294 let s = "A".repeat(MAX_INLINE_VIEW_LEN as usize);
1295 builder.append_option(Some(&s));
1296 }
1297 let array = builder.finish();
1298 let gced = array.gc();
1299 // Since all views fit inline, data buffer is empty
1300 assert_eq!(
1301 gced.data_buffers().len(),
1302 0,
1303 "Should have no data buffers for inline values"
1304 );
1305 assert_eq!(gced.len(), 100);
1306 // verify element-wise equality
1307 array.iter().zip(gced.iter()).for_each(|(orig, got)| {
1308 assert_eq!(orig, got, "Inline value mismatch after gc");
1309 });
1310 }
1311
1312 /// 3) All large values (> INLINE_LEN): each must be copied into the new data buffer
1313 #[test]
1314 fn test_gc_all_large() {
1315 let mut builder = StringViewBuilder::new().with_fixed_block_size(BLOCK_SIZE);
1316 let large_str = "X".repeat(MAX_INLINE_VIEW_LEN as usize + 5);
1317 // append multiple large strings
1318 for _ in 0..50 {
1319 builder.append_option(Some(&large_str));
1320 }
1321 let array = builder.finish();
1322 let gced = array.gc();
1323 // New data buffers should be populated (one or more blocks)
1324 assert!(
1325 !gced.data_buffers().is_empty(),
1326 "Expected data buffers for large values"
1327 );
1328 assert_eq!(gced.len(), 50);
1329 // verify that every large string emerges unchanged
1330 array.iter().zip(gced.iter()).for_each(|(orig, got)| {
1331 assert_eq!(orig, got, "Large view mismatch after gc");
1332 });
1333 }
1334
1335 /// 4) All null elements: ensure null bitmap handling path is correct
1336 #[test]
1337 fn test_gc_all_nulls() {
1338 let mut builder = StringViewBuilder::new().with_fixed_block_size(BLOCK_SIZE);
1339 for _ in 0..20 {
1340 builder.append_null();
1341 }
1342 let array = builder.finish();
1343 let gced = array.gc();
1344 // length and null count match
1345 assert_eq!(gced.len(), 20);
1346 assert_eq!(gced.null_count(), 20);
1347 // data buffers remain empty for null-only array
1348 assert!(
1349 gced.data_buffers().is_empty(),
1350 "No data should be stored for nulls"
1351 );
1352 }
1353
1354 /// 5) Random mix of inline, large, and null values with slicing tests
1355 #[test]
1356 fn test_gc_random_mixed_and_slices() {
1357 let mut rng = StdRng::seed_from_u64(42);
1358 let mut builder = StringViewBuilder::new().with_fixed_block_size(BLOCK_SIZE);
1359 // Keep a Vec of original Option<String> for later comparison
1360 let mut original: Vec<Option<String>> = Vec::new();
1361
1362 for _ in 0..200 {
1363 if rng.random_bool(0.1) {
1364 // 10% nulls
1365 builder.append_null();
1366 original.push(None);
1367 } else {
1368 // random length between 0 and twice the inline limit
1369 let len = rng.random_range(0..(MAX_INLINE_VIEW_LEN * 2));
1370 let s: String = "A".repeat(len as usize);
1371 builder.append_option(Some(&s));
1372 original.push(Some(s));
1373 }
1374 }
1375
1376 let array = builder.finish();
1377 // Test multiple slice ranges to ensure offset logic is correct
1378 for (offset, slice_len) in &[(0, 50), (10, 100), (150, 30)] {
1379 let sliced = array.slice(*offset, *slice_len);
1380 let gced = sliced.gc();
1381 // Build expected slice of Option<&str>
1382 let expected: Vec<Option<&str>> = original[*offset..(*offset + *slice_len)]
1383 .iter()
1384 .map(|opt| opt.as_deref())
1385 .collect();
1386
1387 assert_eq!(gced.len(), *slice_len, "Slice length mismatch");
1388 // Compare element-wise
1389 gced.iter().zip(expected.iter()).for_each(|(got, expect)| {
1390 assert_eq!(got, *expect, "Value mismatch in mixed slice after gc");
1391 });
1392 }
1393 }
1394
1395 #[test]
1396 fn test_eq() {
1397 let test_data = [
1398 Some("longer than 12 bytes"),
1399 None,
1400 Some("short"),
1401 Some("again, this is longer than 12 bytes"),
1402 ];
1403
1404 let array1 = {
1405 let mut builder = StringViewBuilder::new().with_fixed_block_size(8);
1406 test_data.into_iter().for_each(|v| builder.append_option(v));
1407 builder.finish()
1408 };
1409 let array2 = {
1410 // create a new array with the same data but different layout
1411 let mut builder = StringViewBuilder::new().with_fixed_block_size(100);
1412 test_data.into_iter().for_each(|v| builder.append_option(v));
1413 builder.finish()
1414 };
1415 assert_eq!(array1, array1.clone());
1416 assert_eq!(array2, array2.clone());
1417 assert_eq!(array1, array2);
1418 }
1419
1420 /// Integration tests for `inline_key_fast` covering:
1421 ///
1422 /// 1. Monotonic ordering across increasing lengths and lexical variations.
