tidaldb/docs/planning/milestone-1/phase-2/OVERVIEW.md

Milestone 1, Phase 2: Write-Ahead Log

Status: COMPLETE

Phase Deliverable

A durable, append-only signal event log. Every signal write (view, like, skip, completion) is appended to the WAL before any aggregation occurs. Signal aggregates, decay scores, and windowed counts are derived state — the WAL is the source of truth. Group commit amortizes fsync cost across concurrent writers. Content-addressed events via per-event BLAKE3 hash for deduplication. Crash recovery scans forward from last checkpoint and truncates corrupted tails.

Acceptance Criteria

• Batch-oriented wire format: 64-byte cache-aligned header (magic 0x54494C44, version, event count, first sequence number, batch timestamp, payload length, BLAKE3 checksum) followed by tightly-packed 21-byte event records (entity_id u64 LE, signal_type u8, weight f32 LE, timestamp u64 LE)
• BLAKE3 hash covers header[0..32] || all_event_bytes — corrupted batches detected at recovery
• Group commit: dedicated writer thread via crossbeam::channel::bounded(10_000) with recv_deadline; batch fills at 100 events or 10ms timeout, whichever comes first; one fsync per batch
• Segment files: 16 MB rotation, named wal-{first_seq:020}.seg; list_segments() returns ordered list
• Two-phase crash recovery: Phase 1 — verify magic and payload bounds; Phase 2 — verify BLAKE3; truncate at first invalid batch boundary
• WalHandle::open() returns (handle, replayed_events) — caller gets events since last checkpoint for signal materializer replay
• Sequence numbers are monotonically increasing u64, starting at 1; persist across close/reopen
• Deduplication via double-buffered HashSet<u128> (first 128 bits of per-event BLAKE3); 30-second rotation window; duplicate returns Ok(0)
• WalHandle::checkpoint(seq) writes checkpoint.meta atomically with last-materialized sequence number and timestamp
• WalHandle::truncate_before(seq) dispatches to writer thread (no race with segment writes); deletes segments whose last sequence < seq
• WalHandle::shutdown() flushes remaining events, fsyncs, and joins writer thread
• WalHandle implements Drop for best-effort shutdown
• #![forbid(unsafe_code)] — entirely safe Rust; crossbeam unsafe is in the dependency, not the WAL code
• cargo fmt clean, cargo clippy -D warnings clean

Dependencies

• Requires: m1p1 (types: EntityId, Timestamp encoding patterns) — WAL uses u64 entity IDs and nanosecond timestamps directly
• Blocks: m1p4 (Signal Ledger — WAL replay feeds the materializer; WalHandle is SignalLedger's durability backend)

Research References

• docs/research/tidaldb_wal.md — batch-oriented format (Section 1, Approach 3), group commit with crossbeam (Section 3, Pattern 4), BLAKE3 + length-prefix crash detection (Section 4, Approach 3), segment rotation (Section 5), bounded sliding window dedup (Section 6, Approach 3), full implementation blueprint
• thoughts.md — Part II.1 (WAL convergence), Part V.5 (quarantine-first), Part V.6 (group commit)

Spec References

• CODING_GUIDELINES.md — Section 7 (error handling), Section 10 (dependency policy for crossbeam)

Task Index

#	Task	Delivers	Depends On	Complexity	Status
01	WAL Wire Format and Segment Files	BatchHeader, EventRecord, SegmentWriter, WalError	None	M	COMPLETE
02	Group Commit Writer	WriterConfig, WalCommand, run_writer loop	Task 01	M	COMPLETE
03	Crash Recovery and Replay	WalReader, recover(), partial-write truncation	Task 01	M	COMPLETE
04	Deduplication, Checkpoint, and Public API	DedupWindow, CheckpointManager, WalHandle, SignalEvent	Task 02, Task 03	M	COMPLETE

Task Dependency DAG

Task 01: Wire Format + Segment Files
    |
    +-------------------------------+
    |                               |
    v                               v
Task 02: Group Commit Writer    Task 03: Crash Recovery + Replay
    |                               |
    +---------------+---------------+
                    |
                    v
    Task 04: Dedup + Checkpoint + WalHandle (Public API)

Tasks 02 and 03 are parallelizable — both depend only on Task 01's types.

File Layout

tidal/src/
  wal/
    mod.rs          -- Task 04: WalHandle, WalConfig, SignalEvent (public API)
    format.rs       -- Task 01: BatchHeader, EventRecord encode/decode
    segment.rs      -- Task 01: SegmentWriter, list_segments
    error.rs        -- Task 01: WalError enum
    writer.rs       -- Task 02: WalCommand, WriterConfig, run_writer
    reader.rs       -- Task 03: WalReader, RecoveryResult, recover()
    dedup.rs        -- Task 04: DedupWindow
    checkpoint.rs   -- Task 04: CheckpointManager
  lib.rs            -- pub mod wal (added)

Open Questions (Resolved)

1. oneshot channels — Resolved: used crossbeam::channel::bounded(1) per-append as the reply channel. Zero additional dependencies.

2. Segment pre-allocation — Resolved: not implemented in m1p2. Deferred until disk write performance becomes a measured bottleneck.

3. WAL compression — Resolved: deferred. At 10K events/sec the write rate (~210 KB/sec) is nowhere near a disk bandwidth constraint.

4. Multi-batch fsync — Resolved: single fsync per batch (as designed). The 10ms timeout at low write rates makes multi-batch accumulation unnecessary.

5. Interaction with fjall WAL — Resolved: the two WALs are independent. tidalDB's signal WAL sits in {dir}/wal/; fjall's internal journal sits in the fjall keyspace directory. Recovery order: signal WAL replay → signal state reconstruction → fjall entity store (no cross-dependency in crash recovery).