Atomicity
Each query executes as a single atomic transaction. All mutations within a query either commit together or roll back entirely. There are no partial writes. If any step of a multi-traversal query fails, the entire transaction is aborted and no data is modified.Consistency
The database transitions from one valid state to another on every committed transaction. All constraints (unique IDs, type safety on properties, vector dimension matching, and text index value constraints) are enforced at write time, not retroactively.Isolation
Transactions run with serializable snapshot isolation. Each transaction reads from a consistent point-in-time snapshot. Graph reads, vector search, and text search all observe that same snapshot. Write transactions also get read-your-writes, and indexed reads participate in normal conflict detection.- Read-only queries see a consistent snapshot and are never blocked by writes.
- Write transactions commit through the single writer process, so conflicting writes are detected before commit.
Durability
All committed data is persisted to object storage before the write is acknowledged to the client. Object storage provides the durability guarantee. Local caches (SSD and in-memory) are performance optimizations, not durability mechanisms. A full cache loss does not result in data loss.High Availability
Helix Enterprise can run with fewer than three gateway or database nodes, but those deployments are not high availability and are not recommended for production. For high availability, use at least three database nodes (one writer and two readers) and at least three gateway instances per cluster. This baseline ensures:- Reader redundancy. If a reader becomes unavailable, the remaining readers and the writer continue serving read traffic. A replacement reader is provisioned automatically.
- Gateway redundancy. Gateway instances are stateless and interchangeable. Loss of one gateway has no impact on availability. Traffic is redistributed across the remaining instances.
- Writer recovery. If the writer becomes unavailable, a new writer process is started and recovers its state from object storage. No committed data is lost. Write availability is restored once the new writer is running.
- Infrastructure redundancy. All supporting services (load balancers, metadata stores, coordination layers) run with redundant infrastructure sized for high-availability clusters, eliminating single points of failure across the cluster.