History

In an attempt to extend HDFS's success from Data Analysis to Query Serving workloads (this workload requires low latency), Dhruba Borthakur enhanced HBase and make its latencies twice as slow as MySQL server. Then when flash storage came out, it became clear that a new storage engine was needed to serve a random workload efficiently.

He started seeking for new techniques to build next generation key-value store, especially for serving data that resides on fast storage. Since he used flash storage, the network data access was 50% higher overhead than local data access, which meant that embedded database within an application could have much slower latency than applications that access data across the network.

At that time, there were several existing embedded database: BerkeleyDB, SQLite3, as well as leveldb, which was the fastest according to open-source benchmarks, plus its written in c++, thus leveldb became the first choice for his benchmarking.

Soon he found out that leveldb was not suitable for the Query Serving workloads. leveldb only works well when the size of the database is smaller than the size of RAM. Additionally, leveldb's single-threaded compaction process was insufficient to drive server workloads, and it was not able to consume all the I/Os that were offered by the underlying flash storage.

Finally, he decided that the best path was to fork the leveldb code and change its architecture to suit the needs (use a database that can drive fast storage hardware), then RocksDB was born.

Checkpoints[05]

RocksDB supports snapshot checkpoints on a running database, which were took and stored in a separate directory. Checkpoints can be used for both full and incremental backups.

Concurrency Control[06]

RocksDB supports two kinds of transactions: TransactionDB and OptimisticTransactionDB. Transactions have BEGIN, COMMIT and ROLLBACK APIs, allowing applications to modify the data concurrently while RocksDB is checking conflicts. RocksDB supports both pessimistic and optimistic concurrency control.

Data Model[07]

RocksDB uses embedded key-value store, in which keys and values are arbitrary byte streams. RocksDB organizes all data in sorted order and the common operations are Get(key), Put(key), Delete(key) and NewIterator().

Indexes[08][09]

All RocksDB's persistent data is stored in a collection of SSTs. For a Get() query, RocksDB scans through mutable memtable, list of immutable memtables, and SST files to find the target key. SST files are stored in levels.

Isolation Levels[06]

RocksDB supports SNAPSHOT isolation level.

Joins

Logging[07]

Every update to RocksDB is written to two places: 1) an in-memory data structure (called memtable) and 2) write ahead log on disk. When failure happens, write ahead logs can be used to restore the data in the memtable. By default, RocksDB ensures process crash consistency by fflush()ing the write ahead log after every user write.

Query Compilation

Query Execution[10]

The database provides Put, Delete, and Get methods to modify/query the database. This is not quite the same as query execution in other relational databases.

Query Interface[10]

The database provides Put, Delete, and Get methods to modify/query the database. This is not quite the same as query execution in other relational databases.

Storage Architecture[11]

RocksDB is designed to make full use of the fast accessing speed of flash storage. However, it also supports pure memory storage.

Storage Model[09]

All RocksDB's data is organized in a collection of SSTs. They often use sst, table and sst file interchangeably.

Storage Organization

Stored Procedures

System Architecture[03]

RocksDB supports flexible configuration settings, which may be tuned to run on various production environments, including pure memory, Flash, hard disks or HDFS.

Views