Dragonfly 2.0, the open-source in-memory data store positioning itself as a high-performance alternative to Redis and Valkey, has reached major release status. The project team says the release marks production-ready maturity rather than introducing headline features.
Key changes include: native RDB import from Valkey 9 (including hashes with mixed persistent and expiring fields), the new GEOSEARCHSTORE command for geospatial queries with sorting and ACL checks, and configurable replication backlog retention by age and memory budget.
On the memory management front, Dragonfly now updates per-connection telemetry only per parsing cycle instead of per pipeline command, shrinks client input buffers after periods of low utilization (default max drops from 64 to 32 KiB), and releases oversized RDB serialization buffers after writing. New defragmentation phases and a shared RESP2 read buffer lay groundwork for lower memory overhead at scale.
The release also hardens the RESTORE command against malformed inputs — rejecting duplicate hash fields and data lengths exceeding remaining input — and fixes numerous issues in Streams, consumer groups, RESP3 responses, cluster migrations, tiered storage, and compressed lists.
