// 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 log and the tree once to fill the queue again. After that, only a
-// new message or a new node adds work.
+// 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
// viewFile is the content of view.json.
type viewFile struct {
- Chat viewState `json:"chat"`
- Compaction viewState `json:"compaction"`
+ Chat ViewState `json:"chat"`
+ Compaction ViewState `json:"compaction"`
}
// openMemory opens the chat in root and starts the builds that are ready. The
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)
if err != nil {
return nil, errors.Join(err, m.Close())
}
m.mutex.Lock()
+ m.syncer.Written()
m.dispatch()
m.mutex.Unlock()
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)
- m.cview = newView(compactionHigh, compactionLow)
+ m.chat = newView(viewChatHigh, viewChatLow)
+ m.cview = newView(viewCompactionHigh, viewCompactionLow)
return nil
}
return fmt.Errorf("decode view: %w", err)
}
- m.chat, err = restoreView(file.Chat, chatHigh, chatLow, m.tree)
+ m.chat, err = restoreView(file.Chat, viewChatHigh, viewChatLow, m.tree)
if err != nil {
return fmt.Errorf("chat view: %w", err)
}
- m.cview, err = restoreView(file.Compaction, compactionHigh, compactionLow, m.tree)
+ m.cview, err = restoreView(file.Compaction, viewCompactionHigh, viewCompactionLow, m.tree)
if err != nil {
return fmt.Errorf("compaction view: %w", err)
}
// 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 {
- if m.tree.Built(0, id) {
+ built, err := m.tree.Built(0, id)
+ if err != nil {
+ return err
+ }
+
+ if built {
continue
}
- leaf, free := freeLeaf(m.log.Message(id))
+ 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)
+ err = m.build(leaf)
if err != nil {
return err
}
}
- for _, node := range m.tree.Nodes() {
- if node.Index%2 != 0 {
+ keys, err := m.tree.Keys()
+ if err != nil {
+ return err
+ }
+
+ for _, key := range keys {
+ if key.index%2 != 0 {
continue
}
- err := m.settle(node)
+ err = m.settle(key)
if err != nil {
return err
}
}
for m.chat.end < m.log.Len() {
- err := m.addMessage(m.log.Message(m.chat.end))
+ message, err := m.log.Message(m.chat.end)
+ if err != nil {
+ return err
+ }
+
+ err = m.addMessage(message)
if err != nil {
return err
}
return m.save()
}
-// Close stops the builds, waits for them, closes the files of the chat, and
-// releases its lock.
+// 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()
defer m.mutex.Unlock()
var err error
+ if m.syncer != nil {
+ err = m.syncer.Close()
+ }
+
if m.tree != nil {
- err = m.tree.Close()
+ err = errors.Join(err, m.tree.Close())
}
return errors.Join(err, m.log.Close())
return Message{}, err
}
+ defer m.syncer.Written()
+
err = m.addMessage(message)
if err != nil {
return Message{}, err
}
// 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.
+// 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)
- if free {
- err := m.build(leaf)
+ switch {
+ case built:
+ case free:
+ err = m.build(leaf)
if err != nil {
return err
}
- } else {
+ default:
m.queue.AddLeaf(message.ID)
}
- merged := m.chat.Append(m.tree)
- m.cview.Append(m.tree)
- if merged {
- m.cview.Coarsen(m.chat, m.tree)
+ 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
}
-// build adds node to the tree and settles its parent.
+// 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
}
- return m.settle(node)
+ m.chat.Refresh(node)
+ m.cview.Refresh(node)
+
+ return m.settle(nodeKey{node.Level, node.Index})
}
-// settle looks at the parent of node 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(node Node) error {
+// 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 {
- sibling, built := m.tree.Node(node.Level, node.Index^1)
- if !built || m.tree.Built(node.Level+1, node.Index/2) {
- return nil
+ parent := nodeKey{key.level + 1, key.index / 2}
+ ready, err := m.ready(parent)
+ if err != nil || !ready {
+ return err
}
- left, right := node, sibling
- if node.Index%2 != 0 {
- left, right = sibling, node
+ left, right, err := m.children(parent)
+ if err != nil {
+ return err
}
- parent, free := freeMerge(left, right)
+ node, free := freeMerge(left, right)
if !free {
- m.queue.AddMerge(nodeKey{node.Level + 1, node.Index / 2})
+ m.queue.AddMerge(parent)
return nil
}
- err := m.tree.Add(parent)
+ err = m.tree.Add(node)
if err != nil {
return err
}
- node = parent
+ key = parent
+ }
+}
+
+// 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
+ }
+ }
+
+ built, err := m.tree.Built(key.level, key.index)
+
+ return !built, err
+}
+
+// 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
}
+
+ return children[0], children[1], nil
}
-// dispatch starts the builds that are ready, up to the limit.
+// 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()
return
}
- task := m.task(key)
+ task, err := m.task(key)
+ if err != nil {
+ m.queue.Fail(key)
+ continue
+ }
+
m.workers.Add(1)
go m.run(key, task)
}
}
-// task returns the build task of the node key.
-func (m *Memory) task(key nodeKey) Task {
+// 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,
}
if key.level == 0 {
- task.Input = []string{leafSource(m.log.Message(key.index))}
- return task
+ message, err := m.log.Message(key.index)
+ if err != nil {
+ return Task{}, err
+ }
+
+ task.Input = []string{leafSource(message)}
+
+ return task, nil
}
- for _, index := range []int{2 * key.index, 2*key.index + 1} {
- child, _ := m.tree.Node(key.level-1, index)
+ left, right, err := m.children(key)
+ if err != nil {
+ return Task{}, err
+ }
+
+ for _, child := range []Node{left, right} {
task.Input = append(task.Input, strings.ReplaceAll(child.Text, "\n", " "))
}
- return task
+ return task, nil
}
// run builds one node with the summarizer. A failed build stays in the queue
Text: text,
Size: len(text),
})
+ m.syncer.Written()
}
if err != nil {
}
}
+// Flush returns when every write before the call is durable.
+func (m *Memory) Flush(ctx context.Context) error {
+ return m.syncer.Flush(ctx)
+}
+
// save writes the views to view.json.
func (m *Memory) save() error {
data, err := json.Marshal(viewFile{
m.mutex.Lock()
defer m.mutex.Unlock()
- return m.chat.Render(m.tree)
+ return m.chat.Render()
+}
+
+// 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()
+
+ return m.chat.LinesWithout(skip)
+}
+
+// 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()
+}
+
+// Message reads the message with the given ID.
+func (m *Memory) Message(id int) (Message, error) {
+ m.mutex.Lock()
+ defer m.mutex.Unlock()
+
+ 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)
}
// Context returns the compaction view for the build of node (l, i). It ends
end = nodeEnd(level, index)
}
- return m.cview.Context(m.tree, end)
-}
-
-// writeFileAtomic replaces the file at path with data through a temporary file
-// and a rename. A reader or a crash of the process sees the old file or the new
-// file, never a part of one.
-func writeFileAtomic(path string, data []byte) error {
- file, err := os.CreateTemp(filepath.Dir(path), "."+filepath.Base(path)+".*")
- if err != nil {
- return fmt.Errorf("create temporary file: %w", err)
- }
-
- _, err = file.Write(data)
- err = errors.Join(err, file.Close())
- if err == nil {
- err = os.Rename(file.Name(), path)
- }
-
- if err != nil {
- return errors.Join(fmt.Errorf("replace file: %w", err), os.Remove(file.Name()))
- }
-
- return nil
+ return m.cview.Context(end)
}
projects / onechat / blobdiff