sig_store.rs

/*
 *
 * SPDX-FileCopyrightText: 2023 Sebastiano Vigna
 *
 * SPDX-License-Identifier: Apache-2.0 OR LGPL-2.1-or-later
 */

/*!

Fast sorting and grouping of signatures and values.

A *signature* is a pair of 64-bit integers, and a *value* is a generic type
implementing [`epserde::traits::ZeroCopy`].

A [`SigStore`] acts as a builder for a [`ChunkStore`]: it
accepts signature/value pairs in any order, and when you call
[`SigStore::into_chunk_store`] it returns an immutable  [`ChunkStore`]
that can [iterate on chunks of pairs, where chunks are defined
by the highest bits of signatures](ChunkStore::iter).

The trait [`ToSig`] provides a standard way to generate signatures for a [`SigStore`].

*/

use anyhow::Result;
use epserde::prelude::*;
use mem_dbg::{MemDbg, MemSize};
use rdst::RadixKey;
use std::borrow::Cow;
use std::{collections::VecDeque, fs::File, io::*, marker::PhantomData};

use crate::prelude::spooky_short;

/**

A signature and a value.

*/

#[derive(Epserde, Debug, Clone, Copy, MemDbg, MemSize)]
#[repr(C)]
#[zero_copy]
pub struct SigVal<T: ZeroCopy + 'static> {
    pub sig: [u64; 2],
    pub val: T,
}

impl<T: ZeroCopy + 'static> RadixKey for SigVal<T> {
    const LEVELS: usize = 16;

    fn get_level(&self, level: usize) -> u8 {
        (self.sig[1 - level / 8] >> ((level % 8) * 8)) as u8
    }
}

/**

Trait for types that must be turned into a signature.

We provide implementations for all primitive types and strings
by turning them into slice of bytes and then hashing them with
[crate::utils::spooky::spooky_short], using the given seed.

*/

pub trait ToSig {
    fn to_sig(key: &Self, seed: u64) -> [u64; 2];
}

impl ToSig for String {
    fn to_sig(key: &Self, seed: u64) -> [u64; 2] {
        let spooky = spooky_short(key, seed);
        [spooky[0], spooky[1]]
    }
}

impl ToSig for &String {
    fn to_sig(key: &Self, seed: u64) -> [u64; 2] {
        let spooky = spooky_short(key, seed);
        [spooky[0], spooky[1]]
    }
}

impl ToSig for str {
    fn to_sig(key: &Self, seed: u64) -> [u64; 2] {
        let spooky = spooky_short(key, seed);
        [spooky[0], spooky[1]]
    }
}

impl ToSig for &str {
    fn to_sig(key: &Self, seed: u64) -> [u64; 2] {
        let spooky = spooky_short(key, seed);
        [spooky[0], spooky[1]]
    }
}

macro_rules! to_sig_prim {
    ($($ty:ty),*) => {$(
        impl ToSig for $ty {
            fn to_sig(key: &Self, seed: u64) -> [u64; 2] {
                let spooky = spooky_short(&key.to_ne_bytes(), seed);
                [spooky[0], spooky[1]]
            }
        }
    )*};
}

to_sig_prim!(isize, usize, i8, i16, i32, i64, i128, u8, u16, u32, u64, u128);

macro_rules! to_sig_slice {
    ($($ty:ty),*) => {$(
        impl ToSig for &[$ty] {
            fn to_sig(key: &Self, seed: u64) -> [u64; 2] {
                // Alignemnt to u8 never fails or leave trailing/leading bytes
                let spooky = spooky_short(unsafe {key.align_to::<u8>().1 }, seed);
                [spooky[0], spooky[1]]
            }
        }
    )*};
}

to_sig_slice!(isize, usize, i8, i16, i32, i64, i128, u8, u16, u32, u64, u128);

/**

Accumulates key signatures (i.e., random-looking
hashes associated to keys) and associated values,
grouping them in different disk buffers by the high bits of the hash.
Along the way, it keeps track of the number of signatures with the same
`max_chunk_high_bits` high bits.

The implementation exploits the fact that signatures are randomly distributed,
and thus bucket sorting is very effective: at construction time you specify
the number of high bits to use for bucket sorting (say, 8), and when you
[push](`SigStore::push`) keys they will be stored in different disk buffers
(in this case, 256) depending on their high bits. The buffers will be stored
in a directory created by [`tempfile::TempDir`].

