hyperfile 0.12.0

use std::io::Result;
use std::sync::Arc;
use std::collections::{HashMap, HashSet};
use log::debug;
use aws_sdk_s3::Client;
use tokio::sync::Mutex;
use btree_ondisk::BlockLoader;
use btree_ondisk::node::{BtreeNode, BTREE_NODE_LEVEL_MIN};
use crate::{BlockPtr, BMapUserData};
use crate::meta_format::BlockPtrFormat;
use crate::segment::Segment;
use crate::s3uri::S3Uri;
use crate::file::ReadTiming;
use crate::BlockIndex;
use crate::staging::s3::S3Staging;

pub struct S3BlockLoader {
    pub client: Client,
    pub bucket: String,
    pub root_path: String,
    /// Shared with the file's staging handle so meta-block requests
    /// land in the same counters as data requests.
    read_timing: Arc<ReadTiming>,
    /// The most node requests of one `read_batch` in flight at once.
    concurrency: usize,
    /// Nodes whose parent is loaded and that the walk will want, by the object
    /// holding them. By object, not checkpoint: under the stream format one
    /// checkpoint's leaves are in several objects, and the first fetched would
    /// take the others' entries and leave them to be fetched once per node.
    backlog: Arc<Mutex<HashMap<crate::SegmentId, HashSet<BlockPtr>>>>,
}

impl Clone for S3BlockLoader {
    fn clone(&self) -> Self {
        Self {
            client: self.client.to_owned(),
            bucket: self.bucket.clone(),
            root_path: self.root_path.clone(),
            read_timing: self.read_timing.clone(),
            concurrency: self.concurrency,
            backlog: self.backlog.clone(),
        }
    }
}

impl S3BlockLoader {
    pub fn new(client: &Client, bucket: &str, root_path: &str, read_timing: Arc<ReadTiming>) -> Self {
        Self {
            client: client.to_owned(),
            bucket: bucket.to_string(),
            root_path: root_path.to_string(),
            backlog: Arc::new(Mutex::new(HashMap::new())),
            read_timing,
            concurrency: crate::DEFAULT_META_READ_MAX_CONCURRENCY,
        }
    }

    /// Fill each of `wanted` — nodes of object `segid` — from one fetch of that
    /// object's index chunk, and return the other nodes of the chunk the walk is
    /// waiting for (see `backlog`).
    async fn read_segment(&self, segid: crate::SegmentId, wanted: &mut [(BlockPtr, &mut [u8])], ud: &BMapUserData)
        -> Result<Vec<(BlockPtr, Vec<u8>)>>
    {
        let meta_block_size = wanted[0].1.len();
        let key = format!("{}/{}", self.root_path, Segment::segid_to_staging_file_id(segid));

        debug!("read meta blocks chunk from s3://{}/{} from s3", self.bucket, &key);
        // The chunk fetch issues its own requests and counts them
        // itself, since it may make two.
        let (meta_block_offset, seg_meta_block_size, mut meta_blocks, meta_block_buf) =
            S3Staging::do_fetch_meta_blocks_chunk(&self.client, &self.bucket, &key, segid,
                crate::staging::s3::Layout::Known { stream: ud.blk_ptr_format.is_stream() }, Some(&self.read_timing)).await?;
        assert!(meta_block_size == seg_meta_block_size);
        for (v, buf) in wanted.iter_mut() {
            // Metadata blocks are in the summary part, so a pointer to one carries
            // that part, and its offset is into the object this chunk is from.
            let (_, offset) = BlockPtrFormat::decode(v, &ud.blk_ptr_format);
            assert!(offset >= meta_block_offset);
            buf.copy_from_slice(&meta_block_buf[offset - meta_block_offset..offset - meta_block_offset + meta_block_size]);
        }

        let mut v_nonleaf_meta_block_ptr: Vec<(crate::SegmentId, BlockPtr)> = Vec::new();
        // for each meta block:
        // 1. get all next level block ptr
        // 2. rebuild block ptr for meta block itself
        let mut block_seq = 0;
        let mut file_off = meta_block_offset;
        let mut output = Vec::new();
        // TODO: more efficient way to split meta_block_buf into meta_block
        for meta_block_slice in meta_block_buf.chunks(meta_block_size) {
            let meta_block = meta_block_slice.to_vec();

