go-store/docs/RFC-STORE.md

go-store RFC — SQLite Key-Value Store

An agent should be able to use this store from this document alone.

Module: dappco.re/go/store Repository: core/go-store Files: 8

---

1. Overview

SQLite-backed key-value store with TTL, namespace isolation, reactive events, and quota enforcement. Pure Go (no CGO). Used by core/ide for memory caching and by agents for workspace state.

---

2. Architecture

File	Purpose
store.go	Core Store: CRUD on (grp, key) compound PK, TTL via expires_at (Unix ms), background purge (60s), text/template rendering, iter.Seq2 iterators
events.go	Watch/Unwatch (buffered chan, cap 16, non-blocking sends) + OnChange callbacks (synchronous)
scope.go	ScopedStore wraps *Store, prefixes groups with namespace:. Quota enforcement (MaxKeys/MaxGroups)
workspace.go	Workspace buffer: DuckDB-backed mutable accumulation, atomic commit to journal
journal.go	InfluxDB journal: write completed units, query time-series, retention
compact.go	Cold archive: compress journal entries to JSONL.gz
store_test.go	Store unit tests
workspace_test.go	Workspace buffer tests

---

3. Key Design Decisions

• Single-connection SQLite. MaxOpenConns(1) because SQLite pragmas (WAL, busy_timeout) are per-connection — a pool would hand out unpragma'd connections causing SQLITE_BUSY
• TTL is triple-layered: lazy delete on Get, query-time WHERE filtering, background purge goroutine
• LIKE queries use escapeLike() with ^ as escape char to prevent SQL wildcard injection

---

4. Store Struct

// Store is the SQLite KV store with optional InfluxDB journal backing.
type Store struct {
    db      *sql.DB            // SQLite connection (single, WAL mode)
    journal influxdb2.Client   // InfluxDB client (nil if no journal configured)
    bucket  string             // InfluxDB bucket name
    org     string             // InfluxDB org
    mu      sync.RWMutex
    watchers map[string][]chan Event
}

// Event is emitted on Watch channels when a key changes.
type Event struct {
    Group string
    Key   string
    Value string
}

// New creates a store. Journal is optional — pass WithJournal() to enable.
//
//   st, _ := store.New(":memory:")                           // SQLite only
//   st, _ := store.New("/path/to/db", store.WithJournal(    // SQLite + InfluxDB
//       "http://localhost:8086", "core-org", "core-bucket",
//   ))
func New(path string, opts ...StoreOption) (*Store, error) { }

type StoreOption func(*Store)

func WithJournal(url, org, bucket string) StoreOption { }

---

5. API

st, _ := store.New(":memory:")      // or store.New("/path/to/db")
defer st.Close()

st.Set("group", "key", "value")
st.SetWithTTL("group", "key", "value", 5*time.Minute)
val, _ := st.Get("group", "key")    // lazy-deletes expired

// Iteration
for key, val := range st.AllSeq("group") { ... }
for group := range st.GroupsSeq() { ... }

// Events
ch := st.Watch("group")
st.OnChange(func(event store.Event) { ... })

---

6. ScopedStore

// ScopedStore wraps a Store with a namespace prefix and optional quotas.
//
//   scoped := store.NewScoped(st, "mynamespace")
//   scoped.Set("key", "value")           // stored as group "mynamespace:default", key "key"
//   scoped.SetIn("mygroup", "key", "v")  // stored as group "mynamespace:mygroup", key "key"
type ScopedStore struct {
    store     *Store
    namespace string  // validated: ^[a-zA-Z0-9-]+$
    MaxKeys   int     // 0 = unlimited
    MaxGroups int     // 0 = unlimited
}

func NewScoped(st *Store, namespace string) *ScopedStore { }

// Set stores a value in the default group ("namespace:default")
func (ss *ScopedStore) Set(key, value string) error { }

// SetIn stores a value in an explicit group ("namespace:group")
func (ss *ScopedStore) SetIn(group, key, value string) error { }

// Get retrieves a value from the default group
func (ss *ScopedStore) Get(key string) (string, error) { }

// GetFrom retrieves a value from an explicit group
func (ss *ScopedStore) GetFrom(group, key string) (string, error) { }

• Namespace regex: ^[a-zA-Z0-9-]+$
• Default group: Set(key, value) uses literal "default" as group, prefixed: "mynamespace:default"
• SetIn(group, key, value) allows explicit group within the namespace
• Quota: MaxKeys, MaxGroups — checked before writes, upserts bypass

---

7. Event System

• Watch(group string) <-chan Event — returns buffered channel (cap 16), non-blocking sends drop events
• Unwatch(group string, ch <-chan Event) — remove a watcher
• OnChange(callback) — synchronous callback in writer goroutine
• Deadlock warning: notify() holds s.mu read-lock — calling Watch/Unwatch/OnChange from inside a callback will deadlock

---

8. Workspace Buffer

Stateful work accumulation over time. A workspace is a named DuckDB buffer for mutable work-in-progress. When a unit of work completes, the full state commits atomically to a time-series journal (InfluxDB). A summary updates the identity store (the existing SQLite store or an external database).