After all key signatures and values have been accumulated, you must
call [`SigStore::into_chunk_store`] to flush the buffers and obtain a
[`ChunkStore`]. [`SigStore::into_chunk_store`] takes the the number of high bits
to use for grouping signatures into chunks, and the necessary buffer splitting or merging
will be handled automatically by the resulting [`ChunkStore`].

*/
#[derive(Debug)]
pub struct SigStore<T> {
    /// Number of keys added so far.
    len: usize,
    /// The number of high bits used for bucket sorting (i.e., the number of files).
    buckets_high_bits: u32,
    /// The maximum number of high bits used for defining chunks in the call to
    /// [`SigStore::into_chunk_store`]. Chunk sizes will be computed incrementally
    /// for chunks defined by this number of high bits.
    max_chunk_high_bits: u32,
    /// A mask for the lowest `buckets_high_bits` bits.
    buckets_mask: u64,
    // A mask for the lowest `max_chunk_high_bits` bits.
    max_chunk_mask: u64,
    /// The writers associated to the buckets.
    writers: VecDeque<BufWriter<File>>,
    /// The number of keys in each bucket.
    bucket_sizes: Vec<usize>,
    /// The number of keys with the same `max_chunk_high_bits` high bits.
    chunk_sizes: Vec<usize>,
    _marker: PhantomData<T>,
}

/**

An container for the signatures and values accumulated by a [`SigStore`], with
the ability to [enumerate them grouped in chunks](ChunkStore::iter).

*/
#[derive(Debug)]
pub struct ChunkStore<T> {
    /// The number of high bits used for bucket sorting.
    bucket_high_bits: u32,
    /// The number of high bits defining a chunk.
    chunk_high_bits: u32,
    /// The files associated to the buckets.
    files: Vec<File>,
    /// The number of keys in each bucket.
    buf_sizes: Vec<usize>,
    /// The number of keys in each chunk.
    chunk_sizes: Vec<usize>,
    _marker: PhantomData<T>,
}

impl<T: ZeroCopy + 'static> ChunkStore<T> {
    /// Return the chunk sizes.
    pub fn chunk_sizes(&self) -> &Vec<usize> {
        &self.chunk_sizes
    }

    /// Return an iterator on chunks.
    ///
    /// This method can be called multiple times.
    pub fn iter(&mut self) -> Result<ChunkIterator<'_, T>> {
        Ok(ChunkIterator {
            store: self,
            next_file: 0,
            next_chunk: 0,
            chunks: VecDeque::from(vec![]),
            _marker: PhantomData,
        })
    }
}

/**

Enumerate chunks in a [`ChunkStore`].

A [`ChunkIterator`] handles the mapping between buckets and chunks. If a chunk is made
by one or more buckets, it will aggregate them as necessary; if a bucket contains
several chunks, it will split the bucket into chunks. In all cases, each chunk
is sorted and tested for duplicates: if duplicates are detected, a fake pair
containing `usize::MAX` and an empty chunk will be returned.

Note that a [`ChunkIterator`] returns an owned variant of [`Cow`]. The reason for
using [`Cow`] is easier interoperability with in-memory construction methods, which
usually return borrowed variants.

*/
#[derive(Debug)]
pub struct ChunkIterator<'a, T: ZeroCopy + 'static> {
    store: &'a mut ChunkStore<T>,
    /// The next file to examine.
    next_file: usize,
    /// The index of the next chunk to return.
    next_chunk: usize,
    /// The remaining chunks to emit, if there are several chunks per bucket.
    chunks: VecDeque<Vec<SigVal<T>>>,
    _marker: PhantomData<T>,
}

impl<'a, T: ZeroCopy + Send + Sync + 'static> Iterator for ChunkIterator<'a, T> {
    type Item = (usize, Cow<'a, [SigVal<T>]>);

    fn next(&mut self) -> Option<Self::Item> {
        let store = &mut self.store;
        if store.bucket_high_bits >= store.chunk_high_bits {
            // We need to aggregate one or more buckets to get a chunk
            if self.next_file >= store.files.len() {
                return None;
            }

            let to_aggr = 1 << (store.bucket_high_bits - store.chunk_high_bits);

            let len = store.chunk_sizes[self.next_chunk];
            let mut chunk = Vec::<SigVal<T>>::with_capacity(len);

            // SAFETY: we just allocated this vector so it is safe to set the length.
            // read_exact guarantees that the vector will be filled with data.
            #[allow(clippy::uninit_vec)]
            unsafe {
                chunk.set_len(len);
            }

