Presto is an open source distributed SQL query engine for running interactive analytic queries against heterogeneous data sources. It was open sourced by Facebook in 2013. Although it is also known as PrestoDB, Presto is not a general-purpose database management system (DBMS). It does not manage the storage of data. Instead, Presto is a query engine which allows querying data where it lives, including Hive, Cassandra, Kafka, and relational databases. A single Presto query is able to combine data from multiple sources. Presto was designed, built and optimized for interactive queries. In comparison, both Presto and Hive support SQL queries against HDFS, while Presto is targeted at interactive queries and Hive is suitable for batch processing. Presto supports ANSI-compatible SQL statements.
Presto started out as a project at Facebook. Prior to Presto, Facebook primarily used Hive, but it wasn't optimized for high speed interactive queries. The development of Presto finished in 2012, and it was rolled out to the company in early 2013. Facebook open-sourced Presto in November 2013 under the Apache Software License.
Hash Join Broadcast Join Shuffle Join
Presto has two types of join distributions. It can support both broadcast join and partitioned (shuffle) join. The join distribution can either be specified by the user or be decided by the cost-based optimization strategies that are supported by Presto.
At each node level, Presto performs a hash-based join.
Presto is designed to query data from sources including Hadoop environments and other relational database systems, so it does not directly take the role of data storage. All data and the intermediate results are stored in-memory whenever possible. For communication between nodes, data is also stored in in-memory buffers and sent through the network. This avoids the high cost of I/O operations and speeds up the execution.
For memory-intensive queries, Presto also offers the functionality of spilling data to disk. But this is not a primary function of Presto and it is assumed that most of the query operations should be performed in-memory completely.
As similar to many classic MPP (massively parallel process) database management systems, Presto utilizes a shared-nothing system architecture.
Presto is deployed on a cluster of nodes. A node can take the role of either a coordinator or a worker. Each node has its own private disk and memory, and the user can configure the memory usage of each node. Since Presto does not store data directly, the disk of each node is used minimally for storing logs only, and all communications are done through the network.
Decomposition Storage Model (Columnar)
To execute a query, Presto splits the assignments to each worker, and the workers fetch the data from the data sources. The unit of data that Presto locally operates on is called a page. The page is a columnar of a sequence of rows.
Dictionary Encoding Run-Length Encoding
Presto can operate on dictionary and run-length-encoded blocks from connectors. When generating intermediate results, Presto also produces compressed data in the form of dictionary or run-length-encoded blocks.
Read Uncommitted Read Committed Serializable Repeatable Read
Depending on the underlying source of data, whether or not transaction is supported depends on the implementation of the specific connector. For the connectors that support transactions, Presto API supports 4 different types of isolation levels. The isolation level is to be specified when a transaction is started.
https://github.com/prestosql/presto
https://prestosql.io/docs/current/
Presto Open Source Community
2013
Accumulo, Cassandra, Elasticsearch, Hive, Kudu, MongoDB, MySQL, Pinot, PostgreSQL, Redis, Redshift