History[07][08][09]

RisingWave Database is built by RisingWave Labs (formerly known as Singularity Data), a database systems startup founded in 2021 by former IBM researcher and Amazon Redshift engineer Yingjun Wu.

While working at Amazon Redshift, Wu noticed that existing database systems cannot process streaming data efficiently and existing streaming systems were too complicated for most companies to use. These observations motivated Wu to found RisingWave Labs with a mission to “democratize stream processing”.

Checkpoints[10][11][12]

RisingWave uses the Chandy–Lamport algorithm to create consistent checkpoints.

To ensure that data is correct and consistent, read queries always fetch data from the most recent checkpoint. This means RisingWave does not ensure read-after-write consistency.

A local shared buffer is used to stage uncommitted write batches submitted by operators. The storage manager will notify all operators to commit their buffered writes into the shared storage when the checkpoint trigger message has reached all operators.

Compression[13][14]

RisingWave applies both naive compression and prefix compression at the block-level. It uses LZ4 and Zstd for naive compression.

Concurrency Control[15]

RisingWave does not support concurrency control in its compute engine. However, it does employ Multi-version Concurrency Control (MVCC) in its storage engine.

Data Model[16]

RisingWave uses a relational data model. Its relational tables are composed of a list of strongly-typed columns. All columns are implicitly nullable. The supported primitive data types are boolean, integer, fixed-point and floating-point numbers, strings, and temporals. Composite data types of struct and list are also supported.

Foreign Keys

Indexes[17][15]

RisingWave does not support traditional index data structures. Instead, indexes are implemented as specialized materialized views. RisingWave stores materialized views as key-value pairs in log-structured merge trees.

Isolation Levels

Joins[18][19][20]

RisingWave supports hash join, nested loop join, and index nested loop join (also called lookup join). RisingWave has two execution modes: the batch-query mode and the streaming mode. All three join strategies are used in the batch-query mode, while only hash join and lookup join are used in the streaming mode. The supported join types are inner join, left outer join, right outer join, and full outer join. Window joins are also available as part of RisingWave's support for time window functions.

Logging[11]

Parallel Execution[21][22][23][24]

RisingWave supports bushy parallelism in both its batch-query and streaming modes. Consistent hashing is employed to partition data for parallel execution.

Query Compilation

Query Execution[25][16]

RisingWave adopts the top-to-bottom vectorized processing model. Operators emit a Data Chunk in batch-query mode and a Stream Chunk in streaming mode. A Data Chunk consists of multiple columns and a visibility array (represents each row's visibility status for row filtering purposes). A Stream Chunk consists of multiple columns, a visibility array, and an additional ops column marking an operation on each row. Each entry in the ops column can be one of Delete, Insert, UpdateDelete, or UpdateInsert.

Query Interface[26][27]

The RisingWave SQL query interface is mostly compatible with PostgreSQL. It has client libraries in Java, Node.js, Python, and Go. RisingWave is wire compatible with PostgreSQL and can be accessed with PostgreSQL terminal psql. As a cloud-native database, RisingWave can be integrated with cloud services such as Confluent Cloud, DataStax, and Grafana Cloud. To support stream processing, RisingWave can be integrated with the following message brokers or streaming services: Kafka, Redpanda, Apache Pulsar, DataStax Astra Streaming, StreamNative Cloud, and Kinesis Data Streams.

Storage Architecture[15][12]

RisingWave has a disk-oriented storage architecture. Files are directly written to an S3-compatible shared storage service by default. RisingWave also supports local drives, Google Cloud Storage, and HDFS/WebHDFS as shared storage destinations.

Storage Model[15][28][16]

RisingWave uses the row store format. Each row is encoded into key-value entries.

Storage Organization[15]

RisingWave uses a LSM-Tree based key-value storage engine that provides MVCC read and write capabilities. All key-value pairs are stored in block-based sorted strings tables (SST). Each SST consists of two files:

• The .data file is composed of 64KB blocks containing the actual key-value pairs.
• The .meta file contains metadata such as min-max index, Bloom filter, and per-block metadata.

Stored Procedures[29]

RisingWave offers UDF implemented as external functions in Python.

System Architecture[21][30]

RisingWave adopts a disaggregated architecture to support elastically scaling the compute nodes without migrating the storage. A RisingWave database is composed of four key layers:

• A serving layer that parses, plans, and optimizes SQL queries.
• A processing layer that performs all the data computations and updates.
• A persistence layer that stores and retrieves data on an object storage service (such as S3).
• A meta server that coordinates the operations between the serving layer and the processing layer.

The serving, processing, and persistent layers each consist of multiple nodes. Compute nodes in the processing layer execute the optimized query plan (generated by the serving layer) in parallel. Compactor nodes in the persistent layer compact data with consistent hashing and upload the resulting SST files to shared storage.

The meta server is responsible for many tasks, including but not limited to:

• Managing the stream graph of RisingWave
• Managing barrier injection and collection for initiating checkpoint capture
• Manages the cluster information, such as the address and status of nodes
• Detecting failure by periodically sending heartbeats to serving-layer and processing-layer nodes

Views[31][32][33]

RisingWave supports both views and materialized views. New materialized views can be created based on existing materialized views. RisingWave updates materialized views incrementally as new data is fed into the database.