            {
                let (pre, mut buf, post) = unsafe { chunk.align_to_mut::<u8>() };
                assert!(pre.is_empty());
                assert!(post.is_empty());
                for i in self.next_file..self.next_file + to_aggr {
                    let mut reader = &store.files[i];
                    let bytes = store.buf_sizes[i] * core::mem::size_of::<SigVal<T>>();
                    reader.read_exact(&mut buf[..bytes]).unwrap();
                    buf = &mut buf[bytes..];
                }
            }

            let res = (self.next_chunk, Cow::Owned(chunk));
            self.next_file += to_aggr;
            self.next_chunk += 1;
            Some(res)
        } else {
            // We need to split buckets in several chunks
            if self.chunks.is_empty() {
                if self.next_file == store.files.len() {
                    return None;
                }

                let split_into = 1 << (store.chunk_high_bits - store.bucket_high_bits);

                // Index of the first chunk we are going to retrieve
                let chunk_offset = self.next_file * split_into;
                for chunk in chunk_offset..chunk_offset + split_into {
                    self.chunks
                        .push_back(Vec::with_capacity(store.chunk_sizes[chunk]));
                }

                let mut len = store.buf_sizes[self.next_file];
                let buf_size = 1024;
                let mut buffer = Vec::<SigVal<T>>::with_capacity(buf_size);
                #[allow(clippy::uninit_vec)]
                unsafe {
                    buffer.set_len(buf_size);
                }
                let chunk_mask = (1 << store.chunk_high_bits) - 1;
                store.files[self.next_file]
                    .seek(SeekFrom::Start(0))
                    .unwrap();

                while len > 0 {
                    let to_read = buf_size.min(len);
                    unsafe {
                        buffer.set_len(to_read);
                    }
                    let (pre, buf, after) = unsafe { buffer.align_to_mut::<u8>() };
                    debug_assert!(pre.is_empty());
                    debug_assert!(after.is_empty());

                    store.files[self.next_file].read_exact(buf).unwrap();

                    // We move each signature/value pair into its chunk
                    for &v in &buffer {
                        let chunk = (v.sig[0].rotate_left(store.chunk_high_bits) as usize
                            & chunk_mask)
                            - chunk_offset;
                        self.chunks[chunk].push(v);
                    }
                    len -= to_read;
                }

                self.next_file += 1;
            }

            let res = (
                self.next_chunk,
                Cow::Owned(self.chunks.pop_front().unwrap()),
            );
            self.next_chunk += 1;
            Some(res)
        }
    }
}

impl<'a, T: ZeroCopy + Send + Sync> ExactSizeIterator for ChunkIterator<'a, T> {
    fn len(&self) -> usize {
        self.store.chunk_sizes.len() - self.next_chunk
    }
}

fn write_binary<T: ZeroCopy>(writer: &mut impl Write, tuples: &[SigVal<T>]) -> std::io::Result<()> {
    let (pre, buf, post) = unsafe { tuples.align_to::<u8>() };
    debug_assert!(pre.is_empty());
    debug_assert!(post.is_empty());
    writer.write_all(buf)
}

impl<T: ZeroCopy + 'static> SigStore<T> {
    /// Create a new store with 2<sup>`buckets_high_bits`</sup> buffers, keeping
    /// counts for chunks defined by at most `max_chunk_high_bits` high bits.
    pub fn new(buckets_high_bits: u32, max_chunk_high_bits: u32) -> Result<Self> {
        let temp_dir = tempfile::TempDir::new()?;
        let mut writers = VecDeque::new();
        for i in 0..1 << buckets_high_bits {
            let file = File::options()
                .read(true)
                .write(true)
                .create(true)
                .truncate(true)
                .open(temp_dir.path().join(format!("{}.tmp", i)))?;
            writers.push_back(BufWriter::new(file));
        }
        Ok(Self {
            len: 0,
            buckets_high_bits,
            max_chunk_high_bits,
            buckets_mask: (1u64 << buckets_high_bits) - 1,
            max_chunk_mask: (1u64 << max_chunk_high_bits) - 1,
            writers,
            bucket_sizes: vec![0; 1 << buckets_high_bits],
            chunk_sizes: vec![0; 1 << max_chunk_high_bits],
            _marker: PhantomData,
        })
    }

