github.com/steveyen/gkvlite

gkvlite

gkvlite is a simple, ordered, ACID, key-value persistence library for Go.

Overview

gkvlite is a library that provides a simple key-value persistence store, inspired by SQLite and CouchDB/Couchstore.

gkvlite has the following features...

• 100% implemented in the Go Language (golang).
• Open source license - MIT.
• Keys are ordered, so range scans are supported.
• On-disk storage for a "Store" is a single file.
• ACID properties are supported via a simple, append-only, copy-on-write storage design.
• Concurrent goroutine support (multiple readers, single mutator).

Key concepts

• A Store is held in a single file.
• A Store can have zero or more Collections.
• A Collection can have zero or more Items.
• An Item is a key and value.
• A key is a []byte, max length 64KB (length is uint16).
• A value is a []byte, max length 4GB (length is uint32).

ACID properties

• Atomicity - all unpersisted changes from all Collections during a Store.Flush() will be persisted atomically.
• Consistency - simple key-value level consistency is supported.
• Isolation - mutations won't affect concurrent readers or snapshots.
• Durability - you control when you want to Flush() & fsync so your application can address its performance-vs-safety tradeoffs appropriately.

Performance

• In general, performance is similar to probabilistic balanced binary tree performance.
• O(log N) performance for item retrieval, insert, update, delete.
• O(log N) performance to find the smallest or largest items (by key).
• Range iteration performance is same as binary tree traversal performance.
• You can optionally retrieve just keys only, to save I/O & memory resources.

Snapshots

• Read-only Store snapshots are supported.
• Mutations on the original Store won't be seen by snapshots.
• Snapshot creation is a fast O(1) operation per Collection.

Concurrency

The simplest way to use gkvlite is in single-threaded fashion, such as by using a go channel or other application-provided locking to serialize access to a Store.

More advanced users may want to use gkvlite's support for concurrent goroutines. The idea is that multiple read-only goroutines, a single read-write (mutation) goroutine, and a single persistence (flusher) goroutine do not need to block each other.

More specifically, you should have only a single read-write (or mutation) goroutine per Store, and should have only a single persistence goroutine per Store (doing Flush()'s). And, you may have multiple, concurrent read-only goroutines per Store (doing read-only operations like Get()'s, Visit()'s, Snapshot()'s, CopyTo()'s, etc).

A read-only goroutine that performs a long or slow read operation, like a Get() that must retrieve from disk or a range scan, will see a consistent, isolated view of the collection. That is, mutations that happened after the slow read started will not be seen by the reader.

IMPORTANT: In concurrent usage, the user must provide a StoreFile implementation that is concurrent safe.

Note that os.File is not a concurrent safe implementation of the StoreFile interface. You will need to provide your own implementation of the StoreFile interface, such as by using a channel to serialize StoreFile method invocations.

Finally, advanced users may also use a read-write (mutation) goroutine per Collection instead of per Store. There should only be, though, only a single persistence (flusher) goroutine per Store.

Other features

• In-memory-only mode is supported, where you can use the same API but without any persistence.
• You provide the os.File - this library just uses the os.File you provide.
• You provide the os.File.Sync() - if you want to fsync your file, call file.Sync() after you do a Flush().
• Similar to SQLite's VFS feature, you can supply your own StoreFile interface implementation instead of an actual os.File, for your own advanced testing or I/O interposing needs (e.g., compression, checksums, I/O statistics, caching, enabling concurrency, etc).
• You can specify your own KeyCompare function. The default is bytes.Compare(). See also the StoreCallbacks.KeyCompareForCollection() callback function.
• Collections are written to file sorted by Collection name. This allows users with advanced concurrency needs to reason about how concurrent flushes interact with concurrent mutations. For example, if you have a main data collection and a secondary-index collection, with clever collection naming you can know that the main collection will always be "ahead of" the secondary-index collection even with concurrent flushing.
• A Store can be reverted using the FlushRevert() API to revert the last Flush(). This brings the state of a Store back to where it was as of the next-to-last Flush(). This allows the application to rollback or undo changes on a persisted file.
• Reverted snapshots (calling FlushRevert() on a snapshot) does not affect (is isolated from) the original Store and does not affect the underlying file. Calling FlushRevert() on the main Store, however, will adversely affect any active snapshots; where the application should stop using any snapshots that were created before the FlushRevert() invocation on the main Store.
• To evict O(log N) number of items from memory, call Collection.EvictSomeItems(), which traverses a random tree branch and evicts any clean (already persisted) items found during that traversal. Eviction means clearing out references to those clean items, which means those items can be candidates for GC.
• You can control item priority to access hotter items faster by shuffling them closer to the top of balanced binary trees (warning: intricate/advanced tradeoffs here).
• Tree depth is provided by using the VisitItemsAscendEx() or VisitItemsDescendEx() methods.
• You can associate transient, ephemeral (non-persisted) data with your items. If you do use the Item.Transient field, you should use sync/atomic pointer functions for concurrency correctness. In general, Item's should be treated as immutable, except for the Item.Transient field.
• Application-level Item.Val buffer management is possible via the optional ItemValAddRef/ItemValDecRef() store callbacks, to help reduce garbage memory. Libraries like github.com/steveyen/go-slab may be helpful here.
• Errors from file operations are propagated all the way back to your code, so your application can respond appropriately.
• Tested - "go test" unit tests.
• Docs - "go doc" documentation.

