Jimexist commented on code in PR #3057: URL: https://github.com/apache/arrow-rs/pull/3057#discussion_r1020827896
########## parquet/src/bloom_filter/mod.rs: ########## @@ -0,0 +1,217 @@ +// 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. + +//! Bloom filter implementation specific to Parquet, as described +//! in the [spec](https://github.com/apache/parquet-format/blob/master/BloomFilter.md) + +use crate::errors::ParquetError; +use crate::file::metadata::ColumnChunkMetaData; +use crate::format::{ + BloomFilterAlgorithm, BloomFilterCompression, BloomFilterHash, BloomFilterHeader, +}; +use std::hash::Hasher; +use std::io::{Read, Seek, SeekFrom}; +use thrift::protocol::TCompactInputProtocol; +use twox_hash::XxHash64; + +/// Salt as defined in the [spec](https://github.com/apache/parquet-format/blob/master/BloomFilter.md#technical-approach) +const SALT: [u32; 8] = [ + 0x47b6137b_u32, + 0x44974d91_u32, + 0x8824ad5b_u32, + 0xa2b7289d_u32, + 0x705495c7_u32, + 0x2df1424b_u32, + 0x9efc4947_u32, + 0x5c6bfb31_u32, +]; + +/// Each block is 256 bits, broken up into eight contiguous "words", each consisting of 32 bits. +/// Each word is thought of as an array of bits; each bit is either "set" or "not set". +type Block = [u32; 8]; + +/// takes as its argument a single unsigned 32-bit integer and returns a block in which each +/// word has exactly one bit set. +fn mask(x: u32) -> Block { + let mut result = [0_u32; 8]; + for i in 0..8 { + // wrapping instead of checking for overflow + let y = x.wrapping_mul(SALT[i]); + let y = y >> 27; + result[i] = 1 << y; + } + result +} + +/// setting every bit in the block that was also set in the result from mask +fn block_insert(block: &mut Block, hash: u32) { + let mask = mask(hash); + for i in 0..8 { + block[i] |= mask[i]; + } +} + +/// returns true when every bit that is set in the result of mask is also set in the block. +fn block_check(block: &Block, hash: u32) -> bool { + let mask = mask(hash); + for i in 0..8 { + if block[i] & mask[i] == 0 { + return false; + } + } + true +} + +/// A split block Bloom filter +pub struct Sbbf(Vec<Block>); + +impl Sbbf { + fn new(bitset: &[u8]) -> Self { + let data = bitset + .chunks_exact(4 * 8) + .map(|chunk| { + let mut block = [0_u32; 8]; + for (i, word) in chunk.chunks_exact(4).enumerate() { + block[i] = u32::from_le_bytes(word.try_into().unwrap()); + } + block + }) + .collect::<Vec<Block>>(); + Self(data) + } + + pub fn read_from_column_chunk<R: Read + Seek>( + column_metadata: &ColumnChunkMetaData, + mut reader: &mut R, + ) -> Result<Self, ParquetError> { + let offset = column_metadata.bloom_filter_offset().ok_or_else(|| { + ParquetError::General("Bloom filter offset is not set".to_string()) + })? as u64; + reader.seek(SeekFrom::Start(offset))?; + // deserialize header + let mut prot = TCompactInputProtocol::new(&mut reader); + let header = BloomFilterHeader::read_from_in_protocol(&mut prot)?; + + match header.algorithm { + BloomFilterAlgorithm::BLOCK(_) => { + // this match exists to future proof the singleton algorithm enum + } + } + match header.compression { + BloomFilterCompression::UNCOMPRESSED(_) => { + // this match exists to future proof the singleton compression enum + } + } + match header.hash { + BloomFilterHash::XXHASH(_) => { + // this match exists to future proof the singleton hash enum + } + } + // length in bytes + let length: usize = header.num_bytes.try_into().map_err(|_| { + ParquetError::General("Bloom filter length is invalid".to_string()) + })?; + let mut buffer = vec![0_u8; length]; + reader.read_exact(&mut buffer).map_err(|e| { + ParquetError::General(format!("Could not read bloom filter: {}", e)) + })?; + Ok(Self::new(&buffer)) + } + + #[inline] + fn hash_to_block_index(&self, hash: u64) -> usize { + // unchecked_mul is unstable, but in reality this is safe, we'd just use saturating mul + // but it will not saturate + (((hash >> 32).saturating_mul(self.0.len() as u64)) >> 32) as usize Review Comment: yes this is per [spec](https://github.com/apache/parquet-format/blob/master/BloomFilter.md) > The filter_insert operation first uses the most significant 32 bits of its argument to select a block to operate on. Call the argument "h", and recall the use of "z" to mean the number of blocks. Then a block number i between 0 and z-1 (inclusive) to operate on is chosen as follows: ```rust unsigned int64 h_top_bits = h >> 32; unsigned int64 z_as_64_bit = z; unsigned int32 i = (h_top_bits * z_as_64_bit) >> 32; ``` > The first line extracts the most significant 32 bits from h and assignes them to a 64-bit unsigned integer. The second line is simpler: it just sets an unsigned 64-bit value to the same value as the 32-bit unsigned value z. The purpose of having both h_top_bits and z_as_64_bit be 64-bit values is so that their product is a 64-bit value. That product is taken in the third line, and then the most significant 32 bits are extracted into the value i, which is the index of the block that will be operated on. > After this process to select i, filter_insert uses the least significant 32 bits of h as the argument to block_insert called on block i. > The technique for converting the most significant 32 bits to an integer between 0 and z-1 (inclusive) avoids using the modulo operation, which is often very slow. This trick can be found in [Kenneth A. Ross's 2006 IBM research report, "Efficient Hash Probes on Modern Processors"](https://domino.research.ibm.com/library/cyberdig.nsf/papers/DF54E3545C82E8A585257222006FD9A2/$File/rc24100.pdf) -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
