Grid is a distributed SQL database that integrates in-memory storage for hot data with disk-based logs and sealed GridMap files. The system uses a custom consensus protocol called ORCHID to manage data consistency and write operations across nodes without traditional leader election mechanisms like those found in Raft.
The ORCHID protocol relies on the Kuramoto model of coupled oscillators. Each node maintains a phase parameter, and the cluster calculates an order parameter R to assess synchronization. If R falls below a threshold of 0.85, write operations are rejected. The node with the lowest ID among synchronized participants acts as the proposer.
For write operations, the proposer calculates a checksum of the operation and its sequence. It requires a majority of configured participants to agree on the checksum before committing to the disk-based OpLog. If a network partition occurs and a majority cannot be reached, the system rejects writes to prevent data divergence.
The system supports both JDBC and reactive stacks for SQL operations. Performance benchmarks on a host with an Intel Core i7 1240H, 64 GiB RAM, and NVMe storage show write speeds of approximately 4922 operations per second for two nodes with fsync enabled, and up to 7095 operations per second for a single node.
Grid has undergone formal verification using TLA+/TLC and Jepsen testing to ensure consistency under network and process failures. Future development plans include expanding the SQL dialect, improving the transparency of the sealed GridMap storage, and implementing distributed locking mechanisms.
