Datafuse is an open source elastic and scalable cloud warehouse, it offers blazing fast query and combines elasticity, simplicity, low cost of the cloud, built to make the Data Cloud easy.
Datafuse design principles:
- Elastic In Datafuse, storage and compute resources can dynamically scale up and down on demand.
- Secure All data files and network traffic in Datafuse is encrypted end-to-end, and provide Role Based Access Control in SQL level.
- User-friendly Datafuse is an ANSI SQL compliant cloud warehouse, it is easy for data scientist and engineers to use.
- Cost-efficient Datafuse processes queries with high performance, and the user only pays for what is actually used.
Datafuse consists of three components: meta service layer, and the decoupled compute and storage layers.
The meta service is a layer to service multiple tenants. In current implementation, the meta service has components:
- Metadata - Which manages all metadata of databases, tables, clusters, the transaction, etc.
- Administration Which stores user info, user management, access control information, usage statistics, etc.
- Security Which performs authorization and authentication to protect the privacy of users' data.
The compute layer is the clusters that running computing workloads, each cluster have many nodes, each node has components:
- Planner - Builds execution plan from the user's SQL statement.
- Optimizer - Optimizer rules like predicate push down or pruning of unused columns.
- Processors - Vector-based query execution pipeline, which is build by planner instructions.
- Cache - Caching Data and Indexes based on the version.
Many clusters can attach the same database, so they can serve the query in parallel by different users.
Datafuse stores data in an efficient, columnar format as Parquet files. For efficient pruning, Datafuse also creates indexes for each Parquet file to speed up the queries.
Datafuse is currently in Alpha and is not ready to be used in production, Roadmap 2021
Datafuse is licensed under Apache 2.0.