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

   1 package main
   2
   3 import (
   4         "cmp"
   5         "context"
   6         "encoding/binary"
   7         "encoding/json"
   8         "errors"
   9         "fmt"
  10         "math/bits"
  11         "os"
  12         "path/filepath"
  13         "slices"
  14         "time"
  15         "unicode/utf8"
  16 )
  17
  18 // nodeBytes is the size limit of a node text in UTF-8 bytes. A model can go a
  19 // few bytes over it, because the view measures the real sizes.
  20 const nodeBytes = 512
  21
  22 // treeRecordBytes is the size of a record of index/tree.idx: the position of
  23 // the node line, the node index, and the node level.
  24 const treeRecordBytes = 32
  25
  26 // openTree opens the tree of the chat in root and indexes its new lines. A chat
  27 // of total messages cannot have a node that covers a later message.
  28 func openTree(ctx context.Context, root string, total int, clock func() time.Time) (*Tree, error) {
  29         for _, dir := range []string{"tree", "index"} {
  30                 err := os.MkdirAll(filepath.Join(root, dir), 0o700)
  31                 if err != nil {
  32                         return nil, fmt.Errorf("create %s directory: %w", dir, err)
  33                 }
  34         }
  35
  36         t := &Tree{
  37                 clock: clock,
  38                 journal: Journal{
  39                         dir: filepath.Join(root, "tree"),
  40                 },
  41                 dir:    filepath.Join(root, "index"),
  42                 levels: make(map[int]*Records),
  43         }
  44
  45         err := t.open(ctx, total)
  46         if err != nil {
  47                 return nil, errors.Join(fmt.Errorf("index tree: %w", err), t.Close())
  48         }
  49
  50         return t, nil
  51 }
  52
  53 // leafSource returns the text that the leaf of message holds or summarizes.
  54 func leafSource(message Message) string {
  55         return message.Kind + ": " + message.Text
  56 }
  57
  58 // freeLeaf returns a text-only leaf when its source fits in a node. Images need
  59 // a model description even when their text fits.
  60 func freeLeaf(message Message) (Node, bool) {
  61         text := leafSource(message)
  62         if len(text) > nodeBytes || len(imageFiles(message.Text)) > 0 {
  63                 return Node{}, false
  64         }
  65
  66         return Node{
  67                 Level: 0,
  68                 Index: message.ID,
  69                 Text:  text,
  70                 Size:  len(text),
  71         }, true
  72 }
  73
  74 // freeMerge returns the parent of two sibling nodes when both texts, joined by
  75 // a newline, fit in a node. Such a parent needs no model call.
  76 func freeMerge(left, right Node) (Node, bool) {
  77         if left.Size+1+right.Size > nodeBytes {
  78                 return Node{}, false
  79         }
  80
  81         text := left.Text + "\n" + right.Text
  82
  83         return Node{
  84                 Level: left.Level + 1,
  85                 Index: left.Index / 2,
  86                 Text:  text,
  87                 Size:  len(text),
  88         }, true
  89 }
  90
  91 // nodeEnd returns the ID after the last message of node (l, i).
  92 func nodeEnd(level, index int) int {
  93         return (index + 1) << level
  94 }
  95
  96 // Tree holds the built nodes of one chat. It writes one JSON line for each node
  97 // to tree/YYYY-MM-DD.jsonl, named for the local day of the build. Each node is
  98 // built once and never again. A node of a higher level is built only after both
  99 // of its children. After a failed write, the tree refuses all writes until it
 100 // is opened again. The caller holds the lock of the chat and serializes all
 101 // calls.
 102 //
 103 // The tree keeps no node in memory. The index index/tree.idx lists the node
 104 // lines in file order, and index/tree-L.idx holds the position of node (L, i)
 105 // in record i, with a zero record for a node that is not built. At open the
 106 // tree indexes the lines after the last indexed one, so a missing index is
 107 // built again from the day files.
 108 type Tree struct {
 109         clock   func() time.Time
 110         journal Journal
 111         dir     string
 112         order   *Records
 113         levels  map[int]*Records
 114         failure error
 115 }
 116
 117 // Close closes the files of the tree.
 118 func (t *Tree) Close() error {
 119         err := t.journal.Close()
 120         if t.order != nil {
 121                 err = errors.Join(err, t.order.Close())
 122                 t.order = nil
 123         }
 124
 125         for level, index := range t.levels {
 126                 err = errors.Join(err, index.Close())
 127
 128                 delete(t.levels, level)
 129         }
 130
 131         return err
 132 }
 133
 134 // Add writes a new node to the tree.
