git.giacomolli.uk - onechat/blob - memory.go

   1 package main
   2
   3 import (
   4         "context"
   5         "encoding/json"
   6         "errors"
   7         "fmt"
   8         "log"
   9         "os"
  10         "path/filepath"
  11         "strings"
  12         "sync"
  13         "time"
  14 )
  15
  16 var (
  17         // errMemoryNoView tells that a chat has messages but no saved view.
  18         errMemoryNoView = errors.New("the chat has messages but no saved view")
  19
  20         // errMemoryUnsettled tells that a line of the view stays pending,
  21         // because its build failed.
  22         errMemoryUnsettled = errors.New("a summary of the view failed, and it is tried again at the next message")
  23 )
  24
  25 // openMemory opens the chat in root and starts the builds that are ready. The
  26 // clock gives the dates of new messages and the days of new nodes. logger gets
  27 // one line for each failed build. The builds stop when ctx ends or at Close.
  28 func openMemory(ctx context.Context, root string, clock func() time.Time, summarizer Summarizer, logger *log.Logger) (*Memory, error) {
  29         log, err := openLog(ctx, root, clock)
  30         if err != nil {
  31                 return nil, err
  32         }
  33
  34         m := &Memory{
  35                 root:       root,
  36                 log:        log,
  37                 queue:      newBuildQueue(),
  38                 summarizer: summarizer,
  39                 logger:     logger,
  40                 changed:    make(chan struct{}),
  41         }
  42
  43         m.ctx, m.cancel = context.WithCancel(ctx)
  44
  45         m.syncer, err = openSyncer(root)
  46         if err != nil {
  47                 return nil, errors.Join(err, m.Close())
  48         }
  49
  50         m.tree, err = openTree(ctx, root, log.Len(), clock)
  51         if err != nil {
  52                 return nil, errors.Join(err, m.Close())
  53         }
  54
  55         err = m.load()
  56         if err != nil {
  57                 return nil, errors.Join(fmt.Errorf("load view: %w", err), m.Close())
  58         }
  59
  60         err = m.restoreQueue()
  61         if err != nil {
  62                 return nil, errors.Join(fmt.Errorf("restore build queue: %w", err), m.Close())
  63         }
  64
  65         err = m.catchUp()
  66         if err != nil {
  67                 return nil, errors.Join(err, m.Close())
  68         }
  69
  70         m.mutex.Lock()
  71         m.syncer.Written()
  72         m.dispatch()
  73         m.mutex.Unlock()
  74
  75         return m, nil
  76 }
  77
  78 // Summarizer builds the text of one node with a model.
  79 type Summarizer interface {
  80         Summarize(ctx context.Context, task Task) (string, error)
  81 }
  82
  83 // Memory is the memory of one chat: the message log, the summary tree, the chat
  84 // view, the compaction view, and the queue of the builds. The compaction view
  85 // is the chat view merged further, at the limits of 16 to 32 KB. It takes the
  86 // same new lines, and it merges again in one batch when it is too large or when
  87 // the chat view merges. So the compactions share its prefix in the prompt
  88 // cache. It saves the views to view.json after each change and loads them at
  89 // start. It never builds the views again from the log, because a new view
  90 // differs from the live one and so loses the prompt cache.
  91 //
  92 // Each write goes to the operating system at once, and a syncer makes the
  93 // writes durable by group commit in the background. Flush waits until the
  94 // writes so far are durable.
  95 //
  96 // The summarizer builds the nodes in the queue in the background. The start
  97 // reads the indexes once to fill the queue again. After that, only a new
  98 // message or a new node adds work. Memory keeps the nodes of the views and the
  99 // tasks of the running builds. It reads all other messages and nodes from their
 100 // day files through the indexes.
 101 type Memory struct {
 102         mutex      sync.Mutex
 103         root       string
 104         log        *Log
 105         syncer     *Syncer
 106         tree       *Tree
 107         chat       *View
 108         cview      *View
 109         queue      *BuildQueue
 110         summarizer Summarizer
 111         logger     *log.Logger
 112         ctx        context.Context
 113         cancel     context.CancelFunc
 114         workers    sync.WaitGroup
 115         changed    chan struct{}
 116 }
 117
 118 // Close stops the builds, waits for them, makes the writes durable, closes the
 119 // files of the chat, and releases its lock.
 120 func (m *Memory) Close() error {
 121         m.cancel()
 122         m.workers.Wait()
 123
 124         m.mutex.Lock()
 125         defer m.mutex.Unlock()
 126
 127         var err error
 128         if m.syncer != nil {
 129                 err = m.syncer.Close()
 130         }
 131
 132         if m.tree != nil {
 133                 err = errors.Join(err, m.tree.Close())
 134         }
 135
 136         return errors.Join(err, m.log.Close())
 137 }
 138
 139 // Append logs a new message, adds it to the tree and the views, and starts the
 140 // builds that are ready, also the failed ones.
