parquet/column/writer/
encoder.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 bytes::Bytes;
19
20use crate::basic::{ConvertedType, Encoding, LogicalType, Type};
21use crate::bloom_filter::Sbbf;
22use crate::column::writer::{
23    compare_greater, fallback_encoding, has_dictionary_support, is_nan, update_max, update_min,
24};
25use crate::data_type::DataType;
26use crate::data_type::private::ParquetValueType;
27use crate::encodings::encoding::{DictEncoder, Encoder, get_encoder};
28use crate::errors::{ParquetError, Result};
29use crate::file::properties::{EnabledStatistics, WriterProperties};
30use crate::geospatial::accumulator::{GeoStatsAccumulator, try_new_geo_stats_accumulator};
31use crate::geospatial::statistics::GeospatialStatistics;
32use crate::schema::types::{BasicTypeInfo, ColumnDescPtr};
33
34/// A collection of [`ParquetValueType`] encoded by a [`ColumnValueEncoder`]
35pub trait ColumnValues {
36    /// The number of values in this collection
37    fn len(&self) -> usize;
38}
39
40#[cfg(feature = "arrow")]
41impl ColumnValues for dyn arrow_array::Array {
42    fn len(&self) -> usize {
43        arrow_array::Array::len(self)
44    }
45}
46
47impl<T: ParquetValueType> ColumnValues for [T] {
48    fn len(&self) -> usize {
49        self.len()
50    }
51}
52
53/// The encoded data for a dictionary page
54pub struct DictionaryPage {
55    pub buf: Bytes,
56    pub num_values: usize,
57    pub is_sorted: bool,
58}
59
60/// The encoded values for a data page, with optional statistics
61pub struct DataPageValues<T> {
62    pub buf: Bytes,
63    pub num_values: usize,
64    pub encoding: Encoding,
65    pub min_value: Option<T>,
66    pub max_value: Option<T>,
67    pub nan_count: Option<u64>,
68    pub variable_length_bytes: Option<i64>,
69}
70
71/// A generic encoder of [`ColumnValues`] to data and dictionary pages used by
72/// [`super::GenericColumnWriter`]
73pub trait ColumnValueEncoder {
74    /// The underlying value type of [`Self::Values`]
75    ///
76    /// Note: this avoids needing to fully qualify `<Self::Values as ColumnValues>::T`
77    type T: ParquetValueType;
78
79    /// The values encoded by this encoder
80    type Values: ColumnValues + ?Sized;
81
82    /// Create a new [`ColumnValueEncoder`]
83    fn try_new(descr: &ColumnDescPtr, props: &WriterProperties) -> Result<Self>
84    where
85        Self: Sized;
86
87    /// Write the corresponding values to this [`ColumnValueEncoder`]
88    fn write(&mut self, values: &Self::Values, offset: usize, len: usize) -> Result<()>;
89
90    /// Write the values at the indexes in `indices` to this [`ColumnValueEncoder`]
91    fn write_gather(&mut self, values: &Self::Values, indices: &[usize]) -> Result<()>;
92
93    /// Returns the largest `k` such that the first `k` values in
94    /// `values[offset..offset + len]` encode to at most `byte_budget`
95    /// bytes — i.e. how many values fit in a single page byte budget.
96    ///
97    /// Returns `len` if every value fits. Returns at least 1 if a single
98    /// value alone exceeds the budget, matching parquet's "at least one
99    /// value per data page" rule.
100    ///
101    /// `None` means "no cheap estimate available"; the caller stays on
102    /// the batched fast path and lets the post-write
103    /// `should_add_data_page` check handle bounding.
104    ///
105    /// Implementations should short-circuit aggressively: the typical
106    /// case is "everything fits, return `len`", and the next-most-common
107    /// case is "one wide value, return 1." The variable-width walk only
108    /// needs to be precise when the chunk is genuinely near the budget.
109    fn count_values_within_byte_budget(
110        _values: &Self::Values,
111        _offset: usize,
112        _len: usize,
113        _byte_budget: usize,
114    ) -> Option<usize> {
115        None
116    }
117
118    /// As [`Self::count_values_within_byte_budget`] but using gather
119    /// `indices` rather than a contiguous range. Returns the number of
120    /// `indices` that fit, not the maximum index value.
