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

   1 package main
   2
   3 import (
   4         "context"
   5         "encoding/json"
   6         "errors"
   7         "fmt"
   8         "os"
   9         "path/filepath"
  10         "time"
  11         "unicode/utf8"
  12 )
  13
  14 // logKinds holds the logKinds that a message can have.
  15 var logKinds = map[string]bool{
  16         "user":  true,
  17         "reply": true,
  18         "tool":  true,
  19         "echo":  true,
  20         "work":  true,
  21         "note":  true,
  22 }
  23
  24 // Log is the append-only message log of one chat. It writes one JSON line for
  25 // each message to main/YYYY-MM-DD.jsonl, named for the local day of the
  26 // message. It never edits or deletes a line. Each line goes to the operating
  27 // system before Append returns, so a crash of the process loses no message. An
  28 // exclusive lock on the lock file makes sure that only one process writes the
  29 // log. After a failed write, the end of the file is not known, so the log
  30 // refuses all writes until it is opened again. The caller serializes all calls.
  31 //
  32 // The log keeps no message in memory. The index index/main.idx holds the
  33 // position of message i in record i, so a read of a message is one read of the
  34 // index and one of its day file. At open the log indexes the lines after the
  35 // last indexed one, so a missing index is built again from the day files.
  36 type Log struct {
  37         clock   func() time.Time
  38         lock    *os.File
  39         journal Journal
  40         index   *Records
  41         failure error
  42 }
  43
  44 // Message is one entry of the chat log. ID is its position in the log, from 0.
  45 // A rebuild assigns positions in the merged log. Size is the UTF-8 byte length
  46 // of Text. Date is the message time in RFC 3339 format, supplied at append or
  47 // taken from the clock.
  48 type Message struct {
  49         ID   int    `json:"i"`
  50         Kind string `json:"kind"`
  51         Text string `json:"text"`
  52         Size int    `json:"size"`
  53         Date string `json:"date"`
  54 }
  55
  56 // openLog locks the chat in root and indexes its new lines. The clock gives the
  57 // date of each new message.
  58 func openLog(ctx context.Context, root string, clock func() time.Time) (*Log, error) {
  59         for _, dir := range []string{"main", "index"} {
  60                 err := os.MkdirAll(filepath.Join(root, dir), 0o700)
  61                 if err != nil {
  62                         return nil, fmt.Errorf("create %s directory: %w", dir, err)
  63                 }
  64         }
  65
  66         lock, err := lockFile(filepath.Join(root, "lock"))
  67         if err != nil {
  68                 return nil, err
  69         }
  70
  71         l := &Log{
  72                 clock: clock,
  73                 lock:  lock,
  74                 journal: Journal{
  75                         dir: filepath.Join(root, "main"),
  76                 },
  77         }
  78
  79         l.index, err = openRecords(filepath.Join(root, "index", "main.idx"), positionBytes)
  80         if err != nil {
  81                 return nil, errors.Join(err, l.Close())
  82         }
  83
  84         err = l.catchUp(ctx)
  85         if err != nil {
  86                 return nil, errors.Join(fmt.Errorf("index message log: %w", err), l.Close())
  87         }
  88
  89         return l, nil
  90 }
  91
  92 // Close releases the lock. The log refuses all calls after Close.
  93 func (l *Log) Close() error {
  94         err := l.journal.Close()
  95         if l.index != nil {
  96                 err = errors.Join(err, l.index.Close())
  97                 l.index = nil
  98         }
  99
 100         if l.lock != nil {
 101                 err = errors.Join(err, l.lock.Close())
 102                 l.lock = nil
 103         }
 104
 105         return err
 106 }
 107
 108 // catchUp indexes the lines after the last indexed line.
 109 func (l *Log) catchUp(ctx context.Context) error {
 110         start := Position{}
 111         if l.index.count > 0 {
 112                 record, _, err := l.index.Read(l.index.count - 1)
 113                 if err != nil {
 114                         return err
 115                 }
 116
 117                 last := decodePosition(record)
 118                 start = Position{
 119                         day:    last.day,
 120                         offset: last.offset + uint64(last.length),
 121                 }
 122         }
 123
 124         return l.journal.Scan(ctx, start, func(line []byte, place Position) error {
 125                 var message Message
 126
 127                 err := json.Unmarshal(line, &message)
 128                 if err != nil {
 129                         return fmt.Errorf("decode log message: %w", err)
 130                 }
 131
 132                 if message.ID != l.index.count {
 133                         return fmt.Errorf("expected message %d, found %d", l.index.count, message.ID)
 134                 }
 135
 136                 err = logValidateMessage(message)
 137                 if err != nil {
 138                         return fmt.Errorf("message %d: %w", message.ID, err)
 139                 }
 140
 141                 return l.addIndex(place)
 142         })
 143 }
 144
 145 func (l *Log) addIndex(place Position) error {
 146         record := make([]byte, positionBytes)
 147         encodePosition(record, place)
 148
 149         return l.index.Append(record)
 150 }
 151
 152 // Append writes a new message to the log and returns it with its ID, size, and
 153 // date.
