Roadmap
Thermocline v4.0 is an AI-native document database with built-in vector search and automatic hot/cold storage tiering. Development is organized into six sequential phases spanning approximately 17 months.
This roadmap is subject to change based on community feedback and adoption patterns.
Key Capabilities at v4.0 GA
- Native LSM-tree storage engine with WAL and crash recovery
- Automatic hot/cold data tiering (NVMe to Parquet on object storage)
- Built-in vector search (HNSW + DiskANN, up to 1 billion vectors)
- ACID transactions with snapshot isolation (>50K TPS)
- Raft consensus replication (<5s failover)
- MVCC time travel queries
- WAL-based change streams
- MongoDB wire protocol compatibility (all drivers, zero app changes)
- SSPL v1 license
Phase 1: Foundation (4 months) — In Development
Core storage engine and wire protocol.
- Write-Ahead Log (WAL) with configurable sync modes
- LSM-tree memtable and SST file management
- BSON document encoding/decoding
- Leveled and universal compaction strategies
- MongoDB wire protocol (OP_MSG, OP_QUERY, OP_GET_MORE)
- Basic CRUD operations (insert, find, update, delete)
- Gateway with TLS termination and SCRAM-SHA-256 auth
- etcd-based metadata store
- S3 cold storage adapter
- Helm chart v0.1
- Docker Compose for local development
- Crash recovery from WAL replay
Exit criteria: CRUD tests pass, wire protocol compatible with mongosh and all official MongoDB drivers, crash recovery verified, point read <1ms p50.
Phase 2: Query Engine (3 months)
Full MQL support and cold storage queries.
- All MongoDB query operators on native storage
- MQL-to-DataFusion translation for cold queries
- Full aggregation pipeline ($match, $group, $sort, $project, $lookup, $unwind, $facet, etc.)
- Tiered query execution with result merging
- Archival pipeline (BSON to Parquet with zone maps and bloom filters)
- B-tree indexes with online builds
- Cursor management and streaming results
- Predicate pushdown and column pruning for Parquet
- CollectionPolicy CRD for archival rules
Exit criteria: >95% MongoDB query compatibility, correct tiered query results, valid Parquet output, cold query on 1M docs <2s p50.
Phase 3: Transactions & Replication (3 months)
ACID transactions and Raft consensus.
- MVCC version store with configurable retention
- Transaction manager (begin, commit, abort)
- Lock manager with deadlock detection (wait-for graph)
- Snapshot isolation and read committed levels
- Raft consensus (leader election, log replication)
- Membership changes (add/remove nodes)
- Snapshot transfer for new replicas
- Automatic failover with <5s target
- Read replicas with bounded staleness
Exit criteria: ACID verified under concurrent load, Raft failover <5s with zero committed data loss, >50K TPS on 3-node cluster, deadlock resolution <100ms.
Phase 4: AI Features (3 months)
Vector search and hybrid queries.
- HNSW index for hot vectors (in-memory, <50M vectors)
- Distance metrics: cosine, euclidean (L2), dot product
-
$vectorSearchaggregation stage (Atlas-compatible syntax) - Hybrid search (pre-filter and post-filter modes)
- SQ8 scalar quantization (~4x memory reduction)
- Cold vector search on archived Parquet embeddings
- Online index builds without blocking writes
- Vector index persistence and recovery
Exit criteria: >95% recall at <50ms p50 for 1M vectors, hybrid search correctness, Atlas syntax compatibility, SQ8 <2% recall loss.
Phase 5: Advanced Features (2 months)
Time travel, change streams, and sharding.
- Time travel queries (
readConcern.atClusterTime) - Named snapshots (pin a point-in-time)
- Configurable version retention (1-365 days)
- WAL-based change streams with resume tokens
- Change stream filtering and projection
- RAG embedding triggers (auto-generate embeddings on write)
- Hash-based sharding (up to 1024 shards)
- Shard-aware query routing
Exit criteria: Correct historical reads, change stream ordering and resume, even shard distribution (<10% skew), RAG trigger overhead <100ms.
Phase 6: Scale & Polish (2 months)
Billion-scale vectors, migration tooling, production hardening.
- DiskANN index for billion-scale vectors (disk-resident)
- Scale testing (100TB hot, 1PB cold, 1B vectors)
- MongoDB migration CLI (
thermocline-cli migrate) - Migration tested against MongoDB 4.4-7.0
- Full documentation
- Kubernetes operator with CRDs
- Grafana dashboards
- Production hardening and benchmarks
- Product quantization for vectors
- Multi-tenant resource isolation
Exit criteria: 100M vector search <200ms p50, 100TB hot data stable 72+ hours, migration from MongoDB 4.4-7.0, docs complete, K8s operator manages full lifecycle.
Post-v4.0 Future
| Feature | Priority |
|---|---|
| Range sharding | High |
| GPU-accelerated vector search | Medium |
| Row-level security | Medium |
| Full-text search | Medium |
| Graph queries ($graphLookup) | Medium |
| Geospatial queries | Medium |
| Time series collections | Medium |
| Multi-region replication | Medium |
| Predictive ML-based tiering | Low |
| Column-level encryption | Low |
Driver Compatibility
All official MongoDB drivers work with Thermocline using standard connection strings — no modifications needed:
| Driver | Status |
|---|---|
| Node.js / Mongoose | Supported |
| Python / PyMongo / Motor | Supported |
| Java | Supported |
| Go | Supported |
| Rust | Supported |
| C# / .NET | Supported |
| C / C++ | Supported |
| Ruby | Supported |
| PHP | Supported |
How to Influence the Roadmap
- Vote on issues — Use reactions on GitHub issues to show interest
- Open feature requests — Describe your use case and requirements
- Contribute — Implementation PRs are always welcome
- Discuss — Join GitHub Discussions to share your perspective
Release Cadence
- Patch releases — As needed for bug fixes and security patches
- Minor releases — Every 2-3 months with new features
- Major releases — When breaking changes are necessary (rare)