 141 func (m *Memory) Append(kind, text string) (Message, error) {
 142         return m.AppendDated(kind, text, "")
 143 }
 144
 145 // AppendDated appends a message with its RFC 3339 date and builds its nodes. An
 146 // empty date uses the clock, as Append does.
 147 func (m *Memory) AppendDated(kind, text, date string) (Message, error) {
 148         m.mutex.Lock()
 149         defer m.mutex.Unlock()
 150
 151         message, err := m.log.AppendDated(kind, text, date)
 152         if err != nil {
 153                 return Message{}, fmt.Errorf("append dated log message: %w", err)
 154         }
 155
 156         defer m.syncer.Written()
 157
 158         err = m.addMessage(message)
 159         if err != nil {
 160                 return Message{}, fmt.Errorf("add dated message to memory: %w", err)
 161         }
 162
 163         err = m.save()
 164         if err != nil {
 165                 return Message{}, fmt.Errorf("save dated message view: %w", err)
 166         }
 167
 168         m.queue.Retry()
 169         m.dispatch()
 170
 171         return message, nil
 172 }
 173
 174 // children reads the two children of the node key.
 175 func (m *Memory) children(key NodeKey) (Node, Node, error) {
 176         var children [2]Node
 177         for n := range children {
 178                 child, built, err := m.tree.Node(key.level-1, 2*key.index+n)
 179                 if err != nil {
 180                         return Node{}, Node{}, err
 181                 }
 182
 183                 if !built {
 184                         return Node{}, Node{}, fmt.Errorf("node (%d, %d) is not built", key.level-1, 2*key.index+n)
 185                 }
 186
 187                 children[n] = child
 188         }
 189
 190         return children[0], children[1], nil
 191 }
 192
 193 // Wait returns when no build runs. The failed builds wait for the next message.
 194 func (m *Memory) Wait(ctx context.Context) error {
 195         for {
 196                 m.mutex.Lock()
 197                 running := m.queue.Running()
 198                 changed := m.changed
 199                 m.mutex.Unlock()
 200
 201                 if running == 0 {
 202                         return nil
 203                 }
 204
 205                 select {
 206                 case <-changed:
 207                 case <-ctx.Done():
 208                         return ctx.Err()
 209                 }
 210         }
 211 }
 212
 213 // Settle returns when the chat view has no pending line, so that every message
 214 // so far is summarized. A pending line with no build that runs belongs to a
 215 // failed build, which waits for the next message, so Settle returns
 216 // errMemoryUnsettled in place of a wait that never ends.
 217 func (m *Memory) Settle(ctx context.Context) error {
 218         for {
 219                 m.mutex.Lock()
 220                 pending := m.chat.Pending()
 221                 running := m.queue.Running()
 222                 changed := m.changed
 223                 m.mutex.Unlock()
 224
 225                 if !pending {
 226                         return nil
 227                 }
 228
 229                 if running == 0 {
 230                         return errMemoryUnsettled
 231                 }
 232
 233                 select {
 234                 case <-changed:
 235                 case <-ctx.Done():
 236                         return ctx.Err()
 237                 }
 238         }
 239 }
 240
 241 // Flush returns when every write before the call is durable.
 242 func (m *Memory) Flush(ctx context.Context) error {
 243         return m.syncer.Flush(ctx)
 244 }
 245
 246 // ViewLines returns the lines of the chat view, each with its newline, without
 247 // the leaf lines of the messages in skip.
 248 func (m *Memory) ViewLines(skip map[int]bool) []string {
 249         m.mutex.Lock()
 250         defer m.mutex.Unlock()
 251
 252         return m.chat.LinesWithout(skip)
 253 }
 254
 255 // Len returns the number of messages in the log.
 256 func (m *Memory) Len() int {
 257         m.mutex.Lock()
 258         defer m.mutex.Unlock()
 259
 260         return m.log.Len()
 261 }
 262
 263 // Message reads the message with the given ID.
 264 func (m *Memory) Message(id int) (Message, error) {
 265         m.mutex.Lock()
 266         defer m.mutex.Unlock()
 267
 268         return m.log.Message(id)
 269 }
 270
 271 // Node reads the node (l, i) if it is built.
 272 func (m *Memory) Node(level, index int) (Node, bool, error) {
 273         m.mutex.Lock()
 274         defer m.mutex.Unlock()
 275
 276         return m.tree.Node(level, index)
 277 }
 278
 279 // load reads the saved views.
