parquet 60.1.0

// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

use bytes::Bytes;

use crate::basic::{ConvertedType, Encoding, LogicalType, Type};
use crate::bloom_filter::Sbbf;
use crate::column::writer::{fallback_encoding, has_dictionary_support, update_max, update_min};
use crate::data_type::DataType;
use crate::data_type::private::ParquetValueType;
use crate::encodings::encoding::{DictEncoder, Encoder, get_encoder_with_properties};
use crate::errors::{ParquetError, Result};
use crate::file::properties::{EnabledStatistics, ResolvedColumnProperties, WriterProperties};
use crate::geospatial::accumulator::{GeoStatsAccumulator, try_new_geo_stats_accumulator};
use crate::geospatial::statistics::GeospatialStatistics;
use crate::schema::types::ColumnDescPtr;

/// A collection of [`ParquetValueType`] encoded by a [`ColumnValueEncoder`]
pub trait ColumnValues {
    /// The number of values in this collection
    fn len(&self) -> usize;
}

#[cfg(feature = "arrow")]
impl ColumnValues for dyn arrow_array::Array {
    fn len(&self) -> usize {
        arrow_array::Array::len(self)
    }
}

impl<T: ParquetValueType> ColumnValues for [T] {
    fn len(&self) -> usize {
        self.len()
    }
}

/// The encoded data for a dictionary page
pub struct DictionaryPage {
    pub buf: Bytes,
    pub num_values: usize,
    pub is_sorted: bool,
}

/// The encoded values for a data page, with optional statistics
pub struct DataPageValues<T> {
    pub buf: Bytes,
    pub num_values: usize,
    pub encoding: Encoding,
    pub min_value: Option<T>,
    pub max_value: Option<T>,
    pub nan_count: Option<u64>,
    pub variable_length_bytes: Option<i64>,
}

/// A generic encoder of [`ColumnValues`] to data and dictionary pages used by
/// [`super::GenericColumnWriter`]
pub trait ColumnValueEncoder {
    /// The underlying value type of [`Self::Values`]
    ///
    /// Note: this avoids needing to fully qualify `<Self::Values as ColumnValues>::T`
    type T: ParquetValueType;

    /// The values encoded by this encoder
    type Values: ColumnValues + ?Sized;

    /// Create a new [`ColumnValueEncoder`]
    ///
    /// `column_props` holds the settings in `props` that apply specifically to
    /// `descr`, already resolved by the caller.
    #[expect(
        private_interfaces,
        reason = "this trait is not nameable outside the crate"
    )]
    fn try_new(
        descr: &ColumnDescPtr,
        props: &WriterProperties,
        column_props: &ResolvedColumnProperties,
    ) -> Result<Self>
    where
        Self: Sized;

    /// Write the corresponding values to this [`ColumnValueEncoder`]
    fn write(&mut self, values: &Self::Values, offset: usize, len: usize) -> Result<()>;

    /// Write the values at the indexes in `indices` to this [`ColumnValueEncoder`]
    fn write_gather(&mut self, values: &Self::Values, indices: &[usize]) -> Result<()>;

    /// Returns the largest `k` such that the first `k` values in
    /// `values[offset..offset + len]` encode to at most `byte_budget`
    /// bytes — i.e. how many values fit in a single page byte budget.
    ///
    /// Returns `len` if every value fits. Returns at least 1 if a single
    /// value alone exceeds the budget, matching parquet's "at least one
    /// value per data page" rule.
    ///
    /// `None` means "no cheap estimate available"; the caller stays on
    /// the batched fast path and lets the post-write
    /// `should_add_data_page` check handle bounding.
    ///
    /// Implementations should short-circuit aggressively: the typical
    /// case is "everything fits, return `len`", and the next-most-common
    /// case is "one wide value, return 1." The variable-width walk only
    /// needs to be precise when the chunk is genuinely near the budget.
    fn count_values_within_byte_budget(
        _values: &Self::Values,
        _offset: usize,
        _len: usize,
        _byte_budget: usize,
    ) -> Option<usize> {
        None
    }