 135 func (t *Tree) Add(node Node) error {
 136         if t.failure != nil {
 137                 return fmt.Errorf("tree needs a reopen after a failed write: %w", t.failure)
 138         }
 139
 140         err := t.check(node)
 141         if err != nil {
 142                 return err
 143         }
 144
 145         line, err := json.Marshal(node)
 146         if err != nil {
 147                 return fmt.Errorf("encode node: %w", err)
 148         }
 149
 150         place, err := t.journal.Write(t.clock().Format(time.DateOnly), append(line, '\n'))
 151         if err == nil {
 152                 err = t.addIndex(NodeKey{node.Level, node.Index}, place)
 153         }
 154
 155         if err != nil {
 156                 t.failure = err
 157                 return fmt.Errorf("write node (%d, %d): %w", node.Level, node.Index, err)
 158         }
 159
 160         return nil
 161 }
 162
 163 // Built tells if the node (l, i) is built.
 164 func (t *Tree) Built(level, index int) (bool, error) {
 165         _, built, err := t.place(level, index)
 166         return built, err
 167 }
 168
 169 // Node reads the node (l, i) if it is built.
 170 func (t *Tree) Node(level, index int) (Node, bool, error) {
 171         place, built, err := t.place(level, index)
 172         if err != nil || !built {
 173                 return Node{}, false, err
 174         }
 175
 176         line, err := t.journal.Read(place)
 177         if err != nil {
 178                 return Node{}, false, fmt.Errorf("read node (%d, %d): %w", level, index, err)
 179         }
 180
 181         var node Node
 182
 183         err = json.Unmarshal(line, &node)
 184         if err != nil {
 185                 return Node{}, false, fmt.Errorf("decode node (%d, %d): %w", level, index, err)
 186         }
 187
 188         if node.Level != level || node.Index != index {
 189                 return Node{}, false, fmt.Errorf("index of node (%d, %d) points to node (%d, %d)", level, index, node.Level, node.Index)
 190         }
 191
 192         return node, true, nil
 193 }
 194
 195 // Keys returns the keys of the built nodes, ordered by level and then by index.
 196 func (t *Tree) Keys() ([]NodeKey, error) {
 197         keys := make([]NodeKey, 0, t.order.count)
 198         for n := range t.order.count {
 199                 record, _, err := t.order.Read(n)
 200                 if err != nil {
 201                         return nil, err
 202                 }
 203
 204                 key, _ := treeDecodeRecord(record)
 205                 keys = append(keys, key)
 206         }
 207
 208         slices.SortFunc(keys, func(a, b NodeKey) int {
 209                 return cmp.Or(cmp.Compare(a.level, b.level), cmp.Compare(a.index, b.index))
 210         })
 211
 212         return keys, nil
 213 }
 214
 215 func (t *Tree) open(ctx context.Context, total int) error {
 216         var err error
 217
 218         t.order, err = openRecords(filepath.Join(t.dir, "tree.idx"), treeRecordBytes)
 219         if err != nil {
 220                 return err
 221         }
 222
 223         start := Position{}
 224         if t.order.count == 0 {
 225                 err = t.removeLevels()
 226                 if err != nil {
 227                         return err
 228                 }
 229         } else {
 230                 record, _, err := t.order.Read(t.order.count - 1)
 231                 if err != nil {
 232                         return err
 233                 }
 234
 235                 // A crash can stop an add between the two index writes, so the
 236                 // last node gets its level record again.
 237                 node, last := treeDecodeRecord(record)
 238                 err = t.writeLevel(node, last)
 239                 if err != nil {
 240                         return err
 241                 }
 242
 243                 start = Position{
 244                         day:    last.day,
 245                         offset: last.offset + uint64(last.length),
 246                 }
 247         }
 248
 249         return t.journal.Scan(ctx, start, func(line []byte, place Position) error {
 250                 var node Node
 251
 252                 err := json.Unmarshal(line, &node)
 253                 if err != nil {
 254                         return fmt.Errorf("decode tree node: %w", err)
 255                 }
 256
 257                 err = t.check(node)
 258                 if err != nil {
 259                         return err
 260                 }
 261
 262                 if nodeEnd(node.Level, node.Index) > total {
 263                         return fmt.Errorf("node (%d, %d) covers messages after the log", node.Level, node.Index)
 264                 }
 265
 266                 return t.addIndex(NodeKey{node.Level, node.Index}, place)
 267         })
 268 }
 269
 270 // check makes sure that node is valid and can join the tree now.
