sled likes eating data! it's alpha

An alpha modern embedded database.

extern crate sled;

use sled::{ConfigBuilder, Tree};

let config = ConfigBuilder::new()
  .path(path)
  .build();

let tree = Tree::start(config).unwrap();

// set and get
tree.set(k, v1);
assert_eq!(tree.get(&k), Ok(Some(v1)));

// compare and swap
tree.cas(k, Some(v1), Some(v2));

// scan forward
let mut iter = tree.scan(k);
assert_eq!(iter.next(), Some(Ok((k, v2))));
assert_eq!(iter.next(), None);

// deletion
tree.del(&k);

We also support merge operators.

fn concatenate_merge(
  _key: &[u8],               // the key being merged
  old_value: Option<&[u8]>,  // the previous value, if one existed
  merged_bytes: &[u8]        // the new bytes being merged in
) -> Option<Vec<u8>> {       // set the new value, return None to delete
  let mut ret = old_value
    .map(|ov| ov.to_vec())
    .unwrap_or_else(|| vec![]); 

  ret.extend_from_slice(merged_bytes);

  Some(ret)
}

let config = ConfigBuilder::new()
  .temporary(true)
  .merge_operator(concatenate_merge)
  .build();

let tree = Tree::start(config).unwrap();

tree.set(k, vec![0]);
tree.merge(k, vec![1]);
tree.merge(k, vec![2]);
assert_eq!(tree.get(&k), Ok(Some(vec![0, 1, 2])));

// sets replace previously merged data,
// bypassing the merge function.
tree.set(k, vec![3]);
assert_eq!(tree.get(&k), Ok(Some(vec![3])));

// merges on non-present values will add them
tree.del(&k);
tree.merge(k, vec![4]);
assert_eq!(tree.get(&k), Ok(Some(vec![4])));

features

ordered map API
fully atomic single-key operations, supports CAS
merge operators
zstd compression (use the zstd build feature)
cpu-scalable lock-free implementation
SSD-optimized log-structured storage

goals

don't make the user think. the interface should be obvious.
don't surprise users with performance traps.
don't wake up operators. bring reliability techniques from academia into real-world practice.
don't use so much electricity. our data structures should play to modern hardware's strengths.

plans

beat LSM trees for reads and traditional B+ trees for writes
MVCC, transactions, and snapshots provided via a higher-level Db versioned-key interface
form the iron core of a linearizable store and a flexible location-agnostic store
forward-compatible binary format
bindings for other languages

known issues, warnings

keys and values must fit into io_buf_size divided by min_items_per_segment.
quite young, should be considered unstable for the time being
the C API is likely to change rapidly
the on-disk format is going to change in non-forward compatible ways before the 1.0.0 release! after that, we will always support forward migrations.
has not yet received much attention for performance tuning, it has an extremely high theoretical performance but there is a bit of tuning to get there. currently only around 200k operations per second with mixed workloads, and 7 million/s for read-only workloads on tiny keys. this will be improving dramatically soon!
32 bit architectures require Rust nightly with the nightly feature enabled.

contribution welcome!

want to help advance the state of the art in open source embedded databases? check out CONTRIBUTING.md!

architecture

lock-free tree on a lock-free pagecache on a lock-free log. the pagecache scatters partial page fragments across the log, rather than rewriting entire pages at a time as B+ trees for spinning disks historically have. on page reads, we concurrently scatter-gather reads across the log to materialize the page from its fragments.

spacejam / sled