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