 154 func (l *Log) Append(kind, text string) (Message, error) {
 155         return l.AppendDated(kind, text, "")
 156 }
 157
 158 // AppendDated writes a message with date in RFC 3339 format. An empty date uses
 159 // the clock. The day file uses the local day of the given date.
 160 func (l *Log) AppendDated(kind, text, date string) (Message, error) {
 161         if l.lock == nil {
 162                 return Message{}, fmt.Errorf("append dated message: %w", os.ErrClosed)
 163         }
 164
 165         if l.failure != nil {
 166                 return Message{}, fmt.Errorf("log needs a reopen after a failed write: %w", l.failure)
 167         }
 168
 169         var err error
 170
 171         now := l.clock()
 172         if date == "" {
 173                 date = now.Format(time.RFC3339Nano)
 174         } else {
 175                 now, err = time.Parse(time.RFC3339Nano, date)
 176                 if err != nil {
 177                         return Message{}, fmt.Errorf("parse message date: %w", err)
 178                 }
 179
 180                 now = now.In(time.Local)
 181         }
 182
 183         message := Message{
 184                 ID:   l.index.count,
 185                 Kind: kind,
 186                 Text: text,
 187                 Size: len(text),
 188                 Date: date,
 189         }
 190
 191         err = logValidateMessage(message)
 192         if err != nil {
 193                 return Message{}, fmt.Errorf("validate dated message: %w", err)
 194         }
 195
 196         line, err := json.Marshal(message)
 197         if err != nil {
 198                 return Message{}, fmt.Errorf("encode message: %w", err)
 199         }
 200
 201         place, err := l.journal.Write(now.Format(time.DateOnly), append(line, '\n'))
 202         if err == nil {
 203                 err = l.addIndex(place)
 204         }
 205
 206         if err != nil {
 207                 l.failure = err
 208                 return Message{}, fmt.Errorf("write message %d: %w", message.ID, err)
 209         }
 210
 211         return message, nil
 212 }
 213
 214 // Len returns the number of messages in the log.
 215 func (l *Log) Len() int {
 216         return l.index.count
 217 }
 218
 219 // Message reads the message with the given ID from its day file.
 220 func (l *Log) Message(id int) (Message, error) {
 221         record, found, err := l.index.Read(id)
 222         if err != nil {
 223                 return Message{}, err
 224         }
 225
 226         if !found {
 227                 return Message{}, fmt.Errorf("no message %d", id)
 228         }
 229
 230         line, err := l.journal.Read(decodePosition(record))
 231         if err != nil {
 232                 return Message{}, fmt.Errorf("read message %d: %w", id, err)
 233         }
 234
 235         var message Message
 236
 237         err = json.Unmarshal(line, &message)
 238         if err != nil {
 239                 return Message{}, fmt.Errorf("decode message %d: %w", id, err)
 240         }
 241
 242         if message.ID != id {
 243                 return Message{}, fmt.Errorf("index of message %d points to message %d", id, message.ID)
 244         }
 245
 246         return message, nil
 247 }
 248
 249 // logValidateMessage makes sure that a message has a known kind, valid UTF-8
 250 // text, the size of its text, and a valid date.
 251 func logValidateMessage(message Message) error {
 252         if !logKinds[message.Kind] {
 253                 return fmt.Errorf("unknown message kind %q", message.Kind)
 254         }
 255
 256         if !utf8.ValidString(message.Text) {
 257                 return errors.New("message text is not valid UTF-8")
 258         }
 259
 260         if message.Size != len(message.Text) {
 261                 return fmt.Errorf("message size %d is not the text size %d", message.Size, len(message.Text))
 262         }
 263
 264         _, err := time.Parse(time.RFC3339Nano, message.Date)
 265         if err != nil {
 266                 return fmt.Errorf("parse message date: %w", err)
 267         }
 268
 269         return nil
 270 }