    /// As [`Self::count_values_within_byte_budget`] but using gather
    /// `indices` rather than a contiguous range. Returns the number of
    /// `indices` that fit, not the maximum index value.
    fn count_values_within_byte_budget_gather(
        _values: &Self::Values,
        _indices: &[usize],
        _byte_budget: usize,
    ) -> Option<usize> {
        None
    }

    /// Returns the number of buffered values
    fn num_values(&self) -> usize;

    /// Returns true if this encoder has a dictionary page
    fn has_dictionary(&self) -> bool;

    /// Returns true if the encoder compresses each value against the value
    /// immediately before it, within the current page.
    ///
    /// For such encodings a page boundary is not free: flushing discards the
    /// previous value, so the first value of the next page is stored in full.
    /// [`GenericColumnWriter::should_add_data_page`] uses this to exempt a
    /// page's mandatory first value from the data page byte limit.
    ///
    /// Per encoding:
    /// * `DELTA_BYTE_ARRAY`: true. Each value is stored as the length of the
    ///   prefix it shares with its predecessor plus the remaining suffix.
    /// * Everything else: false, the default. `PLAIN` and
    ///   `DELTA_LENGTH_BYTE_ARRAY` store a value at the same cost wherever it
    ///   lands, and a dictionary outlives the pages that index into it, so no
    ///   page boundary makes a value more expensive.
    ///
    /// [`GenericColumnWriter::should_add_data_page`]: crate::column::writer::GenericColumnWriter::should_add_data_page
    fn compresses_against_previous_value(&self) -> bool {
        false
    }

    /// Returns the estimated total memory usage of the encoder
    ///
    fn estimated_memory_size(&self) -> usize;

    /// Returns an estimate of the encoded size of dictionary page size in bytes, or `None` if no dictionary
    fn estimated_dict_page_size(&self) -> Option<usize>;

    /// Returns an estimate of the encoded data page size in bytes
    ///
    /// This should include:
    /// <already_written_encoded_byte_size> + <estimated_encoded_size_of_unflushed_bytes>
    fn estimated_data_page_size(&self) -> usize;

    /// Flush the dictionary page for this column chunk if any. Any subsequent calls to
    /// [`Self::write`] will not be dictionary encoded
    ///
    /// Note: [`Self::flush_data_page`] must be called first, as this will error if there
    /// are any pending page values
    fn flush_dict_page(&mut self) -> Result<Option<DictionaryPage>>;

    /// Flush the next data page for this column chunk
    fn flush_data_page(&mut self) -> Result<DataPageValues<Self::T>>;

    /// Flushes bloom filter if enabled and returns it, otherwise returns `None`. Subsequent writes
    /// will *not* be tracked by the bloom filter as it is empty since. This should be called once
    /// near the end of encoding.
    fn flush_bloom_filter(&mut self) -> Option<Sbbf>;

    /// Computes [`GeospatialStatistics`], if any, and resets internal state such that any internal
    /// accumulator is prepared to accumulate statistics for the next column chunk.
    fn flush_geospatial_statistics(&mut self) -> Option<Box<GeospatialStatistics>>;
}

pub struct ColumnValueEncoderImpl<T: DataType> {
    encoder: Box<dyn Encoder<T>>,
    dict_encoder: Option<DictEncoder<T>>,
    descr: ColumnDescPtr,
    num_values: usize,
    statistics_enabled: EnabledStatistics,
    min_value: Option<T::T>,
    max_value: Option<T::T>,
    nan_count: Option<u64>,
    bloom_filter: Option<Sbbf>,
    bloom_filter_target_fpp: f64,
    variable_length_bytes: Option<i64>,
    geo_stats_accumulator: Option<Box<dyn GeoStatsAccumulator>>,
}

impl<T: DataType> ColumnValueEncoderImpl<T> {
    fn is_floating_point_column(&self) -> bool {
        matches!(self.descr.physical_type(), Type::FLOAT | Type::DOUBLE)
            || self.descr.logical_type_ref() == Some(&LogicalType::Float16)
    }