LICENSE

Open source - MIT licensed.

Examples

import (
    "os"
    "github.com/steveyen/gkvlite"
)

f, err := os.Create("/tmp/test.gkvlite")
s, err := gkvlite.NewStore(f)
c := s.SetCollection("cars", nil)

// You can also retrieve the collection, where c == cc.
cc := s.GetCollection("cars")

// Insert values.
c.Set([]byte("tesla"), []byte("$$$"))
c.Set([]byte("mercedes"), []byte("$$"))
c.Set([]byte("bmw"), []byte("$"))

// Replace values.
c.Set([]byte("tesla"), []byte("$$$$"))

// Retrieve values.
mercedesPrice, err := c.Get([]byte("mercedes"))

// One of the most priceless vehicles is not in the collection.
thisIsNil, err := c.Get([]byte("the-apollo-15-moon-buggy"))

// Iterate through items.
c.VisitItemsAscend([]byte("ford"), true, func(i *gkvlite.Item) bool {
    // This visitor callback will be invoked with every item
    // with key "ford" and onwards, in key-sorted order.
    // So: "mercedes", "tesla" are visited, in that ascending order,
    // but not "bmw".
    // If we want to stop visiting, return false;
    // otherwise return true to keep visiting.
    return true
})

// Let's get a snapshot.
snap := s.Snapshot()
snapCars := snap.GetCollection("cars")

// The snapshot won't see modifications against the original Store.
c.Delete([]byte("mercedes"))
mercedesIsNil, err := c.Get([]byte("mercedes"))
mercedesPriceFromSnapshot, err := snapCars.Get([]byte("mercedes"))

// Persist all the changes to disk.
s.Flush()

f.Sync() // Some applications may also want to fsync the underlying file.

// Now, other file readers can see the data, too.
f2, err := os.Open("/tmp/test.gkvlite")
s2, err := gkvlite.NewStore(f2)
c2 := s.GetCollection("cars")

bmwPrice, err := c2.Get([]byte("bmw"))

Tips

Because all collections are persisted atomically when you flush a store to disk, you can implement consistent secondary indexes by maintaining additional collections per store. For example, a "users" collection can hold a JSON document per user, keyed by userId. Another "userEmails" collection can be used like a secondary index, keyed by "emailAddress:userId", with empty values (e.g., []byte{}).

Bulk inserts or batched mutations are roughly supported in gkvlite where your application should only occasionally invoke Flush() after N mutations or after a given amount of time, as opposed to invoking a Flush() after every Set/Delete().

Implementation / design

The fundamental data structure is an immutable treap (tree + heap).

When used with random heap item priorities, treaps have probabilistic balanced tree behavior with the usual O(log N) performance bounds expected of balanced binary trees.

The persistence design is append-only, using ideas from Apache CouchDB and Couchstore / Couchbase, providing a simple approach to reaching ACID properties in the face of process or machine crashes. On re-opening a file, the implementation scans the file backwards looking for the last good root record and logically "truncates" the file at that point. New mutations are appended from that last good root location. This follows the MVCC (multi-version concurrency control) and "the log is the database" approach of CouchDB / Couchstore / Couchbase.

TRADEOFF: the append-only persistence design means file sizes will grow until there's a compaction. To get a compacted file, use CopyTo() with a high "flushEvery" argument.

The append-only file format allows the FlushRevert() API (undo the changes on a file) to have a simple implementation of scanning backwards in the file for the next-to-last good root record and physically truncating the file at that point.

TRADEOFF: the append-only design means it's possible for an advanced adversary to corrupt a gkvlite file by cleverly storing the bytes of a valid gkvlite root record as a value; however, they would need to know the size of the containing gkvlite database file in order to compute a valid gkvlite root record and be able to force a process or machine crash after their fake root record is written but before the next good root record is written/sync'ed.

The immutable, copy-on-write treap plus the append-only persistence design allows for fast and efficient MVCC snapshots.

TRADEOFF: the immutable, copy-on-write design means more memory garbage may be created than other designs, meaning more work for the garbage collector (GC).

TODO / ideas

• TODO: Performance: consider splitting item storage from node storage, so we're not mixing metadata and data in same cache pages. Need to measure how much win this could be in cases like compaction. Tradeoff as this could mean no more single file simplicity.

• TODO: Performance: persist items as log, and don't write treap nodes on every Flush().

• TODO: Keep stats on misses, disk fetches & writes, etc.

• TODO: Provide public API for O(log N) collection spliting & joining.

• TODO: Provide O(1) MidItem() or TopItem() implementation, so that users can split collections at decent points.

• TODO: Provide item priority shifting during CopyTo().

• TODO: Build up perfectly balanced treap from large batch of (potentially externally) sorted items.

• TODO: Allow users to retrieve an item's value size (in bytes) without having to first fetch the item into memory.

• TODO: Allow more fine-grained cached item and node eviction. Node and item objects are currently cached in-memory by gkvlite for higher retrieval performance, only for the nodes & items that you either have updated or have fetched from disk. Certain applications may need that memory instead, though, for more important tasks. The current coarse workaround is to drop all your references to any relevant Stores and Collections, start brand new Store/Collection instances, and let GC reclaim memory.

• See more TODO's throughout codebase / grep.