7.1 The Problem

Writing every micro-event directly to a time-series makes deltas meaningless — 4000 writes of "+1" produces noise. A mutable buffer accumulates the work, then commits once as a complete unit. The time-series only sees finished work, so deltas between entries represent real change.

7.2 Three Layers

Store (SQLite): "this thing exists"     — identity, current summary
Buffer (DuckDB): "this thing is working" — mutable temp state, atomic
Journal (InfluxDB): "this thing completed" — immutable, delta-ready

Layer	Store	Mutability	Lifetime
Identity	SQLite (go-store)	Mutable	Permanent
Hot	DuckDB (temp file)	Mutable	Session/cycle
Journal	InfluxDB	Append-only	Retention policy
Cold	Compressed JSONL	Immutable	Archive

7.3 Workspace API

// Workspace is a named DuckDB buffer for mutable work-in-progress.
// It holds a reference to the parent Store for identity updates and journal writes.
//
//   ws, _ := st.NewWorkspace("scroll-session-2026-03-30")
//   ws.Put("like", map[string]any{"user": "@handle", "post": "video_123"})
//   ws.Commit()  // atomic → journal + identity summary
type Workspace struct {
    name  string
    store *Store   // parent store for identity updates + journal config
    db    *sql.DB  // DuckDB via database/sql driver (temp file, deleted on commit/discard)
}

// NewWorkspace creates a workspace buffer. The DuckDB file is created at .core/state/{name}.duckdb.
//
//   ws, _ := st.NewWorkspace("scroll-session-2026-03-30")
func (s *Store) NewWorkspace(name string) (*Workspace, error) { }

// Put accumulates an entry in the workspace buffer. Returns error on write failure.
//
//   err := ws.Put("like", map[string]any{"user": "@handle"})
func (ws *Workspace) Put(kind string, data map[string]any) error { }

// Aggregate returns a summary of the current workspace state
//
//   summary := ws.Aggregate()  // {"like": 4000, "profile_match": 12}
func (ws *Workspace) Aggregate() map[string]any { }

// Commit writes the aggregated state to the journal and updates the identity store
//
//   result := ws.Commit()
func (ws *Workspace) Commit() core.Result { }

// Discard drops the workspace without committing
//
//   ws.Discard()
func (ws *Workspace) Discard() { }

// Query runs SQL against the buffer for ad-hoc analysis.
// Returns core.Result where Value is []map[string]any (rows as maps).
//
//   result := ws.Query("SELECT kind, COUNT(*) as n FROM entries GROUP BY kind")
//   rows := result.Value.([]map[string]any)  // [{"kind": "like", "n": 4000}]
func (ws *Workspace) Query(sql string) core.Result { }

7.4 Journal

Commit writes a single point per completed workspace. One point = one unit of work.

// CommitToJournal writes aggregated state as a single InfluxDB point.
// Called by Workspace.Commit() internally, but exported for testing.
//
//   s.CommitToJournal("scroll-session", fields, tags)
func (s *Store) CommitToJournal(measurement string, fields map[string]any, tags map[string]string) core.Result { }

// QueryJournal runs a Flux query against the time-series.
// Returns core.Result where Value is []map[string]any (rows as maps).
//
//   result := s.QueryJournal(`from(bucket: "core") |> range(start: -7d)`)
//   rows := result.Value.([]map[string]any)
func (s *Store) QueryJournal(flux string) core.Result { }

Because each point is a complete unit, queries naturally produce meaningful results without complex aggregation.

7.5 Cold Archive

When journal entries age past retention, they compact to cold storage:

// CompactOptions controls cold archive generation.
type CompactOptions struct {
    Before time.Time // archive entries before this time
    Output string    // output directory (default: .core/archive/)
    Format string    // gzip or zstd (default: gzip)
}

// Compact archives journal entries to compressed JSONL
//
//   st.Compact(store.CompactOptions{Before: time.Now().Add(-90*24*time.Hour), Output: "/archive/"})
func (s *Store) Compact(opts CompactOptions) core.Result { }

Output: gzip JSONL files. Each line is a complete unit of work — ready for training data ingestion, CDN publishing, or long-term analytics.

7.6 File Lifecycle

DuckDB files are ephemeral:

Created:   workspace opens → .core/state/{name}.duckdb
Active:    Put() accumulates entries
Committed: Commit() → journal write → identity update → file deleted
Discarded: Discard() → file deleted
Crashed:   Orphaned .duckdb files detected on next New() call

Orphan recovery on New():

// New() scans .core/state/ for leftover .duckdb files.
// Each orphan is opened, aggregated, and discarded (not committed).
// The caller decides whether to commit orphan data via RecoverOrphans().
//
//   orphans := st.RecoverOrphans(".core/state/")
//   for _, ws := range orphans {
//       // inspect ws.Aggregate(), decide whether to commit or discard
//       ws.Discard()
//   }
func (s *Store) RecoverOrphans(stateDir string) []*Workspace { }

---

9. Reference Material

Resource	Location
Core Go RFC	code/core/go/RFC.md
IO RFC	code/core/go/io/RFC.md