    fn write_slice(&mut self, slice: &[T::T]) -> Result<()> {
        if self.statistics_enabled != EnabledStatistics::None
            // INTERVAL, Geometry, and Geography have undefined sort order, so don't write min/max stats for them
            && self.descr.converted_type() != ConvertedType::INTERVAL
        {
            if let Some(accumulator) = self.geo_stats_accumulator.as_deref_mut() {
                update_geo_stats_accumulator(accumulator, slice.iter());
            } else if let Some((min, max, nan_count)) =
                T::T::min_max(self.descr.get_basic_info(), slice)
            {
                update_min(&self.descr, &min, &mut self.min_value);
                update_max(&self.descr, &max, &mut self.max_value);
                if self.is_floating_point_column() {
                    *self.nan_count.get_or_insert(0) += nan_count;
                }
            }

            if let Some(var_bytes) = T::T::variable_length_bytes(slice) {
                *self.variable_length_bytes.get_or_insert(0) += var_bytes;
            }
        }

        // While a dictionary is in use the filter is populated from its distinct values
        // in `flush_dict_page`, so each value is hashed once rather than once per row.
        match &mut self.dict_encoder {
            Some(encoder) => encoder.put(slice),
            _ => {
                if let Some(bloom_filter) = &mut self.bloom_filter {
                    for value in slice {
                        bloom_filter.insert(value);
                    }
                }
                self.encoder.put(slice)
            }
        }
    }
}

impl<T: DataType> ColumnValueEncoder for ColumnValueEncoderImpl<T> {
    type T = T::T;

    type Values = [T::T];

    fn flush_bloom_filter(&mut self) -> Option<Sbbf> {
        let mut sbbf = self.bloom_filter.take()?;
        sbbf.fold_to_target_fpp(self.bloom_filter_target_fpp);
        Some(sbbf)
    }

    #[expect(
        private_interfaces,
        reason = "this trait is not nameable outside the crate"
    )]
    fn try_new(
        descr: &ColumnDescPtr,
        props: &WriterProperties,
        column_props: &ResolvedColumnProperties,
    ) -> Result<Self> {
        let dict_supported =
            column_props.dictionary_enabled && has_dictionary_support(T::get_physical_type());
        let dict_encoder = dict_supported.then(|| DictEncoder::new(descr.clone()));

        // Set either main encoder or fallback encoder.
        let encoder = get_encoder_with_properties(
            column_props
                .encoding
                .unwrap_or_else(|| fallback_encoding(T::get_physical_type(), props)),
            descr,
            column_props,
        )?;

        let statistics_enabled = column_props.statistics_enabled;

        let (bloom_filter, bloom_filter_target_fpp) = create_bloom_filter(column_props)?;

        let geo_stats_accumulator = try_new_geo_stats_accumulator(descr);

        Ok(Self {
            encoder,
            dict_encoder,
            descr: descr.clone(),
            num_values: 0,
            statistics_enabled,
            bloom_filter,
            bloom_filter_target_fpp,
            min_value: None,
            max_value: None,
            nan_count: None,
            variable_length_bytes: None,
            geo_stats_accumulator,
        })
    }

    fn write(&mut self, values: &[T::T], offset: usize, len: usize) -> Result<()> {
        self.num_values += len;

        let slice = values.get(offset..offset + len).ok_or_else(|| {
            general_err!(
                "Expected to write {} values, but have only {}",
                len,
                values.len() - offset
            )
        })?;

        self.write_slice(slice)
    }

    fn write_gather(&mut self, values: &Self::Values, indices: &[usize]) -> Result<()> {
        self.num_values += indices.len();
        let slice: Vec<_> = indices.iter().map(|idx| values[*idx].clone()).collect();
        self.write_slice(&slice)
    }

