21 stable releases (9 major)

15.0.7 Jul 11, 2023
15.0.4 Jun 25, 2023
15.0.2 Nov 27, 2022
14.0.3 Oct 2, 2022
0.0.0 Mar 14, 2020

#35 in Database implementations


Used in 3 crates (2 directly)

MIT/Apache

95KB
1.5K SLoC

Marble

Garbage-collecting disk-based object-store. See examples/kv.rs for a minimal key-value store built on top of this.

Marble is sled's future storage engine.

Supports 4 methods:

  • read: designed for low-latency, concurrent reads of objects
  • write_batch: designed for bulky, high-throughput writes of large batches of objects, ideal for compacting write-ahead logs into a set of updates to specific objects. Does not block calls to read except for brief moments where metadata is being updated.
  • maintenance: compacts backing storage files that have become fragmented. Blocks concurrent calls to write_batch but does not block readers any more than write_batch does. Returns the number of successfully rewritten objects.
  • stats: returns statistics about live and total objects in the backing storage files.

Marble is a low-level object store that can be used to build your own storage engines and databases on top of.

At a high-level, it supports atomic batch writes and single-object reads. Garbage collection is manual. All operations are blocking. Nothing is cached in-memory except for zstd dictionaries and file handles to all storage files. Objects may be sharded upon GC by providing a custom Config::partition_function. Partitioning is not performed on the write batch when it is initially written, because the write batch must be stored in a single file for it to be atomic. But when future calls to Marble::maintenance defragment the storage files by rewriting objects that are still live, it will use this function to assign the rewritten objects into a particular partition.

You should think of Marble as the heap that you flush your write-ahead logs into periodically. It will create a new file for each write batch, and this might actually expand to more files after garbage collection if the batch is significantly larger than the Config::target_file_size.

Marble does not create any threads or call Marble::maintenance automatically under any conditions. You should probably create a background thread that calls this periodically.

Pretty much the only "fancy" thing that Marble does is that it can be configured to create a zstd dictionary that is tuned specifically to your write batches. This is disabled by default and can be configured by setting the Config::zstd_compression_level to something other than None (the level is passed directly to zstd during compression). Compression is bypassed if batches have fewer than 8 items or the average item length is less than or equal to 8.

Examples

let marble = marble::open("heap").unwrap();

// Write new data keyed by a `u64` object ID.
// Batches contain insertions and deletions
// based on whether the value is a Some or None.
marble.write_batch([
    (0_u64, Some(&[32_u8] as &[u8])),
    (4_u64, None),
]).unwrap();

// read it back
assert_eq!(marble.read(0).unwrap(), Some(vec![32].into_boxed_slice()));
assert_eq!(marble.read(4).unwrap(), None);
assert_eq!(marble.read(6).unwrap(), None);

// after a few more batches that may have caused fragmentation
// by overwriting previous objects, perform maintenance which
// will defragment the object store based on `Config` settings.
let objects_defragmented = marble.maintenance().unwrap();

// print out system statistics
dbg!(marble.stats());

which prints out something like

marble.stats() = Stats {
    live_objects: 1048576,
    stored_objects: 1181100,
    dead_objects: 132524,
    live_percent: 88,
    files: 11,
}

If you want to customize the settings passed to Marble, you may specify your own Config:

let config = marble::Config {
    path: "my_path".into(),
    zstd_compression_level: Some(7),
    fsync_each_batch: true,
    target_file_size: 64 * 1024 * 1024,
    file_compaction_percent: 50,
    ..Default::default()
};

let marble = config.open().unwrap();

A custom GC sharding function may be provided for partitioning objects based on the object ID and size. This may be useful if your higher-level system allocates certain ranges of object IDs for certain types of objects that you would like to group together in the hope of grouping items together that have similar fragmentation properties (similar expected lifespan etc...). This will only shard objects when they are defragmented through the Marble::maintenance method, because each new write batch must be written together in one file to retain write batch atomicity in the face of crashes.

// This function shards objects into partitions
// similarly to a slab allocator that groups objects
// into size buckets based on powers of two.
fn shard_by_size(object_id: u64, object_size: usize) -> u8 {
    let next_po2 = object_size.next_power_of_two();
    u8::try_from(next_po2.trailing_zeros()).unwrap()
}

let config = marble::Config {
    path: "my_sharded_path".into(),
    partition_function: shard_by_size,
    ..Default::default()
};

let marble = config.open().unwrap();

Defragmentation is always generational, and will group rewritten objects together. Written objects can be further sharded based on a configured partition_function which allows you to shard objects by ObjectId and the size of the object raw bytes.

Marble solves a pretty basic problem in database storage: storing arbitrary bytes on-disk, getting them back, and defragmenting files.

You can think of it as a KV where keys are non-zero u64's, and values are arbitrary blobs of raw bytes.

Writes are meant to be performed in bulk by some background process. Each call to Marble::write_batch creates at least one new file that stores the objects being written. Multiple calls to fsync occur for each call to write_batch. It is blocking. Object metadata is added to the backing wait-free pagetable incrementally, not atomically, so if you rely on batch atomicity, you should serve the batch's objects directly from a cache of your own until write_batch returns. However, upon crash, batches are recovered atomically.

Reads can continue mostly unblocked while batch writes and maintenance are being handled.

You are responsible for:

  • calling Marble::maintenance at appropriate intervals to defragment storage files.
  • choosing appropriate configuration tunables for your desired space and write amplification.
  • ensuring the Config.partition_function is set to a function that appropriately shards your objects based on their ObjectId and/or size. Ideally, objects that have expected death times will be colocated in a shard so that work spent copying live objects is minimized.
  • allocating and managing free ObjectId's.

If you want to create an industrial database on top of Marble, you will probably also want to add:

  • logging and a write cache for accumulating updates that occasionally get flushed to Marble via write_batch. Remember, each call to write_batch creates at least one new file and fsyncs multiple times, so you should batch calls appropriately. Once the log or write cache has reached an appropriate size, you can have a background thread write a corresponding batch of objects to its storage, and once write_batch returns, the corresponding log segments and write cache can be deleted, as the objects will be available via Marble::read.
  • an appropriate read cache. Marble::read always reads directly from disk.
  • for maximum SSD friendliness, your own log should be configurable to be written to a separate storage device, to avoid comingling writes that have vastly different expected death times.
  • dictionary-based compression for efficiently compressing objects that may be smaller than 64k.

Ideas for getting great garbage collection performance:

  • give certain kinds of objects a certain ObjectId range. for example, tree index nodes can be above 1<<63, and tree leaf nodes can be below that point. The Config.partition_function can return the shard 0 for leaf nodes, and 1 for index nodes, and they will always be written to separate files.
  • WiscKey-style sharding of large items from other items, based on the size of the object. Assign a shard ID based on which power of 2 the object size is.
  • Basically any sharding strategy that tends to group items together that exhibit some amount of locality in terms of expected mutations or overall lifespan.

In short, you get to focus on a bunch of the fun parts of building your own database, without so much effort spent on boring file garbage collection.

Dependencies

~3–8.5MB
~70K SLoC