 280 func (m *Memory) load() error {
 281         data, err := os.ReadFile(m.viewPath())
 282         if errors.Is(err, os.ErrNotExist) && m.log.Len() == 0 {
 283                 m.chat = newView(viewChatHigh, viewChatLow)
 284                 m.cview = newView(viewCompactionHigh, viewCompactionLow)
 285
 286                 return nil
 287         }
 288
 289         if errors.Is(err, os.ErrNotExist) {
 290                 return errMemoryNoView
 291         }
 292
 293         if err != nil {
 294                 return fmt.Errorf("read view: %w", err)
 295         }
 296
 297         var file MemoryViewFile
 298         err = json.Unmarshal(data, &file)
 299         if err != nil {
 300                 return fmt.Errorf("decode view: %w", err)
 301         }
 302
 303         m.chat, err = restoreView(file.Chat, viewChatHigh, viewChatLow, m.tree)
 304         if err != nil {
 305                 return fmt.Errorf("chat view: %w", err)
 306         }
 307
 308         m.cview, err = restoreView(file.Compaction, viewCompactionHigh, viewCompactionLow, m.tree)
 309         if err != nil {
 310                 return fmt.Errorf("compaction view: %w", err)
 311         }
 312
 313         if m.cview.end != m.chat.end {
 314                 return fmt.Errorf("compaction view covers %d messages, but the chat view %d", m.cview.end, m.chat.end)
 315         }
 316
 317         if m.chat.end > m.log.Len() {
 318                 return fmt.Errorf("view covers %d messages, but the log has %d", m.chat.end, m.log.Len())
 319         }
 320
 321         return nil
 322 }
 323
 324 // restoreQueue adds the work of the messages in the view to the queue. It
 325 // builds each node that needs no model call and that a crash left out.
 326 func (m *Memory) restoreQueue() error {
 327         for id := range m.chat.end {
 328                 built, err := m.tree.Built(0, id)
 329                 if err != nil {
 330                         return err
 331                 }
 332
 333                 if built {
 334                         continue
 335                 }
 336
 337                 message, err := m.log.Message(id)
 338                 if err != nil {
 339                         return err
 340                 }
 341
 342                 leaf, free := freeLeaf(message)
 343                 if !free {
 344                         m.queue.AddLeaf(id)
 345                         continue
 346                 }
 347
 348                 err = m.build(leaf)
 349                 if err != nil {
 350                         return err
 351                 }
 352         }
 353
 354         keys, err := m.tree.Keys()
 355         if err != nil {
 356                 return err
 357         }
 358
 359         for _, key := range keys {
 360                 if key.index%2 != 0 {
 361                         continue
 362                 }
 363
 364                 err = m.settle(key)
 365                 if err != nil {
 366                         return err
 367                 }
 368         }
 369
 370         return nil
 371 }
 372
 373 // catchUp adds the messages that the view does not cover yet. A crash can stop
 374 // an append after the log write and before the view save.
 375 func (m *Memory) catchUp() error {
 376         if m.chat.end == m.log.Len() {
 377                 return nil
 378         }
 379
 380         for m.chat.end < m.log.Len() {
 381                 message, err := m.log.Message(m.chat.end)
 382                 if err != nil {
 383                         return err
 384                 }
 385
 386                 err = m.addMessage(message)
 387                 if err != nil {
 388                         return err
 389                 }
 390         }
 391
 392         return m.save()
 393 }
 394
 395 // dispatch starts the builds that are ready, up to the limit. A build whose
 396 // input cannot be read fails and waits for the next message.
 397 func (m *Memory) dispatch() {
 398         for m.ctx.Err() == nil {
 399                 key, ready := m.queue.Next()
 400                 if !ready {
 401                         return
 402                 }
 403
 404                 task, err := m.task(key)
 405                 if err != nil {
 406                         m.failed(key, fmt.Errorf("read its input: %w", err))
 407                         m.queue.Fail(key)
 408
 409                         continue
 410                 }
 411
 412                 m.workers.Add(1)
 413                 go m.run(key, task)
 414         }
 415 }
 416
 417 // addMessage builds the leaf of message when it needs no model call, or else
 418 // adds the leaf to the queue. Then it appends the leaf to the views. A crash can
 419 // stop an append after the leaf build and before the view save, so a leaf that
 420 // is built already stays as it is, and restoreQueue settles its parent.
