X-Git-Url: https://git.giacomolli.uk/onechat/blobdiff_plain/9f17bfb1173524553fff5301b0664267b430e6e3..6e3e9c5c10f3d35bae7e9846334974a09418aef7:/memory.go?ds=sidebyside

diff --git a/memory.go b/memory.go
index 0be2fde..0fd7d83 100644
--- a/memory.go
+++ b/memory.go
@@ -7,41 +7,42 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"strings"
 	"sync"
 	"time"
 )
 
-// errNoView tells that a chat has messages but no saved view.
-var errNoView = errors.New("the chat has messages but no saved view")
+var (
+	// errMemoryNoView tells that a chat has messages but no saved view.
+	errMemoryNoView = errors.New("the chat has messages but no saved view")
 
-// Memory is the memory of one chat: the message log, the summary tree, and the
-// view. It saves the view to view.json after each change and loads it at start.
-// It never builds the view again from the log, because a new view differs from
-// the live one and so loses the prompt cache.
-type Memory struct {
-	mutex sync.Mutex
-	root  string
-	log   *Log
-	tree  *Tree
-	chat  *View
-}
-
-// viewFile is the content of view.json.
-type viewFile struct {
-	Chat viewState `json:"chat"`
-}
+	// errMemoryUnsettled tells that a line of the view stays pending,
+	// because its build failed.
+	errMemoryUnsettled = errors.New("a summary of the view failed, and it is tried again at the next message")
+)
 
-// openMemory opens the chat in root. The clock gives the dates of new messages
-// and the days of new nodes.
-func openMemory(ctx context.Context, root string, clock func() time.Time) (*Memory, error) {
+// openMemory opens the chat in root and starts the builds that are ready. The
+// clock gives the dates of new messages and the days of new nodes. The builds
+// stop when ctx ends or at Close.
+func openMemory(ctx context.Context, root string, clock func() time.Time, summarizer Summarizer) (*Memory, error) {
 	log, err := openLog(ctx, root, clock)
 	if err != nil {
 		return nil, err
 	}
 
 	m := &Memory{
-		root: root,
-		log:  log,
+		root:       root,
+		log:        log,
+		queue:      newBuildQueue(),
+		summarizer: summarizer,
+		changed:    make(chan struct{}),
+	}
+
+	m.ctx, m.cancel = context.WithCancel(ctx)
+
+	m.syncer, err = openSyncer(root)
+	if err != nil {
+		return nil, errors.Join(err, m.Close())
 	}
 
 	m.tree, err = openTree(ctx, root, log.Len(), clock)
@@ -54,141 +55,398 @@ func openMemory(ctx context.Context, root string, clock func() time.Time) (*Memo
 		return nil, errors.Join(fmt.Errorf("load view: %w", err), m.Close())
 	}
 
+	err = m.restoreQueue()
+	if err != nil {
+		return nil, errors.Join(fmt.Errorf("restore build queue: %w", err), m.Close())
+	}
+
+	err = m.catchUp()
+	if err != nil {
+		return nil, errors.Join(err, m.Close())
+	}
+
+	m.mutex.Lock()
+	m.syncer.Written()
+	m.dispatch()
+	m.mutex.Unlock()
+
 	return m, nil
 }
 