    fn count_values_within_byte_budget(
        values: &[T::T],
        offset: usize,
        len: usize,
        byte_budget: usize,
    ) -> Option<usize> {
        // Clamp so that a caller-supplied `len` that overruns the input
        // (e.g. a level/value mismatch the encoder will reject later)
        // returns an estimate instead of panicking here.
        let end = (offset + len).min(values.len());
        let start = offset.min(end);
        count_within_budget::<T>(
            end - start,
            byte_budget,
            values[start..end].iter().map(Some),
        )
    }

    fn count_values_within_byte_budget_gather(
        values: &[T::T],
        indices: &[usize],
        byte_budget: usize,
    ) -> Option<usize> {
        // `values.get` yields `None` for an out-of-range index (defensive
        // against a level/value mismatch the encoder rejects later); such a
        // position is counted but contributes no bytes.
        count_within_budget::<T>(
            indices.len(),
            byte_budget,
            indices.iter().map(|&i| values.get(i)),
        )
    }

    fn num_values(&self) -> usize {
        self.num_values
    }

    fn has_dictionary(&self) -> bool {
        self.dict_encoder.is_some()
    }

    fn compresses_against_previous_value(&self) -> bool {
        // While dictionary encoding is active `self.encoder` is unused: the
        // data page holds RLE indices, which carry no cross-value state.
        self.dict_encoder.is_none() && self.encoder.encoding() == Encoding::DELTA_BYTE_ARRAY
    }

    fn estimated_memory_size(&self) -> usize {
        let encoder_size = self.encoder.estimated_memory_size();

        let dict_encoder_size = self
            .dict_encoder
            .as_ref()
            .map(|encoder| encoder.estimated_memory_size())
            .unwrap_or_default();

        let bloom_filter_size = self
            .bloom_filter
            .as_ref()
            .map(|bf| bf.estimated_memory_size())
            .unwrap_or_default();

        encoder_size + dict_encoder_size + bloom_filter_size
    }

    fn estimated_dict_page_size(&self) -> Option<usize> {
        Some(self.dict_encoder.as_ref()?.dict_encoded_size())
    }

    fn estimated_data_page_size(&self) -> usize {
        match &self.dict_encoder {
            Some(encoder) => encoder.estimated_data_encoded_size(),
            _ => self.encoder.estimated_data_encoded_size(),
        }
    }

    fn flush_dict_page(&mut self) -> Result<Option<DictionaryPage>> {
        match self.dict_encoder.take() {
            Some(encoder) => {
                if self.num_values != 0 {
                    return Err(general_err!(
                        "Must flush data pages before flushing dictionary"
                    ));
                }

                if let Some(bloom_filter) = &mut self.bloom_filter {
                    for value in encoder.uniques() {
                        bloom_filter.insert(value);
                    }
                }

                let buf = encoder.write_dict()?;

                Ok(Some(DictionaryPage {
                    buf,
                    num_values: encoder.num_entries(),
                    is_sorted: encoder.is_sorted(),
                }))
            }
            _ => Ok(None),
        }
    }

    fn flush_data_page(&mut self) -> Result<DataPageValues<T::T>> {
        let (buf, encoding) = match &mut self.dict_encoder {
            Some(encoder) => (encoder.write_indices()?, Encoding::RLE_DICTIONARY),
            _ => (self.encoder.flush_buffer()?, self.encoder.encoding()),
        };

        Ok(DataPageValues {
            buf,
            encoding,
            num_values: std::mem::take(&mut self.num_values),
            min_value: self.min_value.take(),
            max_value: self.max_value.take(),
            nan_count: self.nan_count.take(),
            variable_length_bytes: self.variable_length_bytes.take(),
        })
    }

    fn flush_geospatial_statistics(&mut self) -> Option<Box<GeospatialStatistics>> {
        self.geo_stats_accumulator.as_mut().map(|a| a.finish())?
    }
}

