Badger started as a key-value store, so the goal was to use the latest research to build a key-value store. It uses a log-structured merge (LSM) tree based implementation, which has an advantage in throughput when compared to B+ tree since background (disk) writes in LSM maintain a sequential access pattern.
BadgerDB was initially a persistent key-value store intended to replace DgraphDB’s dependency on RocksDB, since RocksDB is written in C++ and requires the use of Cgo to be called via Go. It was a (side) project lead by Manish Rai Jain, a chief decision maker at Dgraph. The project github repository received its first commit January 26, 2017. As the project progressed, it upgraded from being a key-value store to a database system, on October 16, 2017.
The user has the option to toggle the flag SyncWrites, to determine how immediately an update is propagated to disk. If SyncWrite is false, writes may not sync to disk immediately and the database might have to be closed before the changes are propagated to disk. In addition to changes being propagate to disk on Close(), changes are also propagated when MaxTableSize is reached.
BadgerDB uses delta encoding to reduce the size of keys, to further reduce the size of the LSMs. Also, during a read, a fingerprint of the key is stored rather than the key itself to also save space. The fingerprint of a key contains the information necessary to identify a unique key, but takes up less space. It is not a 100% accurate and can cause a false negative causing a transaction to abort (and requiring it to be restarted).
BadgerDB is an in-memory DBMS, that supports larger-than-memory databases by writing to disk. Disk is also used to make the BadgerDB persistent and resilient. BadterDB works in-memory until the MaxTableSize is reached, at which point the data is propagated to disk. A database is retrieved from and written on disk upon Open() and Close().
Multi-version Concurrency Control (MVCC) Two-Phase Locking (Deadlock Prevention)
Badger supports ACID transactions and Multi-Version Concurrency Control with Snapshot Isolation. Badger acquires locks on directories when accessing data, so multiple processes cannot open the same database at the same time. If the transaction is read-write, the transaction checks if there is a conflict and return and error if there is one.
BadgerDB supports a LSM tree, where keys are stored in the LSM and values are store in a write-ahead log called a value log. Since key tend to be smaller than values, this allows the system to maintain a smaller LSM tree. The LSM make sure to maintain a sorted order, making range queries easier to process, as well.
Serializable Snapshot Isolation
BadgerDB supports snapshot isolation. Badger also has read-only transactions (called Views), that ensures a consistent view of the database system, and that will not include changes made by an uncommitted transaction at the start of this transactions. With read-write transactions, if a conflict occurs, the transaction will written with an error notifying the user of a conflict, giving the user the option to retry the transaction.
Tuple-at-a-Time Model Vectorized Model
BadgerDB is inspired by the simplicity of LevelDB and supports Get, Set, Delete, and Iterate functions. Multiple updates can be batched-up into a single transaction, allowing the user to do a lot of writes at a time.
https://github.com/dgraph-io/badger
https://godoc.org/github.com/dgraph-io/badger
Dgraph
2017