-// load reads the saved view. A crash can stop an append after the log write and
-// before the view save, so load continues the view with the messages that it
-// does not cover yet.
-func (m *Memory) load() error {
-	data, err := os.ReadFile(m.viewPath())
-	if errors.Is(err, os.ErrNotExist) && m.log.Len() == 0 {
-		m.chat = newView(chatHigh, chatLow)
-		return nil
+// Summarizer builds the text of one node with a model.
+type Summarizer interface {
+	Summarize(ctx context.Context, task Task) (string, error)
+}
+
+// Memory is the memory of one chat: the message log, the summary tree, the chat
+// view, the compaction view, and the queue of the builds. The compaction view
+// is the chat view merged further, at the limits of 16 to 32 KB. It takes the
+// same new lines, and it merges again in one batch when it is too large or when
+// the chat view merges. So the compactions share its prefix in the prompt
+// cache. It saves the views to view.json after each change and loads them at
+// start. It never builds the views again from the log, because a new view
+// differs from the live one and so loses the prompt cache.
+//
+// Each write goes to the operating system at once, and a syncer makes the
+// writes durable by group commit in the background. Flush waits until the
+// writes so far are durable.
+//
+// The summarizer builds the nodes in the queue in the background. The start
+// reads the indexes once to fill the queue again. After that, only a new
+// message or a new node adds work. Memory keeps the nodes of the views and the
+// tasks of the running builds. It reads all other messages and nodes from their
+// day files through the indexes.
+type Memory struct {
+	mutex      sync.Mutex
+	root       string
+	log        *Log
+	syncer     *Syncer
+	tree       *Tree
+	chat       *View
+	cview      *View
+	queue      *BuildQueue
+	summarizer Summarizer
+	ctx        context.Context
+	cancel     context.CancelFunc
+	workers    sync.WaitGroup
+	changed    chan struct{}
+}
+
+// Close stops the builds, waits for them, makes the writes durable, closes the
+// files of the chat, and releases its lock.
+func (m *Memory) Close() error {
+	m.cancel()
+	m.workers.Wait()
+
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	var err error
+	if m.syncer != nil {
+		err = m.syncer.Close()
 	}
 
-	if errors.Is(err, os.ErrNotExist) {
-		return errNoView
+	if m.tree != nil {
+		err = errors.Join(err, m.tree.Close())
 	}
 
+	return errors.Join(err, m.log.Close())
+}
+
+// Append logs a new message, adds it to the tree and the views, and starts the
+// builds that are ready, also the failed ones.
+func (m *Memory) Append(kind, text string) (Message, error) {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	message, err := m.log.Append(kind, text)
 	if err != nil {
-		return fmt.Errorf("read view: %w", err)
+		return Message{}, err
 	}
 
-	var file viewFile
-	err = json.Unmarshal(data, &file)
+	defer m.syncer.Written()
+
+	err = m.addMessage(message)
 	if err != nil {
-		return fmt.Errorf("decode view: %w", err)
+		return Message{}, err
 	}
 
-	m.chat, err = restoreView(file.Chat, chatHigh, chatLow, m.tree)
+	err = m.save()
 	if err != nil {
-		return err
+		return Message{}, err
 	}
 
-	if m.chat.end > m.log.Len() {
-		return fmt.Errorf("view covers %d messages, but the log has %d", m.chat.end, m.log.Len())
+	m.queue.Retry()
+	m.dispatch()
+
+	return message, nil
+}
+
+// children reads the two children of the node key.
+func (m *Memory) children(key NodeKey) (Node, Node, error) {
+	var children [2]Node
+	for n := range children {
+		child, built, err := m.tree.Node(key.level-1, 2*key.index+n)
+		if err != nil {
+			return Node{}, Node{}, err
+		}
+
+		if !built {
+			return Node{}, Node{}, fmt.Errorf("node (%d, %d) is not built", key.level-1, 2*key.index+n)
+		}
+
+		children[n] = child
 	}
 
-	if m.chat.end == m.log.Len() {
-		return nil
+	return children[0], children[1], nil
+}
+
+// Wait returns when no build runs. The failed builds wait for the next message.
+func (m *Memory) Wait(ctx context.Context) error {
+	for {
+		m.mutex.Lock()
+		running := m.queue.Running()
+		changed := m.changed
+		m.mutex.Unlock()
+
+		if running == 0 {
+			return nil
+		}
+
+		select {
+		case <-changed:
+		case <-ctx.Done():
+			return ctx.Err()
+		}
 	}
+}
 
-	for m.chat.end < m.log.Len() {
-		err = m.addMessage(m.log.Message(m.chat.end))
-		if err != nil {
-			return err
+// Settle returns when the chat view has no pending line, so that every message
+// so far is summarized. A pending line with no build that runs belongs to a
+// failed build, which waits for the next message, so Settle returns
+// errMemoryUnsettled in place of a wait that never ends.
+func (m *Memory) Settle(ctx context.Context) error {
+	for {
+		m.mutex.Lock()
+		pending := m.chat.Pending()
+		running := m.queue.Running()
+		changed := m.changed
+		m.mutex.Unlock()
+
+		if !pending {
+			return nil
+		}
+
+		if running == 0 {
+			return errMemoryUnsettled
+		}
+
+		select {
+		case <-changed:
+		case <-ctx.Done():
+			return ctx.Err()
 		}
 	}
+}
 