 421 func (m *Memory) addMessage(message Message) error {
 422         built, err := m.tree.Built(0, message.ID)
 423         if err != nil {
 424                 return err
 425         }
 426
 427         leaf, free := freeLeaf(message)
 428         switch {
 429         case built:
 430         case free:
 431                 err = m.build(leaf)
 432                 if err != nil {
 433                         return err
 434                 }
 435         default:
 436                 m.queue.AddLeaf(message.ID)
 437         }
 438
 439         merged, err := m.chat.Append(m.tree)
 440         if err != nil {
 441                 return fmt.Errorf("append to chat view: %w", err)
 442         }
 443
 444         _, err = m.cview.Append(m.tree)
 445         if err == nil && merged {
 446                 err = m.cview.Coarsen(m.chat, m.tree)
 447         }
 448
 449         if err != nil {
 450                 return fmt.Errorf("append to compaction view: %w", err)
 451         }
 452
 453         return nil
 454 }
 455
 456 // save writes the views to view.json.
 457 func (m *Memory) save() error {
 458         data, err := json.Marshal(MemoryViewFile{
 459                 Chat:       m.chat.State(),
 460                 Compaction: m.cview.State(),
 461         })
 462         if err != nil {
 463                 return fmt.Errorf("encode view: %w", err)
 464         }
 465
 466         err = writeFileAtomic(m.viewPath(), append(data, '\n'))
 467         if err != nil {
 468                 return fmt.Errorf("save view: %w", err)
 469         }
 470
 471         return nil
 472 }
 473
 474 func (m *Memory) viewPath() string {
 475         return filepath.Join(m.root, "view.json")
 476 }
 477
 478 // build adds node to the tree, keeps it in the views when it is a pending line,
 479 // and settles its parent.
 480 func (m *Memory) build(node Node) error {
 481         err := m.tree.Add(node)
 482         if err != nil {
 483                 return err
 484         }
 485
 486         m.chat.Refresh(node)
 487         m.cview.Refresh(node)
 488
 489         return m.settle(NodeKey{node.Level, node.Index})
 490 }
 491
 492 // settle looks at the parent of the built node key when both children are
 493 // built. It builds each parent that needs no model call, up the tree, and adds
 494 // the first other parent to the queue.
 495 func (m *Memory) settle(key NodeKey) error {
 496         for {
 497                 parent := NodeKey{key.level + 1, key.index / 2}
 498                 ready, err := m.ready(parent)
 499                 if err != nil || !ready {
 500                         return err
 501                 }
 502
 503                 left, right, err := m.children(parent)
 504                 if err != nil {
 505                         return err
 506                 }
 507
 508                 node, free := freeMerge(left, right)
 509                 if !free {
 510                         m.queue.AddMerge(parent)
 511                         return nil
 512                 }
 513
 514                 err = m.tree.Add(node)
 515                 if err != nil {
 516                         return err
 517                 }
 518
 519                 key = parent
 520         }
 521 }
 522
 523 // task returns the build task of the node key, with its input read from the log
 524 // or the tree.
 525 func (m *Memory) task(key NodeKey) (Task, error) {
 526         task := Task{
 527                 Level:   key.level,
 528                 Index:   key.index,
 529                 Context: m.context(key.level, key.index),
 530         }
 531
 532         if key.level == 0 {
 533                 message, err := m.log.Message(key.index)
 534                 if err != nil {
 535                         return Task{}, err
 536                 }
 537
 538                 task.Input = []string{leafSource(message)}
 539
 540                 return task, nil
 541         }
 542
 543         left, right, err := m.children(key)
 544         if err != nil {
 545                 return Task{}, err
 546         }
 547
 548         for _, child := range []Node{left, right} {
 549                 task.Input = append(task.Input, strings.ReplaceAll(child.Text, "\n", " "))
 550         }
 551
 552         return task, nil
 553 }
 554
 555 // run builds one node with the summarizer. A failed build stays in the queue
 556 // until the next message.
 557 func (m *Memory) run(key NodeKey, task Task) {
 558         defer m.workers.Done()
 559
 560         text, err := m.summarizer.Summarize(m.ctx, task)
 561
 562         m.mutex.Lock()
 563         defer m.mutex.Unlock()
 564
 565         if err == nil {
 566                 err = m.build(Node{
 567                         Level: key.level,
 568                         Index: key.index,
 569                         Text:  text,
 570                         Size:  len(text),
 571                 })
 572                 m.syncer.Written()
 573         }
 574
 575         if err != nil && m.ctx.Err() == nil {
 576                 m.failed(key, err)
 577         }
 578
 579         if err != nil {
 580                 m.queue.Fail(key)
 581         } else {
 582                 m.queue.Done(key)
 583         }
 584
 585         close(m.changed)
 586         m.changed = make(chan struct{})
 587         m.dispatch()
 588 }
 589
 590 // failed logs the failed build of the node key with its cause err.