    /// Adds a signature/value pair to this store.
    pub fn push(&mut self, sig_val: &SigVal<T>) -> std::io::Result<()> {
        self.len += 1;
        // high_bits can be 0
        let buffer =
            ((sig_val.sig[0].rotate_left(self.buckets_high_bits)) & self.buckets_mask) as usize;
        let chunk =
            ((sig_val.sig[0].rotate_left(self.max_chunk_high_bits)) & self.max_chunk_mask) as usize;

        self.bucket_sizes[buffer] += 1;
        self.chunk_sizes[chunk] += 1;

        write_binary(&mut self.writers[buffer], std::slice::from_ref(sig_val))
    }

    /// Adds signature/value pairs to this store.
    pub fn extend(&mut self, iter: impl IntoIterator<Item = SigVal<T>>) -> Result<()> {
        for sig_val in iter {
            self.push(&sig_val)?;
        }
        Ok(())
    }

    /// The number of signature/value pairs added to the store so far.
    pub fn len(&self) -> usize {
        self.len
    }

    pub fn is_empty(&self) -> bool {
        self.len == 0
    }

    /// Flush the buffers and return a pair given by [`ChunkStore`] whose chunks are defined by
    /// the `chunk_high_bits` high bits of the signatures.
    ///
    /// It must hold that
    /// `chunk_high_bits` is at most the `max_chunk_high_bits` value provided
    /// at construction time, or this method will panic.
    pub fn into_chunk_store(mut self, chunk_high_bits: u32) -> Result<ChunkStore<T>> {
        assert!(chunk_high_bits <= self.max_chunk_high_bits);
        let mut files = Vec::with_capacity(self.writers.len());

        // Flush all writers
        for _ in 0..1 << self.buckets_high_bits {
            let mut writer = self.writers.pop_front().unwrap();
            writer.flush()?;
            let mut file = writer.into_inner()?;
            file.seek(SeekFrom::Start(0))?;
            files.push(file);
        }

        // Aggregate chunk sizes as necessary
        let chunk_sizes = self
            .chunk_sizes
            .chunks(1 << (self.max_chunk_high_bits - chunk_high_bits))
            .map(|x| x.iter().sum())
            .collect::<Vec<_>>();
        Ok(ChunkStore {
            bucket_high_bits: self.buckets_high_bits,
            chunk_high_bits,
            files,
            buf_sizes: self.bucket_sizes,
            chunk_sizes,
            _marker: PhantomData,
        })
    }
}

#[test]

fn test_sig_sorter() {
    use rand::prelude::*;
    for max_chunk_bits in [0, 2, 8, 9] {
        for buckets_high_bits in [0, 2, 8, 9] {
            for chunk_high_bits in [0, 2, 8, 9] {
                if chunk_high_bits > max_chunk_bits {
                    continue;
                }
                let mut sig_sorter = SigStore::new(buckets_high_bits, max_chunk_bits).unwrap();
                let mut rand = SmallRng::seed_from_u64(0);

                for _ in (0..10000).rev() {
                    sig_sorter
                        .push(&SigVal {
                            sig: [rand.next_u64(), rand.next_u64()],
                            val: rand.next_u64(),
                        })
                        .unwrap();
                }
                let mut chunk_store = sig_sorter.into_chunk_store(chunk_high_bits).unwrap();
                let mut count = 0;
                let iter = chunk_store.iter().unwrap();
                for chunk in iter {
                    count += 1;
                    for &w in chunk.1.iter() {
                        assert_eq!(
                            chunk.0,
                            w.sig[0].rotate_left(chunk_high_bits) as usize
                                & ((1 << chunk_high_bits) - 1)
                        );
                    }
                }
                assert_eq!(count, 1 << chunk_high_bits);
            }
        }
    }
}

#[test]
fn test_u8() {
    use rand::prelude::*;
    let mut sig_sorter = SigStore::new(2, 2).unwrap();
    let mut rand = SmallRng::seed_from_u64(0);
    for _ in (0..1000).rev() {
        sig_sorter
            .push(&SigVal {
                sig: [rand.next_u64(), rand.next_u64()],
                val: rand.next_u64() as u8,
            })
            .unwrap();
    }
    let mut chunk_store = sig_sorter.into_chunk_store(2).unwrap();
    let mut count = 0;
    let iter = chunk_store.iter().unwrap();
    for chunk in iter {
        count += 1;
        for &w in chunk.1.iter() {
            assert_eq!(chunk.0, w.sig[0].rotate_left(2) as usize & ((1 << 2) - 1));
        }
    }
    assert_eq!(count, 4);
}