-	return m.save()
+// Flush returns when every write before the call is durable.
+func (m *Memory) Flush(ctx context.Context) error {
+	return m.syncer.Flush(ctx)
 }
 
-// Close closes the files of the chat and releases its lock.
-func (m *Memory) Close() error {
+// ViewLines returns the lines of the chat view, each with its newline, without
+// the leaf lines of the messages in skip.
+func (m *Memory) ViewLines(skip map[int]bool) []string {
 	m.mutex.Lock()
 	defer m.mutex.Unlock()
 
-	var err error
-	if m.tree != nil {
-		err = m.tree.Close()
-	}
+	return m.chat.LinesWithout(skip)
+}
 
-	return errors.Join(err, m.log.Close())
+// Len returns the number of messages in the log.
+func (m *Memory) Len() int {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	return m.log.Len()
 }
 
-// Append logs a new message and adds it to the tree and the view.
-func (m *Memory) Append(kind, text string) (Message, error) {
+// Message reads the message with the given ID.
+func (m *Memory) Message(id int) (Message, error) {
 	m.mutex.Lock()
 	defer m.mutex.Unlock()
 
-	message, err := m.log.Append(kind, text)
+	return m.log.Message(id)
+}
+
+// Node reads the node (l, i) if it is built.
+func (m *Memory) Node(level, index int) (Node, bool, error) {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	return m.tree.Node(level, index)
+}
+
+// load reads the saved views.
+func (m *Memory) load() error {
+	data, err := os.ReadFile(m.viewPath())
+	if errors.Is(err, os.ErrNotExist) && m.log.Len() == 0 {
+		m.chat = newView(viewChatHigh, viewChatLow)
+		m.cview = newView(viewCompactionHigh, viewCompactionLow)
+
+		return nil
+	}
+
+	if errors.Is(err, os.ErrNotExist) {
+		return errMemoryNoView
+	}
+
 	if err != nil {
-		return Message{}, err
+		return fmt.Errorf("read view: %w", err)
 	}
 
-	err = m.addMessage(message)
+	var file MemoryViewFile
+	err = json.Unmarshal(data, &file)
 	if err != nil {
-		return Message{}, err
+		return fmt.Errorf("decode view: %w", err)
 	}
 
-	err = m.save()
+	m.chat, err = restoreView(file.Chat, viewChatHigh, viewChatLow, m.tree)
 	if err != nil {
-		return Message{}, err
+		return fmt.Errorf("chat view: %w", err)
 	}
 
-	return message, nil
+	m.cview, err = restoreView(file.Compaction, viewCompactionHigh, viewCompactionLow, m.tree)
+	if err != nil {
+		return fmt.Errorf("compaction view: %w", err)
+	}
+
+	if m.cview.end != m.chat.end {
+		return fmt.Errorf("compaction view covers %d messages, but the chat view %d", m.cview.end, m.chat.end)
+	}
+
+	if m.chat.end > m.log.Len() {
+		return fmt.Errorf("view covers %d messages, but the log has %d", m.chat.end, m.log.Len())
+	}
+
+	return nil
 }
 
-// addMessage builds the leaf of message when it needs no model call and appends
-// it to the view.
-func (m *Memory) addMessage(message Message) error {
-	_, built := m.tree.Node(0, message.ID)
-	leaf, free := freeLeaf(message)
-	if free && !built {
-		err := m.build(leaf)
+// restoreQueue adds the work of the messages in the view to the queue. It
+// builds each node that needs no model call and that a crash left out.
+func (m *Memory) restoreQueue() error {
+	for id := range m.chat.end {
+		built, err := m.tree.Built(0, id)
+		if err != nil {
+			return err
+		}
+
+		if built {
+			continue
+		}
+
+		message, err := m.log.Message(id)
+		if err != nil {
+			return err
+		}
+
+		leaf, free := freeLeaf(message)
+		if !free {
+			m.queue.AddLeaf(id)
+			continue
+		}
+
+		err = m.build(leaf)
 		if err != nil {
 			return err
 		}
 	}
 