 271 func (t *Tree) check(node Node) error {
 272         if node.Level < 0 || node.Level >= bits.UintSize-2 || node.Index < 0 {
 273                 return fmt.Errorf("node (%d, %d) is out of range", node.Level, node.Index)
 274         }
 275
 276         if node.Text == "" || !utf8.ValidString(node.Text) || node.Size != len(node.Text) {
 277                 return fmt.Errorf("node (%d, %d) has no valid text of its size", node.Level, node.Index)
 278         }
 279
 280         built, err := t.Built(node.Level, node.Index)
 281         if err != nil {
 282                 return err
 283         }
 284
 285         if built {
 286                 return fmt.Errorf("node (%d, %d) is already built", node.Level, node.Index)
 287         }
 288
 289         if node.Level == 0 {
 290                 return nil
 291         }
 292
 293         for _, index := range []int{2 * node.Index, 2*node.Index + 1} {
 294                 built, err = t.Built(node.Level-1, index)
 295                 if err != nil {
 296                         return err
 297                 }
 298
 299                 if !built {
 300                         return fmt.Errorf("node (%d, %d) has a child that is not built", node.Level, node.Index)
 301                 }
 302         }
 303
 304         return nil
 305 }
 306
 307 // addIndex adds the node line at place to both indexes.
 308 func (t *Tree) addIndex(node NodeKey, place Position) error {
 309         record := make([]byte, treeRecordBytes)
 310         encodePosition(record, place)
 311         binary.LittleEndian.PutUint64(record[positionBytes:], uint64(node.index))
 312         binary.LittleEndian.PutUint32(record[positionBytes+8:], uint32(node.level))
 313
 314         err := t.order.Append(record)
 315         if err != nil {
 316                 return err
 317         }
 318
 319         return t.writeLevel(node, place)
 320 }
 321
 322 // place returns the position of the node (l, i) if it is built.
 323 func (t *Tree) place(level, index int) (Position, bool, error) {
 324         if level < 0 || level >= bits.UintSize-2 || index < 0 {
 325                 return Position{}, false, nil
 326         }
 327
 328         records, err := t.level(level)
 329         if err != nil {
 330                 return Position{}, false, err
 331         }
 332
 333         record, found, err := records.Read(index)
 334         if err != nil || !found {
 335                 return Position{}, false, err
 336         }
 337
 338         place := decodePosition(record)
 339
 340         return place, place.length > 0, nil
 341 }
 342
 343 // removeLevels removes the level indexes, which are not valid without the order
 344 // index.
 345 func (t *Tree) removeLevels() error {
 346         paths, err := filepath.Glob(filepath.Join(t.dir, "tree-*.idx"))
 347         if err != nil {
 348                 return fmt.Errorf("list level indexes: %w", err)
 349         }
 350
 351         for _, path := range paths {
 352                 err = os.Remove(path)
 353                 if err != nil {
 354                         return fmt.Errorf("remove level index: %w", err)
 355                 }
 356         }
 357
 358         return nil
 359 }
 360
 361 // writeLevel writes the position of node to its level index.
 362 func (t *Tree) writeLevel(node NodeKey, place Position) error {
 363         level, err := t.level(node.level)
 364         if err != nil {
 365                 return err
 366         }
 367
 368         record := make([]byte, positionBytes)
 369         encodePosition(record, place)
 370
 371         return level.Write(node.index, record)
 372 }
 373
 374 // level returns the index of level, and opens it at first use.
 375 func (t *Tree) level(level int) (*Records, error) {
 376         index, open := t.levels[level]
 377         if open {
 378                 return index, nil
 379         }
 380
 381         index, err := openRecords(filepath.Join(t.dir, fmt.Sprintf("tree-%d.idx", level)), positionBytes)
 382         if err != nil {
 383                 return nil, err
 384         }
 385
 386         t.levels[level] = index
 387
 388         return index, nil
 389 }
 390
 391 // Count returns the number of built nodes.
 392 func (t *Tree) Count() int {
 393         return t.order.count
 394 }
 395
 396 // Node is one line of the summary tree. Node (l, i) covers the 2^l messages
 397 // from i*2^l on. A node of level 0 holds one message. A node of a higher level
 398 // merges its two children, (l-1, 2i) and (l-1, 2i+1). Size is the UTF-8 byte
 399 // length of Text.
 400 type Node struct {
 401         Level int    `json:"l"`
 402         Index int    `json:"i"`
 403         Text  string `json:"text"`
 404         Size  int    `json:"size"`
 405 }
 406
 407 // NodeKey names the node (l, i).
 408 type NodeKey struct {
 409         level int
 410         index int
 411 }
 412
 413 // treeDecodeRecord reads the key and the position of a record of
 414 // index/tree.idx.
 415 func treeDecodeRecord(record []byte) (NodeKey, Position) {
 416         key := NodeKey{
 417                 level: int(binary.LittleEndian.Uint32(record[positionBytes+8:])),
 418                 index: int(binary.LittleEndian.Uint64(record[positionBytes:])),
 419         }
 420
 421         return key, decodePosition(record)
 422 }