Grafana Mimir vs nodbi

nodbi presents a single document-store interface — docdb_create, docdb_query, docdb_update — over SQLite, DuckDB, PostgreSQL, MongoDB, CouchDB and Elasticsearch. The engineering reality behind that abstraction is that each backend's JSON support keeps moving, and the releases show it: jsonb_tree adopted as RSQLite 2.4.4 exposes it, json_tree reworked for DuckDB 1.3.0, then avoided entirely for DuckDB listfields because it was too slow. The 0.11.0 release in late 2024 is the one that changed the contract, making docdb_query() return columns of a single consistent type.

Two threads dominate. The first is performance, pursued backend by backend: fast direct NDJSON import moved from DuckDB-only to SQLite and PostgreSQL, query refactors chasing each DuckDB release, and the removal of expensive tree-walking where a cheaper path exists. The second is making results predictable — consistent column types, version checks on the database backend, clearer messages when a Postgres database does not exist yet or when column names contain the dots nodbi reserves for JSON paths.

Given that most recent releases are triggered by DuckDB and RSQLite version changes, the next one likely follows the same pattern — adopting a new JSON function or working around a slow one. The duplicate-_id handling added in 0.14.0 suggests NDJSON ingestion edge cases are the current active area.