-	m.chat.Append(m.tree)
+	keys, err := m.tree.Keys()
+	if err != nil {
+		return err
+	}
+
+	for _, key := range keys {
+		if key.index%2 != 0 {
+			continue
+		}
+
+		err = m.settle(key)
+		if err != nil {
+			return err
+		}
+	}
 
 	return nil
 }
 
-// build adds node to the tree, and then each parent above it that needs no
-// model call.
-func (m *Memory) build(node Node) error {
-	for {
-		err := m.tree.Add(node)
+// catchUp adds the messages that the view does not cover yet. A crash can stop
+// an append after the log write and before the view save.
+func (m *Memory) catchUp() error {
+	if m.chat.end == m.log.Len() {
+		return nil
+	}
+
+	for m.chat.end < m.log.Len() {
+		message, err := m.log.Message(m.chat.end)
 		if err != nil {
 			return err
 		}
 
-		sibling, built := m.tree.Node(node.Level, node.Index^1)
-		if !built {
-			return nil
+		err = m.addMessage(message)
+		if err != nil {
+			return err
 		}
+	}
 
-		left, right := node, sibling
-		if node.Index%2 != 0 {
-			left, right = sibling, node
+	return m.save()
+}
+
+// dispatch starts the builds that are ready, up to the limit. A build whose
+// input cannot be read fails and waits for the next message.
+func (m *Memory) dispatch() {
+	for m.ctx.Err() == nil {
+		key, ready := m.queue.Next()
+		if !ready {
+			return
 		}
 
-		parent, free := freeMerge(left, right)
-		if !free {
-			return nil
+		task, err := m.task(key)
+		if err != nil {
+			m.queue.Fail(key)
+			continue
 		}
 
-		node = parent
+		m.workers.Add(1)
+		go m.run(key, task)
+	}
+}
+
+// addMessage builds the leaf of message when it needs no model call, or else
+// adds the leaf to the queue. Then it appends the leaf to the views. A crash can
+// stop an append after the leaf build and before the view save, so a leaf that
+// is built already stays as it is, and restoreQueue settles its parent.
+func (m *Memory) addMessage(message Message) error {
+	built, err := m.tree.Built(0, message.ID)
+	if err != nil {
+		return err
 	}
+
+	leaf, free := freeLeaf(message)
+	switch {
+	case built:
+	case free:
+		err = m.build(leaf)
+		if err != nil {
+			return err
+		}
+	default:
+		m.queue.AddLeaf(message.ID)
+	}
+
+	merged, err := m.chat.Append(m.tree)
+	if err != nil {
+		return fmt.Errorf("append to chat view: %w", err)
+	}
+
+	_, err = m.cview.Append(m.tree)
+	if err == nil && merged {
+		err = m.cview.Coarsen(m.chat, m.tree)
+	}
+
+	if err != nil {
+		return fmt.Errorf("append to compaction view: %w", err)
+	}
+
+	return nil
 }
 
 // save writes the views to view.json.
 func (m *Memory) save() error {
-	data, err := json.Marshal(viewFile{
-		Chat: m.chat.State(),
+	data, err := json.Marshal(MemoryViewFile{
+		Chat:       m.chat.State(),
+		Compaction: m.cview.State(),
 	})
 	if err != nil {
 		return fmt.Errorf("encode view: %w", err)
@@ -206,55 +464,167 @@ func (m *Memory) viewPath() string {
 	return filepath.Join(m.root, "view.json")
 }
 
-// View returns the chat view as the model reads it.
-func (m *Memory) View() string {
-	m.mutex.Lock()
-	defer m.mutex.Unlock()
+// build adds node to the tree, keeps it in the views when it is a pending line,
+// and settles its parent.
+func (m *Memory) build(node Node) error {
+	err := m.tree.Add(node)
+	if err != nil {
+		return err
+	}
+
+	m.chat.Refresh(node)
+	m.cview.Refresh(node)
 
-	return m.chat.Render(m.tree)
+	return m.settle(NodeKey{node.Level, node.Index})
 }
 
