hive-qdrant

A Qdrant memory backend for the hive ecosystem, built to degrade instead of fail. An unreachable cluster does not turn into an exception at the call site: writes queue, reads answer degraded, and the queue drains when the circuit closes again.

Coordinates

;; deps.edn
io.github.hive-agi/hive-qdrant {:mvn/version "0.1.9"}

What it is

QdrantMemoryStore implements IMemoryStore from hive-spi — the port leaf, not hive-mcp. The published jar therefore carries no host dependency: the core store, queue and migrate namespaces depend only on the ports, and the host integration files (addon.clj, lifecycle.clj) are excluded from both the jar and the pom.

That is the ecosystem rule this library exists to respect: the core owns the ports, an addon owns the adapter, and a new vector backend requires zero edits to the core.

The fail-soft path

Namespace	Role
hive-qdrant.store	IMemoryStore implementation over clj-qdrant
hive-qdrant.circuit	Circuit breaker — :closed (normal), :open (fail fast), :half-open (probing)
hive-qdrant.queue	JVM-wide write queue; on :open mutations enqueue rather than error, reads return a degraded response
hive-qdrant.failure	Closed :error/* taxonomy plus translation to the legacy {:success? false :errors […]} shape
hive-qdrant.config	Typed env wiring — host, port, api-key, TLS. Only the connection surface is operator-tunable
hive-qdrant.deadline	Per-call gRPC deadline on every read RPC: 15 s for retrieve, scroll and count, 60 s for search (QDRANT_READ_DEADLINE_MS, QDRANT_SEARCH_DEADLINE_MS). A stalled store ends the call, and frees its thread, with DEADLINE_EXCEEDED. Writes (upsert, delete, collection and payload index creates) carry 30 s (QDRANT_WRITE_DEADLINE_MS); a write that misses it fails on the circuit and is queued for replay. An unusable value (not a whole number of ms from 1 to 3600000) falls back to the default with one warning and never stops the store from starting
hive-qdrant.connect	What connect! does to qdrant: open the client under a 5 s health check, then ensure the collection and its payload indexes. A store that cannot be reached keeps its client, so the circuit and the queue own its writes; a store that is connected and answers nothing, at the health check or at any probe or create after it, fails connect!
hive-qdrant.migrate	Chroma / Milvus → Qdrant migration, with the extraction step behind a pluggable :source-fn
hive-qdrant.addon	IAddon registration (host-side, not shipped)
hive-qdrant.lifecycle	IShutdownHook, priority 210 — the client band (host-side, not shipped)

When the breaker transitions back to :closed, drain! flushes the queue through a single-writer core.async pipeline, coalescing by (op, id) so only the latest mutation for an entry is replayed.

Multiple instances

The store slot and the addon id are both parameterizable, so several QdrantAddon instances can coexist — carto, kanban, and so on — each backed by its own collection, and each visible to the addon registry as distinct.

License

MIT.