121    fn count_values_within_byte_budget_gather(
122        _values: &Self::Values,
123        _indices: &[usize],
124        _byte_budget: usize,
125    ) -> Option<usize> {
126        None
127    }
128
129    /// Returns the number of buffered values
130    fn num_values(&self) -> usize;
131
132    /// Returns true if this encoder has a dictionary page
133    fn has_dictionary(&self) -> bool;
134
135    /// Returns the estimated total memory usage of the encoder
136    ///
137    fn estimated_memory_size(&self) -> usize;
138
139    /// Returns an estimate of the encoded size of dictionary page size in bytes, or `None` if no dictionary
140    fn estimated_dict_page_size(&self) -> Option<usize>;
141
142    /// Returns an estimate of the encoded data page size in bytes
143    ///
144    /// This should include:
145    /// <already_written_encoded_byte_size> + <estimated_encoded_size_of_unflushed_bytes>
146    fn estimated_data_page_size(&self) -> usize;
147
148    /// Flush the dictionary page for this column chunk if any. Any subsequent calls to
149    /// [`Self::write`] will not be dictionary encoded
150    ///
151    /// Note: [`Self::flush_data_page`] must be called first, as this will error if there
152    /// are any pending page values
153    fn flush_dict_page(&mut self) -> Result<Option<DictionaryPage>>;
154
155    /// Flush the next data page for this column chunk
156    fn flush_data_page(&mut self) -> Result<DataPageValues<Self::T>>;
157
158    /// Flushes bloom filter if enabled and returns it, otherwise returns `None`. Subsequent writes
159    /// will *not* be tracked by the bloom filter as it is empty since. This should be called once
160    /// near the end of encoding.
161    fn flush_bloom_filter(&mut self) -> Option<Sbbf>;
162
163    /// Computes [`GeospatialStatistics`], if any, and resets internal state such that any internal
164    /// accumulator is prepared to accumulate statistics for the next column chunk.
165    fn flush_geospatial_statistics(&mut self) -> Option<Box<GeospatialStatistics>>;
166}
167
168pub struct ColumnValueEncoderImpl<T: DataType> {
169    encoder: Box<dyn Encoder<T>>,
170    dict_encoder: Option<DictEncoder<T>>,
171    descr: ColumnDescPtr,
172    num_values: usize,
173    statistics_enabled: EnabledStatistics,
174    min_value: Option<T::T>,
175    max_value: Option<T::T>,
176    nan_count: Option<u64>,
177    bloom_filter: Option<Sbbf>,
178    bloom_filter_target_fpp: f64,
179    variable_length_bytes: Option<i64>,
180    geo_stats_accumulator: Option<Box<dyn GeoStatsAccumulator>>,
181}
182
183impl<T: DataType> ColumnValueEncoderImpl<T> {
184    fn is_floating_point_column(&self) -> bool {
185        matches!(self.descr.physical_type(), Type::FLOAT | Type::DOUBLE)
186            || self.descr.logical_type_ref() == Some(&LogicalType::Float16)
187    }
188
189    fn write_slice(&mut self, slice: &[T::T]) -> Result<()> {
190        if self.statistics_enabled != EnabledStatistics::None
191            // INTERVAL, Geometry, and Geography have undefined sort order, so don't write min/max stats for them
192            && self.descr.converted_type() != ConvertedType::INTERVAL
193        {
194            if let Some(accumulator) = self.geo_stats_accumulator.as_deref_mut() {
195                update_geo_stats_accumulator(accumulator, slice.iter());
196            } else if let Some((min, max, nan_count)) =
197                get_min_max(self.descr.get_basic_info(), slice.iter())
198            {
199                update_min(&self.descr, &min, &mut self.min_value);
200                update_max(&self.descr, &max, &mut self.max_value);
201                if self.is_floating_point_column() {
202                    *self.nan_count.get_or_insert(0) += nan_count;
203                }
204            }
205
206            if let Some(var_bytes) = T::T::variable_length_bytes(slice) {
207                *self.variable_length_bytes.get_or_insert(0) += var_bytes;
208            }
209        }
210
211        // encode the values into bloom filter if enabled
212        if let Some(bloom_filter) = &mut self.bloom_filter {
213            for value in slice {
214                bloom_filter.insert(value);
215            }
216        }
217
218        match &mut self.dict_encoder {
219            Some(encoder) => encoder.put(slice),
220            _ => self.encoder.put(slice),
221        }
222    }
223}
224
225impl<T: DataType> ColumnValueEncoder for ColumnValueEncoderImpl<T> {
226    type T = T::T;
227
228    type Values = [T::T];
229
230    fn flush_bloom_filter(&mut self) -> Option<Sbbf> {
231        let mut sbbf = self.bloom_filter.take()?;
232        sbbf.fold_to_target_fpp(self.bloom_filter_target_fpp);
233        Some(sbbf)
234    }
235
236    fn try_new(descr: &ColumnDescPtr, props: &WriterProperties) -> Result<Self> {
237        let dict_supported = props.dictionary_enabled(descr.path())
238            && has_dictionary_support(T::get_physical_type(), props);
239        let dict_encoder = dict_supported.then(|| DictEncoder::new(descr.clone()));
240
241        // Set either main encoder or fallback encoder.
