Roadmap
The milestones and tracks, what each one delivers, and its status.
Loam is built engine first: hybrid retrieval with the Qdrant, Elasticsearch and Flight SQL surfaces (M1), then production hardening (M2). Together they are v1.0. Loam Live (track R) and Loam Durable (track D) are built in parallel. The roadmap page is the short version, grouping M5 and M6; the tables below follow the engine's design roadmap and name more of each milestone's scope.
Compatibility is defined by conformance suites, not feature lists: a milestone closes when the named client libraries and framework integrations pass unmodified.
Available means on the engine's main branch with its gates passing; in progress means partly on main or built on a branch that is landing; planned means designed or proposed.
Engine
| Milestone | Scope | Status |
|---|---|---|
| M0 Foundation | Metastore on openraft, object-store I/O with fault injection, the internal log, cache, workers, links with exactly-once apply, primary-key index, GC, simulation harness | Available |
| M1.1 Collection storage | Lance + Tantivy under one manifest, upserts and deletes, sparse vectors, collection GC | Available |
| M1.2 Query engine | The pluggable MetaStore trait, DataFusion operators, vector / BM25 / sparse retrieval, fusion, the tail merge, native REST API, Flight SQL and DoPut ingest | Available |
| M1.3 Hot tier and routing | foyer NVMe cache, qdrant-edge HNSW hot tier, pinned splits, affinity routing, write backpressure | Available |
| M1.4 Qdrant API | Qdrant REST and gRPC | Available |
| M1.5 Elasticsearch API | The subset the LangChain, LlamaIndex and BEIR suites use | In progress |
| M1.6 SDKs and MCP | Python and TypeScript SDKs, an MCP server | In progress |
| M1.7 Exit gates | LangChain and LlamaIndex vector-store tests; BEIR nDCG@10 within 1 point of Elasticsearch BM25; Recall@10 within 1% of Qdrant; ADBC drivers; identical results with the hot tier on or off. Then the rename to Loam | Planned |
| M2 Production hardening = v1.0 | AuthN/Z, tenant quotas, the native stream API with idempotent producers, OTLP logs, Prometheus/OTel, multi-node, more metastore backends, Kubernetes operator | Planned |
| M2.x Cloud and BYOC = v1.1 | Hosted control plane, OpenFGA, SSO, sharding, billing | Planned |
| M3 Graph | Native graph expansion over collections, and over Iceberg tables once M4 lands (CSR sidecars, graph_expand) | Planned |
| M4 Analytics | Iceberg tables through Lakekeeper, read by DuckDB, Trino and Spark | Planned |
| M5 Streams | Stream replay, changelog streams and ecosystem integrations | Planned |
| M6 Scale | The quorum WAL and a sharded metastore | Planned |
The Kafka wire protocol is deferred past v1.0; Neo4j Bolt/Cypher and the ClickHouse HTTP interface are dropped (decisions D42–D45).
Loam Live and Loam Durable
| Track | Scope | Status |
|---|---|---|
| R1 TiKV metastore and reactive core | TiKV metastore, transactions, commit journal, read-set subscriptions (on main); QuickJS server functions, the connect-rust sync API and the TypeScript client (landing) | In progress |
| R2–R4 | Router and point-in-time restore; the collections bridge (change feed into collections); Helm and BYOC | Planned |
| D1 Embedded Resonate | The Resonate server in the binary with SQLite and TiDB backends, an in-process runtime and a linearizability run (on main); the operations API and bulk import (landing) | In progress |
| D2–D4 | TiKV backend and per-namespace tenancy; durable agents; the connect-rust transport | Planned |
Proposed
Serverless functions billed on CPU time (workerd, wasmtime, gVisor), tenant secrets through Dapr's secrets building block, jobs on Loam (Celery, BullMQ v6, PySpark on Sail, Flink SQL on RisingWave; self-hosted clusters need TiKV), the Postgres wire protocol (PG1), self-hosting with GitOps (RustFS, the TiKV operator, Helm, a Loam operator), and the Loam Commons showcase suite (§22). None of these are built. Postgres on your bucket is under evaluation, with no MySQL planned; a branch per agent workspace is an exploration.
M0 results
All three M0 exit gates pass. The full evidence, including the 336-cell object-store fault matrix and the crash-gate runs, is in the M0 exit report.
For every gate, the testing strategy and the risk register, read §12 Roadmap, testing and risks.