/// Creates a bloom filter sized for the column's configured NDV, returning the filter
/// and the target FPP for folding.
pub(crate) fn create_bloom_filter(
    column_props: &ResolvedColumnProperties,
) -> Result<(Option<Sbbf>, f64)> {
    match column_props.bloom_filter_properties.as_ref() {
        Some(bf_props) => Ok((
            Some(Sbbf::new_with_ndv_fpp(bf_props.ndv(), bf_props.fpp())?),
            bf_props.fpp(),
        )),
        None => Ok((None, 0.0)),
    }
}

fn update_geo_stats_accumulator<'a, T, I>(bounder: &mut dyn GeoStatsAccumulator, iter: I)
where
    T: ParquetValueType + 'a,
    I: Iterator<Item = &'a T>,
{
    if bounder.is_valid() {
        for val in iter {
            bounder.update_wkb(val.as_bytes());
        }
    }
}

/// Plain-encoded byte cost of a single value of type `T::T`.
///
/// Derived from [`ParquetValueType::dict_encoding_size`] (which returns
/// `(per-value overhead, value-bytes)`) so we don't add a parallel
/// per-value-size hook to the trait. Mirrors the dispatch in
/// `KeyStorage::push` (`encodings/encoding/dict_encoder.rs`).
///
/// Placed at the end of the module deliberately. Inserting it above the
/// `ColumnValueEncoder` trait shifts the trait and `ColumnValueEncoderImpl`
/// within the compiled module enough to perturb downstream code placement,
/// which measurably regresses unrelated arrow-writer string benchmarks
/// (~5-9% on `string` / `string_and_binary_view`). Defining it last keeps
/// the hot encoder code at the offsets it has on `main`.
#[inline]
fn plain_encoded_byte_size<T: DataType>(value: &T::T) -> usize {
    let (overhead, bytes) = value.dict_encoding_size();
    match <T::T as ParquetValueType>::PHYSICAL_TYPE {
        // Plain BYTE_ARRAY = 4-byte length prefix + payload.
        Type::BYTE_ARRAY => overhead + bytes,
        // Plain FLBA = raw bytes only; `dict_encoding_size`'s length prefix
        // is irrelevant here, so the encoder passes `type_length` directly.
        Type::FIXED_LEN_BYTE_ARRAY => bytes,
        // Numeric/bool are short-circuited by the caller via
        // `mem::size_of`, so this is unreachable in practice; fall back to
        // `overhead` defensively.
        _ => overhead,
    }
}

/// How many leading values fit in `byte_budget` bytes, shared by the two
/// `ColumnValueEncoder::count_values_within_byte_budget*` methods (one walks a
/// contiguous slice, the other gathers by index).
///
/// `n` is the answer when everything fits; `vals` yields each candidate value,
/// or `None` for a position that should still be counted but contributes no
/// bytes (an out-of-range gather index). The boundary value that crosses the
/// budget is included in the count so the caller's page-flush check trips on
/// this mini-batch rather than leaving a sliver for the next page; this also
/// catches a lone outlier wherever it lands among small values.
///
/// Defined at the end of the module alongside `plain_encoded_byte_size` for
/// the same reason — see that function's note on code placement and the
/// `string` / `string_and_binary_view` benchmarks.
#[inline]
fn count_within_budget<'a, T: DataType>(
    n: usize,
    byte_budget: usize,
    vals: impl Iterator<Item = Option<&'a T::T>>,
) -> Option<usize>
where
    T::T: 'a,
{
    // Fixed-size physical types have a constant per-value byte cost, so the
    // answer is one division — no walk needed.
    let phys = <T::T as ParquetValueType>::PHYSICAL_TYPE;
    if phys != Type::BYTE_ARRAY && phys != Type::FIXED_LEN_BYTE_ARRAY {
        let per = std::mem::size_of::<T::T>().max(1);
        return Some((byte_budget / per).max(1).min(n));
    }
    // Variable-width: accumulate, exit at the first value past the budget.
    let mut cum: usize = 0;
    for (i, v) in vals.enumerate() {
        if let Some(v) = v {
            cum = cum.saturating_add(plain_encoded_byte_size::<T>(v));
        }
        if cum > byte_budget {
            return Some(i + 1);
        }
    }
    Some(n)
}