242        let encoder = get_encoder(
243            props
244                .encoding(descr.path())
245                .unwrap_or_else(|| fallback_encoding(T::get_physical_type(), props)),
246            descr,
247        )?;
248
249        let statistics_enabled = props.statistics_enabled(descr.path());
250
251        let (bloom_filter, bloom_filter_target_fpp) = create_bloom_filter(props, descr)?;
252
253        let geo_stats_accumulator = try_new_geo_stats_accumulator(descr);
254
255        Ok(Self {
256            encoder,
257            dict_encoder,
258            descr: descr.clone(),
259            num_values: 0,
260            statistics_enabled,
261            bloom_filter,
262            bloom_filter_target_fpp,
263            min_value: None,
264            max_value: None,
265            nan_count: None,
266            variable_length_bytes: None,
267            geo_stats_accumulator,
268        })
269    }
270
271    fn write(&mut self, values: &[T::T], offset: usize, len: usize) -> Result<()> {
272        self.num_values += len;
273
274        let slice = values.get(offset..offset + len).ok_or_else(|| {
275            general_err!(
276                "Expected to write {} values, but have only {}",
277                len,
278                values.len() - offset
279            )
280        })?;
281
282        self.write_slice(slice)
283    }
284
285    fn write_gather(&mut self, values: &Self::Values, indices: &[usize]) -> Result<()> {
286        self.num_values += indices.len();
287        let slice: Vec<_> = indices.iter().map(|idx| values[*idx].clone()).collect();
288        self.write_slice(&slice)
289    }
290
291    fn count_values_within_byte_budget(
292        values: &[T::T],
293        offset: usize,
294        len: usize,
295        byte_budget: usize,
296    ) -> Option<usize> {
297        // Clamp so that a caller-supplied `len` that overruns the input
298        // (e.g. a level/value mismatch the encoder will reject later)
299        // returns an estimate instead of panicking here.
300        let end = (offset + len).min(values.len());
301        let start = offset.min(end);
302        count_within_budget::<T>(
303            end - start,
304            byte_budget,
305            values[start..end].iter().map(Some),
306        )
307    }
308
309    fn count_values_within_byte_budget_gather(
310        values: &[T::T],
311        indices: &[usize],
312        byte_budget: usize,
313    ) -> Option<usize> {
314        // `values.get` yields `None` for an out-of-range index (defensive
315        // against a level/value mismatch the encoder rejects later); such a
316        // position is counted but contributes no bytes.