 591 func (m *Memory) failed(key NodeKey, err error) {
 592         m.logger.Printf("summary failed, retry at next message: node (%d, %d): %v", key.level, key.index, err)
 593 }
 594
 595 // ready tells if both children of the node key are built and the node is not.
 596 func (m *Memory) ready(key NodeKey) (bool, error) {
 597         for _, index := range []int{2 * key.index, 2*key.index + 1} {
 598                 built, err := m.tree.Built(key.level-1, index)
 599                 if err != nil || !built {
 600                         return false, err
 601                 }
 602         }
 603
 604         built, err := m.tree.Built(key.level, key.index)
 605
 606         return !built, err
 607 }
 608
 609 func (m *Memory) context(level, index int) string {
 610         end := index
 611         if level > 0 {
 612                 end = nodeEnd(level, index)
 613         }
 614
 615         return m.cview.Context(end)
 616 }
 617
 618 // View returns the chat view as the model reads it.
 619 func (m *Memory) View() string {
 620         m.mutex.Lock()
 621         defer m.mutex.Unlock()
 622
 623         return m.chat.Render()
 624 }
 625
 626 // Context returns the compaction view for the build of node (l, i). It ends
 627 // before the message of a leaf, or after the last message of a merge, and it
 628 // stops at the first line that is not built.
 629 func (m *Memory) Context(level, index int) string {
 630         m.mutex.Lock()
 631         defer m.mutex.Unlock()
 632
 633         return m.context(level, index)
 634 }
 635
 636 // rebuildComplete checks the whole tree, not only the visible view. A compactor
 637 // can fail without an HTTP error, for example with an empty line. Such a failed
 638 // merge can be absent from the view and must prevent a swap.
 639 func (m *Memory) rebuildComplete() error {
 640         m.mutex.Lock()
 641         defer m.mutex.Unlock()
 642
 643         if len(m.queue.leaves) != 0 || len(m.queue.queued) != 0 {
 644                 return errors.New("rebuild failed: summary tree has unfinished nodes")
 645         }
 646
 647         return nil
 648 }
 649
 650 // buildModelLines returns the sizes of the nodes that the model wrote, in tree
 651 // order. A node that fits its source with no model call is not one of them.
 652 func (m *Memory) buildModelLines() ([]int, error) {
 653         m.mutex.Lock()
 654         defer m.mutex.Unlock()
 655
 656         keys, err := m.tree.Keys()
 657         if err != nil {
 658                 return nil, fmt.Errorf("list built nodes: %w", err)
 659         }
 660
 661         var sizes []int
 662         for _, key := range keys {
 663                 var node Node
 664
 665                 node, _, err = m.tree.Node(key.level, key.index)
 666                 if err != nil {
 667                         return nil, fmt.Errorf("read built node: %w", err)
 668                 }
 669
 670                 var model bool
 671
 672                 model, err = m.buildBuiltByModel(key)
 673                 if err != nil {
 674                         return nil, fmt.Errorf("check model source: %w", err)
 675                 }
 676
 677                 if model {
 678                         sizes = append(sizes, node.Size)
 679                 }
 680         }
 681
 682         return sizes, nil
 683 }
 684
 685 // buildBuiltByModel tells if the node key needed a model call: a leaf whose
 686 // source is over nodeBytes, or a merge whose children do not fit in one node.
 687 func (m *Memory) buildBuiltByModel(key NodeKey) (bool, error) {
 688         if key.level == 0 {
 689                 message, err := m.log.Message(key.index)
 690                 if err != nil {
 691                         return false, fmt.Errorf("read leaf source: %w", err)
 692                 }
 693
 694                 _, free := freeLeaf(message)
 695
 696                 return !free, nil
 697         }
 698
 699         left, right, err := m.children(key)
 700         if err != nil {
 701                 return false, fmt.Errorf("read merge sources: %w", err)
 702         }
 703
 704         _, free := freeMerge(left, right)
 705
 706         return !free, nil
 707 }
 708
 709 // Task is the build of the node (Level, Index). Context is the compaction view
 710 // for the build. Input holds the source "kind: text" of a leaf, or the two
 711 // child lines of a merge with their newlines as spaces.
 712 type Task struct {
 713         Level   int
 714         Index   int
 715         Context string
 716         Input   []string
 717 }
 718
 719 // MemoryViewFile is the content of view.json.
 720 type MemoryViewFile struct {
 721         Chat       ViewState `json:"chat"`
 722         Compaction ViewState `json:"compaction"`
 723 }