-// writeFileAtomic replaces the file at path with data. A reader or a crash sees
-// the old file or the new file, never a part of one.
-func writeFileAtomic(path string, data []byte) error {
-	dir := filepath.Dir(path)
-	file, err := os.CreateTemp(dir, "."+filepath.Base(path)+".*")
+// settle looks at the parent of the built node key when both children are
+// built. It builds each parent that needs no model call, up the tree, and adds
+// the first other parent to the queue.
+func (m *Memory) settle(key NodeKey) error {
+	for {
+		parent := NodeKey{key.level + 1, key.index / 2}
+		ready, err := m.ready(parent)
+		if err != nil || !ready {
+			return err
+		}
+
+		left, right, err := m.children(parent)
+		if err != nil {
+			return err
+		}
+
+		node, free := freeMerge(left, right)
+		if !free {
+			m.queue.AddMerge(parent)
+			return nil
+		}
+
+		err = m.tree.Add(node)
+		if err != nil {
+			return err
+		}
+
+		key = parent
+	}
+}
+
+// task returns the build task of the node key, with its input read from the log
+// or the tree.
+func (m *Memory) task(key NodeKey) (Task, error) {
+	task := Task{
+		Level:   key.level,
+		Index:   key.index,
+		Context: m.context(key.level, key.index),
+	}
+
+	if key.level == 0 {
+		message, err := m.log.Message(key.index)
+		if err != nil {
+			return Task{}, err
+		}
+
+		task.Input = []string{leafSource(message)}
+
+		return task, nil
+	}
+
+	left, right, err := m.children(key)
 	if err != nil {
-		return fmt.Errorf("create temporary file: %w", err)
+		return Task{}, err
 	}
 
-	_, err = file.Write(data)
-	if err == nil {
-		err = file.Sync()
+	for _, child := range []Node{left, right} {
+		task.Input = append(task.Input, strings.ReplaceAll(child.Text, "\n", " "))
 	}
 
-	err = errors.Join(err, file.Close())
+	return task, nil
+}
+
+// run builds one node with the summarizer. A failed build stays in the queue
+// until the next message.
+func (m *Memory) run(key NodeKey, task Task) {
+	defer m.workers.Done()
+
+	text, err := m.summarizer.Summarize(m.ctx, task)
+
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
 	if err == nil {
-		err = os.Rename(file.Name(), path)
+		err = m.build(Node{
+			Level: key.level,
+			Index: key.index,
+			Text:  text,
+			Size:  len(text),
+		})
+		m.syncer.Written()
 	}
 
 	if err != nil {
-		return errors.Join(fmt.Errorf("replace file: %w", err), os.Remove(file.Name()))
+		m.queue.Fail(key)
+	} else {
+		m.queue.Done(key)
 	}
 
-	return syncDir(dir)
+	close(m.changed)
+	m.changed = make(chan struct{})
+	m.dispatch()
 }
 
-// syncDir flushes the entries of dir to the disk.
-func syncDir(dir string) (err error) {
-	handle, err := os.Open(dir)
-	if err != nil {
-		return fmt.Errorf("open directory: %w", err)
+// ready tells if both children of the node key are built and the node is not.
+func (m *Memory) ready(key NodeKey) (bool, error) {
+	for _, index := range []int{2 * key.index, 2*key.index + 1} {
+		built, err := m.tree.Built(key.level-1, index)
+		if err != nil || !built {
+			return false, err
+		}
 	}
 
-	defer func() {
-		err = errors.Join(err, handle.Close())
-	}()
+	built, err := m.tree.Built(key.level, key.index)
 
-	err = handle.Sync()
-	if err != nil {
-		return fmt.Errorf("flush directory: %w", err)
+	return !built, err
+}
+
+func (m *Memory) context(level, index int) string {
+	end := index
+	if level > 0 {
+		end = nodeEnd(level, index)
 	}
 
-	return nil
+	return m.cview.Context(end)
+}
+
+// View returns the chat view as the model reads it.
+func (m *Memory) View() string {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	return m.chat.Render()
+}
+
+// Context returns the compaction view for the build of node (l, i). It ends
+// before the message of a leaf, or after the last message of a merge, and it
+// stops at the first line that is not built.
+func (m *Memory) Context(level, index int) string {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	return m.context(level, index)
+}
+
+// Task is the build of the node (Level, Index). Context is the compaction view
+// for the build. Input holds the source "kind: text" of a leaf, or the two
+// child lines of a merge with their newlines as spaces.
+type Task struct {
+	Level   int
+	Index   int
+	Context string
+	Input   []string
+}
+
+// MemoryViewFile is the content of view.json.
+type MemoryViewFile struct {
+	Chat       ViewState `json:"chat"`
+	Compaction ViewState `json:"compaction"`
 }