1423 /// 2. Cross-check against `GenericBinaryArray` comparison to ensure semantic equivalence.
1424 ///
1425 /// This also includes a specific test for the “bar” vs. “bar\0” case, demonstrating why
1426 /// the length field is required even when all inline bytes fit in 12 bytes.
1427 ///
1428 /// The test includes strings that verify correct byte order (prevent reversal bugs),
1429 /// and length-based tie-breaking in the composite key.
1430 ///
1431 /// The test confirms that `inline_key_fast` produces keys which sort consistently
1432 /// with the expected lexicographical order of the raw byte arrays.
1433 #[test]
1434 fn test_inline_key_fast_various_lengths_and_lexical() {
1435 /// Helper to create a raw u128 value representing an inline ByteView:
1436 /// - `length`: number of meaningful bytes (must be ≤ 12)
1437 /// - `data`: the actual inline data bytes
1438 ///
1439 /// The first 4 bytes encode length in little-endian,
1440 /// the following 12 bytes contain the inline string data (unpadded).
1441 fn make_raw_inline(length: u32, data: &[u8]) -> u128 {
1442 assert!(length as usize <= 12, "Inline length must be ≤ 12");
1443 assert!(
1444 data.len() == length as usize,
1445 "Data length must match `length`"
1446 );
1447
1448 let mut raw_bytes = [0u8; 16];
1449 raw_bytes[0..4].copy_from_slice(&length.to_le_bytes()); // length stored little-endian
1450 raw_bytes[4..(4 + data.len())].copy_from_slice(data); // inline data
1451 u128::from_le_bytes(raw_bytes)
1452 }
1453
1454 // Test inputs: various lengths and lexical orders,
1455 // plus special cases for byte order and length tie-breaking
1456 let test_inputs: Vec<&[u8]> = vec![
1457 b"a",
1458 b"aa",
1459 b"aaa",
1460 b"aab",
1461 b"abcd",
1462 b"abcde",
1463 b"abcdef",
1464 b"abcdefg",
1465 b"abcdefgh",
1466 b"abcdefghi",
1467 b"abcdefghij",
1468 b"abcdefghijk",
1469 b"abcdefghijkl",
1470 // Tests for byte-order reversal bug:
1471 // Without the fix, "backend one" would compare as "eno dnekcab",
1472 // causing incorrect sort order relative to "backend two".
1473 b"backend one",
1474 b"backend two",
1475 // Tests length-tiebreaker logic:
1476 // "bar" (3 bytes) and "bar\0" (4 bytes) have identical inline data,
1477 // so only the length differentiates their ordering.
1478 b"bar",
1479 b"bar\0",
1480 // Additional lexical and length tie-breaking cases with same prefix, in correct lex order:
1481 b"than12Byt",
1482 b"than12Bytes",
1483 b"than12Bytes\0",
1484 b"than12Bytesx",
1485 b"than12Bytex",
1486 b"than12Bytez",
1487 // Additional lexical tests
1488 b"xyy",
1489 b"xyz",
1490 b"xza",
1491 ];
1492
1493 // Create a GenericBinaryArray for cross-comparison of lex order
1494 let array: GenericBinaryArray<i32> =
1495 GenericBinaryArray::from(test_inputs.iter().map(|s| Some(*s)).collect::<Vec<_>>());
1496
1497 for i in 0..array.len() - 1 {
1498 let v1 = array.value(i);
1499 let v2 = array.value(i + 1);
1500
1501 // Assert the array's natural lexical ordering is correct
1502 assert!(v1 < v2, "Array compare failed: {v1:?} !< {v2:?}");
1503
1504 // Assert the keys produced by inline_key_fast reflect the same ordering
1505 let key1 = GenericByteViewArray::<BinaryViewType>::inline_key_fast(make_raw_inline(
1506 v1.len() as u32,
1507 v1,
1508 ));
1509 let key2 = GenericByteViewArray::<BinaryViewType>::inline_key_fast(make_raw_inline(
1510 v2.len() as u32,
1511 v2,
1512 ));
1513
1514 assert!(
1515 key1 < key2,
1516 "Key compare failed: key({v1:?})=0x{key1:032x} !< key({v2:?})=0x{key2:032x}",
1517 );
1518 }
1519 }
1520}