0
Rust foundation
Stable types, schema, parser, HIR, and relational IR.
Roadmap
What is implemented,
and what is not.
Development follows a vertical sequence. 62 items are complete. The numbered engine sequence ends at 73; later coordinator, columnar, and schema-evolution work is also complete. Serializable isolation, concurrent writers, and MCP remain planned or deferred.
Complete
1
Storage foundation
4 KiB slotted pages, bounded buffer pool, heap insert/scan, and reopen.
2A
Transactions + WAL core
Transaction lifecycle, versioned WAL, LSN / pageLSN, and durable commit.
2B
Crash recovery
Startup analysis, repeat-history redo, reverse undo, and crash-tail handling.
2B.1
Single writer + runtime rollback
Lazy writer ownership, retryable commit/rollback, and before-image undo.
2C
Checkpoint + WAL lifecycle
Quiescent checkpoints, monotonic logical LSNs, and bounded WAL generation recycling.
3A
Typed expressions + NULL
Contextual NULL typing, three-valued logic, and explicit IS NULL.
3B
Typed DML
INSERT / UPDATE / DELETE, stable RowIds, and affected-row results.
3C
INNER JOIN
Qualified columns, aliases, self joins, and nested-loop execution.
3C.5
Schema fingerprints
Canonical table encoding, SHA-256 fingerprints, and identity checks on open.
integrity
Page and WAL integrity
Page v4/v5 CRC32C, WAL v4 checksums, and decoder fuzzing.
3D
Aggregates + sort
ORDER BY, global aggregates, and in-memory GROUP BY.
4A
Versioned RowId
Page v5 slot generations, tombstone reuse, and stale-locator detection.
4B
Heap reuse + safe relocation
Deterministic first-fit and UPDATE relocation that returns the current RowId.
4C1
Persistent B+Tree
Transactional create/insert/point lookup, mixed pages, and split recovery.
4C2
B+Tree delete / rebalance
Exact (key, RowId) delete, merge-only rebalance, and root collapse.
4D1
Index catalog + backfill
Persistent IndexCatalog and transactional create_index backfill.
4D2
Heap / index DML maintenance
Registered indexes stay consistent with heap and SQL DML.
4E
Point IndexScan
Equality and IS NULL predicates can select a point IndexScan.
4F
ANALYZE and costing
Explicit ANALYZE snapshots and deterministic point access-path costs.
5A
Protocol v1
Language-neutral binary framing. Frozen; current servers reject v1 frames.
5B
netbadbd server
Manifest bootstrap, TCP sessions, and a dedicated database worker.
5C
TLS and authorization
Mutual TLS, certificate principals, and per-table operation scopes.
6A
Go Protocol client
Independent Go Protocol v2 client and generated typed bindings.
6C
Rust remote SDK
Synchronous netbadb-sdk::remote client over Protocol v2.
6D
Inspection CLI
Offline catalog and statement inspection; current machine contract is Inspection JSON v7.
6E1
SQL diagnostics LSP
Diagnostics-only language server over SDK Schema Spec v1 or v2.
7A
Performance baseline
Reproducible warm-cache benchmarks without changing engine behavior.
7B
Range IndexScan
Costed bounded Int64/UInt64 range scans.
7C
Predicate-first NestedLoopJoin
Rejected join pairs allocate no combined row.
7D
Costed HashJoin
Analyzed Scan × Scan INNER JOIN may select a simple equi HashJoin.
7E
Validate-once heap scan
Each immutable heap page is fully validated once per sequential scan.
7F
Join predicate prebinding
NLJ and HashJoin bind column positions once before candidate loops.
7G
Borrowed join evaluation
Join predicates borrow column and literal scalars; only computed results are owned.
7H
Inequality bound rejection
NestedLoopJoin skips left probes that cannot match any right row.
7I
Adaptive inequality sweep
Exact candidate counting can replace a full NestedLoopJoin inner loop.