317        count_within_budget::<T>(
318            indices.len(),
319            byte_budget,
320            indices.iter().map(|&i| values.get(i)),
321        )
322    }
323
324    fn num_values(&self) -> usize {
325        self.num_values
326    }
327
328    fn has_dictionary(&self) -> bool {
329        self.dict_encoder.is_some()
330    }
331
332    fn estimated_memory_size(&self) -> usize {
333        let encoder_size = self.encoder.estimated_memory_size();
334
335        let dict_encoder_size = self
336            .dict_encoder
337            .as_ref()
338            .map(|encoder| encoder.estimated_memory_size())
339            .unwrap_or_default();
340
341        let bloom_filter_size = self
342            .bloom_filter
343            .as_ref()
344            .map(|bf| bf.estimated_memory_size())
345            .unwrap_or_default();
346
347        encoder_size + dict_encoder_size + bloom_filter_size
348    }
349
350    fn estimated_dict_page_size(&self) -> Option<usize> {
351        Some(self.dict_encoder.as_ref()?.dict_encoded_size())
352    }
353
354    fn estimated_data_page_size(&self) -> usize {
355        match &self.dict_encoder {
356            Some(encoder) => encoder.estimated_data_encoded_size(),
357            _ => self.encoder.estimated_data_encoded_size(),
358        }
359    }
360
361    fn flush_dict_page(&mut self) -> Result<Option<DictionaryPage>> {
362        match self.dict_encoder.take() {
363            Some(encoder) => {
364                if self.num_values != 0 {
365                    return Err(general_err!(
366                        "Must flush data pages before flushing dictionary"
367                    ));
368                }
369
370                let buf = encoder.write_dict()?;
371
372                Ok(Some(DictionaryPage {
373                    buf,
374                    num_values: encoder.num_entries(),
375                    is_sorted: encoder.is_sorted(),
376                }))
377            }
378            _ => Ok(None),
379        }
380    }
381
382    fn flush_data_page(&mut self) -> Result<DataPageValues<T::T>> {
383        let (buf, encoding) = match &mut self.dict_encoder {
384            Some(encoder) => (encoder.write_indices()?, Encoding::RLE_DICTIONARY),
385            _ => (self.encoder.flush_buffer()?, self.encoder.encoding()),
386        };
387
388        Ok(DataPageValues {
389            buf,
390            encoding,
391            num_values: std::mem::take(&mut self.num_values),
392            min_value: self.min_value.take(),
393            max_value: self.max_value.take(),
394            nan_count: self.nan_count.take(),
395            variable_length_bytes: self.variable_length_bytes.take(),
396        })
397    }
398
399    fn flush_geospatial_statistics(&mut self) -> Option<Box<GeospatialStatistics>> {
400        self.geo_stats_accumulator.as_mut().map(|a| a.finish())?
401    }
402}
403
404// Get min and max values for all values in `iter`.
405//
406// For floating point we need to compare NaN values until we encounter a non-NaN
407// value which then becomes the new min/max. After this, only non-NaN values are
408// evaluated. If all values are NaN, then the min/max NaNs as determined by
409// IEEE 754 total order are returned.
410fn get_min_max<'a, T, I>(basic_type_info: &BasicTypeInfo, mut iter: I) -> Option<(T, T, u64)>
411where
412    T: ParquetValueType + 'a,
413    I: Iterator<Item = &'a T>,
414{
415    let first = iter.next()?;
416    let mut min_max_nan = is_nan(basic_type_info, first);
417    let mut nan_count = min_max_nan as u64;
418
419    let mut min = first;
420    let mut max = first;
421    for val in iter {
422        match (min_max_nan, is_nan(basic_type_info, val)) {
423            // skip NaNs if we've encounter non-NaN
424            (false, true) => {
425                nan_count += 1;
426                continue;
427            }
428            // if min/max are NaN, check for non-NaN and reset
429            (true, false) => {
430                min = val;
431                max = val;
432                min_max_nan = false;
433                continue;
434            }
435            // both are NaN or non-NaN, so do the comparison
436            (_, val_is_nan) => {
437                nan_count += val_is_nan as u64;
438                // we've already initialized min and max, so a single value can't be both
439                // extremes
440                if compare_greater(basic_type_info, min, val) {
441                    min = val;
442                } else if compare_greater(basic_type_info, val, max) {
443                    max = val;
444                }
445            }
446        }
447    }
448
449    Some((min.clone(), max.clone(), nan_count))
450}
451
452/// Creates a bloom filter sized for the column's configured NDV, returning the filter
453/// and the target FPP for folding.