            // decode nonleaf node, get back all next level block ptr
            let node = BtreeNode::<BlockIndex, BlockPtr, BlockPtr>::from_slice_ref(&meta_block)?;
            // we only care about meta data nodes
            if node.get_level() > BTREE_NODE_LEVEL_MIN {
                for idx in 0..node.get_nchild() {
                    let blk_ptr = *node.get_val(idx);
                    let (child, _) = BlockPtrFormat::decode(&blk_ptr, &ud.blk_ptr_format);
                    // The child's object: this one or an earlier one, of this
                    // checkpoint or an older.
                    v_nonleaf_meta_block_ptr.push((child, *node.get_val(idx)));
                }
            }

            // rebuild bloock ptr for myself
            let block_ptr = BlockPtrFormat::encode(segid, file_off, block_seq, &ud.blk_ptr_format);
            output.push((block_ptr, meta_block));
            file_off += meta_block_size;
            block_seq += 1;
            meta_blocks -= 1;
        }
        assert!(meta_blocks == 0);

        // handle backlog
        let mut lock = self.backlog.lock().await;

        // The children of the nodes just read, under the objects holding them.
        for (object, blk_ptr) in v_nonleaf_meta_block_ptr.into_iter() {
            lock.entry(object).or_default().insert(blk_ptr);
        }

        // What of this object the walk is waiting for.
        let this_segid_backlog: Option<HashSet<BlockPtr>> = lock.remove(&segid);

        drop(lock);

        // discard block ptr not in our backlog waiting list
        if let Some(segid_backlog) = this_segid_backlog {
            output.retain(|(blk_ptr, _)| segid_backlog.contains(blk_ptr));
            return Ok(output);
        }

        // if we don't have backlog wait list for this segment, just return blank vec
        Ok(Vec::new())
    }

    /// At most `n` node requests of one batch in flight at once; zero is one.
    pub fn with_concurrency(mut self, n: usize) -> Self {
        self.concurrency = n.max(1);
        self
    }
}

impl BlockLoader<BlockPtr> for S3BlockLoader {
    async fn read(&self, v: BlockPtr, buf: &mut [u8], user_data: u32) -> Result<Vec<(BlockPtr, Vec<u8>)>> {
        let ud = BMapUserData::from_u32(user_data);
        let (segid, _) = BlockPtrFormat::decode(&v, &ud.blk_ptr_format);
        self.read_segment(segid, &mut [(v, buf)], &ud).await
    }

    /// The nodes of one tree level. Those in one object come from one fetch of
    /// its index chunk — the chunk holds them all, and fetching it once for each
    /// would be the same requests again — and the objects are fetched at most
    /// `concurrency` at a time.
    async fn read_batch(&self, ids: &[BlockPtr], bufs: &mut [Vec<u8>], user_data: u32) -> Result<Vec<(BlockPtr, Vec<u8>)>> {
        use futures::stream::{self, StreamExt, TryStreamExt};
        assert_eq!(ids.len(), bufs.len(), "read_batch: ids and bufs length must match");
        let ud = BMapUserData::from_u32(user_data);
        let mut by_object: std::collections::BTreeMap<crate::SegmentId, Vec<(BlockPtr, &mut [u8])>> =
            std::collections::BTreeMap::new();
        for (id, buf) in ids.iter().copied().zip(bufs.iter_mut()) {
            let (segid, _) = BlockPtrFormat::decode(&id, &ud.blk_ptr_format);
            by_object.entry(segid).or_default().push((id, buf.as_mut_slice()));
        }
        let ud = &ud;
        // Made first and then run: a stream mapped by a closure here trips the
        // compiler's check that the whole future is `Send`.
        let reads: Vec<_> = by_object.into_iter()
            .map(|(segid, mut wanted)| async move { self.read_segment(segid, &mut wanted, ud).await }).collect();
        stream::iter(reads)
            .buffer_unordered(self.concurrency)
            .try_concat()
            .await
    }

    fn dup_from_new_path(self, new_path: &str) -> Self {
        let s3uri = S3Uri::parse(new_path).expect("input new path is not a valid s3 uri");
        let mut clone = self.clone();
        clone.bucket = s3uri.bucket.to_string();
        clone.root_path = s3uri.key.to_string();
        clone
    }
}