7J
Required-column decode
Query plans decode only required heap columns; DML still uses full rows.
7K
Move-aware projection
Identity and unique projections move owned values instead of cloning them.
7L
Direct COUNT(column) scan
A lone global COUNT(column) over SeqScan counts presence without materializing rows.
7M
Direct multi-COUNT scan
One exact heap pass shares live-row counts across multiple COUNT outputs.
7N-7V
Filter and COUNT streaming
Borrowed Filter evaluation, streamed SeqScan/Filter/Project, and COUNT specializations.
51
Multi-storage atomic commit
WAL Prepare plus an independent coordinator log for two or more local write storages.
52
Range partitions
PartitionCatalog v1 RANGE partitioning with exact pruning and atomic row movement.
53
LSM storage
Synchronous LSM table storage with MVCC, SSTables, flush, and leveled compaction.
55
Batch execution
Bounded 256-row batches for SeqScan, Filter, Project, Limit, Aggregate, and HashJoin probe.
isolation
Read Committed and Repeatable Read
Explicit isolation handles; implicit statements use Read Committed. Serializable is not available.
72
Index nested-loop join
Analyzed Scan × Scan INNER equality joins may point-probe an ordered right index.
73
IndexJoin cost calibration
Heap point-probe costs stay aligned with managed_page_count SeqScan comparison.
R17
Runtime schema catalog
Database::open_catalog reconstructs the full schema without external TableDefs.
R19
SQL CREATE TABLE
Transactional Heap CREATE TABLE over native SQL.
R21
SQL DROP TABLE
Exact prepared DROP TABLE name over the native SQL frontend.
R7
SQL CREATE / DROP INDEX
Single-column non-unique Heap BTree CREATE INDEX and DROP INDEX.
PTv2
Physical Types v2
Exact-width integers, Float32/Float64, and Bytes; no implicit numeric widening.
5A2
Protocol v2
Current native wire contract. Servers and Rust/Go clients speak v2 only.
R26
SQL ALTER TABLE
Heap rename, nullable ADD, restricted DROP, and SET/DROP NOT NULL over native SQL.
3A-vis
Global snapshot visibility
Gap-free DatabaseCommitSeq and a published Heap/LSM visibility vector. Repeatable Read pins the first vector.
3B-3F
Coordinator pipeline and group commit
Synchronous pipelining, log compaction, and explicit group commit with batched Prepare/Commit/Change Stream barriers.
Col
Derived columnar projections
Opt-in ColumnarScan, change stream, Base+Delta, compaction, lazy I/O, and caller-driven maintenance. Heap and LSM remain authoritative.
Adapt
Adaptive operations
Caller-driven inspection and one-action maintenance. No background scheduler or timer.
M11
Deployment manifest v11
Current netbadbd startup contract. Manifests v1–v10 are rejected.
NBOP7
NBOP v7 operator plane
Local Unix-domain operator protocol for status, Adaptive controls, and explicit physical-design approval.
PG-rm
PostgreSQL wire removed
netbadb-pgwire and netbadbd --postgres were deleted. Native Protocol v2 is the only database network frontend.
next
Serializable isolation
Serializable isolation, concurrent writers, and broader join enumeration remain later work.
6E2
MCP adapter
Deferred: official MCP SDKs currently exceed the Rust 1.85 MSRV.
7+
Further optimization
General ALTER COLUMN TYPE, one-sided/Text range costing, and broader HashJoin eligibility.
Not included in the current release
Serializable isolation, concurrent writers, and cross-process file locking
PRIMARY KEY / UNIQUE / FK constraints, IF EXISTS, and index-only scans
General ALTER COLUMN TYPE, LSM/partitioned DDL composition, and arbitrary ALTER after DML
Full SQL: outer joins, subqueries, HAVING, DISTINCT, window functions
External sort spill, PostgreSQL wire compatibility, and complete information_schema
MCP adapter (deferred past MSRV), SQL EXPLAIN, network CDC, and on-disk format migrations