454pub(crate) fn create_bloom_filter(
455    props: &WriterProperties,
456    descr: &ColumnDescPtr,
457) -> Result<(Option<Sbbf>, f64)> {
458    match props.bloom_filter_properties(descr.path()) {
459        Some(bf_props) => Ok((
460            Some(Sbbf::new_with_ndv_fpp(bf_props.ndv(), bf_props.fpp())?),
461            bf_props.fpp(),
462        )),
463        None => Ok((None, 0.0)),
464    }
465}
466
467fn update_geo_stats_accumulator<'a, T, I>(bounder: &mut dyn GeoStatsAccumulator, iter: I)
468where
469    T: ParquetValueType + 'a,
470    I: Iterator<Item = &'a T>,
471{
472    if bounder.is_valid() {
473        for val in iter {
474            bounder.update_wkb(val.as_bytes());
475        }
476    }
477}
478
479/// Plain-encoded byte cost of a single value of type `T::T`.
480///
481/// Derived from [`ParquetValueType::dict_encoding_size`] (which returns
482/// `(per-value overhead, value-bytes)`) so we don't add a parallel
483/// per-value-size hook to the trait. Mirrors the dispatch in
484/// `KeyStorage::push` (`encodings/encoding/dict_encoder.rs`).
485///
486/// Placed at the end of the module deliberately. Inserting it above the
487/// `ColumnValueEncoder` trait shifts the trait and `ColumnValueEncoderImpl`
488/// within the compiled module enough to perturb downstream code placement,
489/// which measurably regresses unrelated arrow-writer string benchmarks
490/// (~5-9% on `string` / `string_and_binary_view`). Defining it last keeps
491/// the hot encoder code at the offsets it has on `main`.
492#[inline]
493fn plain_encoded_byte_size<T: DataType>(value: &T::T) -> usize {
494    let (overhead, bytes) = value.dict_encoding_size();
495    match <T::T as ParquetValueType>::PHYSICAL_TYPE {
496        // Plain BYTE_ARRAY = 4-byte length prefix + payload.
497        Type::BYTE_ARRAY => overhead + bytes,
498        // Plain FLBA = raw bytes only; `dict_encoding_size`'s length prefix
499        // is irrelevant here, so the encoder passes `type_length` directly.
500        Type::FIXED_LEN_BYTE_ARRAY => bytes,
501        // Numeric/bool are short-circuited by the caller via
502        // `mem::size_of`, so this is unreachable in practice; fall back to
503        // `overhead` defensively.
504        _ => overhead,
505    }
506}
507
508/// How many leading values fit in `byte_budget` bytes, shared by the two
509/// `ColumnValueEncoder::count_values_within_byte_budget*` methods (one walks a
510/// contiguous slice, the other gathers by index).
511///
512/// `n` is the answer when everything fits; `vals` yields each candidate value,
513/// or `None` for a position that should still be counted but contributes no
514/// bytes (an out-of-range gather index). The boundary value that crosses the
515/// budget is included in the count so the caller's page-flush check trips on
516/// this mini-batch rather than leaving a sliver for the next page; this also
517/// catches a lone outlier wherever it lands among small values.
518///
519/// Defined at the end of the module alongside `plain_encoded_byte_size` for
520/// the same reason — see that function's note on code placement and the
521/// `string` / `string_and_binary_view` benchmarks.
522#[inline]
523fn count_within_budget<'a, T: DataType>(
524    n: usize,
525    byte_budget: usize,
526    vals: impl Iterator<Item = Option<&'a T::T>>,
527) -> Option<usize>
528where
529    T::T: 'a,
530{
531    // Fixed-size physical types have a constant per-value byte cost, so the
532    // answer is one division — no walk needed.
533    let phys = <T::T as ParquetValueType>::PHYSICAL_TYPE;
534    if phys != Type::BYTE_ARRAY && phys != Type::FIXED_LEN_BYTE_ARRAY {
535        let per = std::mem::size_of::<T::T>().max(1);
536        return Some((byte_budget / per).max(1).min(n));
537    }
538    // Variable-width: accumulate, exit at the first value past the budget.
539    let mut cum: usize = 0;
540    for (i, v) in vals.enumerate() {
541        if let Some(v) = v {
542            cum = cum.saturating_add(plain_encoded_byte_size::<T>(v));
543        }
544        if cum > byte_budget {
545            return Some(i + 1);
546        }
547    }
548    Some(n)
549}
parquet/column/writer/encoder.rs

parquet/column/writer/
encoder.rs
