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

   1 package main
   2
   3 import (
   4         "bufio"
   5         "cmp"
   6         "context"
   7         "encoding/json"
   8         "errors"
   9         "fmt"
  10         "io"
  11         "maps"
  12         "os"
  13         "slices"
  14         "strconv"
  15         "strings"
  16         "time"
  17 )
  18
  19 // buildLineBytes is the most bytes of one line in a build input or usage log.
  20 const buildLineBytes = 1 << 20
  21
  22 // buildQuotaLimit is the largest increase in used quota points of a run.
  23 const buildQuotaLimit = 15.0
  24
  25 // buildCallLimit is the most model calls of a run.
  26 const buildCallLimit = 660
  27
  28 // BuildRun appends messages, builds their tree, and stops the run by the rules
  29 // of buildStop.
  30 type BuildRun struct {
  31         memory    *Memory
  32         client    *ModelClient
  33         usagePath string
  34 }
  35
  36 // Run appends the entries. It waits for the builds only before a user message,
  37 // as a turn does, so up to buildMax builds run at once in between, and it waits
  38 // for all builds at the end. It returns the number of messages appended and why
  39 // the run stopped early, or "" when it did not. After a stop it waits for the
  40 // calls in flight, so no sent call is cancelled.
  41 func (r *BuildRun) Run(ctx context.Context, entries []DatedMessage) (int, string) {
  42         appended, reason := r.append(ctx, entries)
  43         if reason == "" {
  44                 return appended, ""
  45         }
  46
  47         err := r.memory.Wait(ctx)
  48         if err != nil {
  49                 reason += "; " + err.Error()
  50         }
  51
  52         return appended, reason
  53 }
  54
  55 // append appends the entries and stops at the first stop rule.
  56 func (r *BuildRun) append(ctx context.Context, entries []DatedMessage) (int, string) {
  57         for n, entry := range entries {
  58                 if entry.Kind == "user" {
  59                         reason := r.wait(ctx)
  60                         if reason != "" {
  61                                 return n, reason
  62                         }
  63                 }
  64
  65                 if ctx.Err() != nil {
  66                         return n, ctx.Err().Error()
  67                 }
  68
  69                 _, err := r.memory.AppendDated(entry.Kind, entry.Text, entry.Date)
  70                 if err != nil {
  71                         return n, err.Error()
  72                 }
  73
  74                 reason := r.check()
  75                 if reason != "" {
  76                         return n + 1, reason
  77                 }
  78         }
  79
  80         return len(entries), r.wait(ctx)
  81 }
  82
  83 // wait returns when no build runs, or when a stop rule ends the run. It checks
  84 // the rules each second, so a run stops soon also during a long wait.
  85 func (r *BuildRun) wait(ctx context.Context) string {
  86         for {
  87                 reason := r.check()
  88                 if reason != "" {
  89                         return reason
  90                 }
  91
  92                 err := r.waitBuilds(ctx)
  93                 if err == nil {
  94                         return r.check()
  95                 }
  96
  97                 if ctx.Err() != nil {
  98                         return ctx.Err().Error()
  99                 }
 100         }
 101 }
 102
 103 // check returns why the run must stop, or "" when it can go on.
 104 func (r *BuildRun) check() string {
 105         if r.client.LimitReached() {
 106                 return fmt.Sprintf("call limit of %d reached", r.client.CallLimit())
 107         }
 108
 109         records, err := buildReadUsage(r.usagePath)
 110         if err != nil {
 111                 return err.Error()
 112         }
 113
 114         return buildStop(records)
 115 }
 116
 117 // waitBuilds waits up to one second without leaving a child context open.
 118 func (r *BuildRun) waitBuilds(ctx context.Context) error {
 119         ctx, cancel := context.WithTimeout(ctx, time.Second)
 120         defer cancel()
 121
 122         return r.memory.Wait(ctx)
 123 }
 124
 125 // DatedMessage is one message with an optional RFC 3339 date.
 126 type DatedMessage struct {
 127         Kind string `json:"kind"`
 128         Text string `json:"text"`
 129         Date string `json:"date,omitempty"`
 130 }
 131
 132 // buildReadUsage reads the records of usage.jsonl at path, in the order of the
 133 // start of their calls. The file has them in the order of the end of their
 134 // calls, which differs when calls run at once.
 135 func buildReadUsage(path string) (records []ModelUsageRecord, err error) {
 136         file, err := os.Open(path)
 137         if errors.Is(err, os.ErrNotExist) {
 138                 return nil, nil
 139         }
 140
 141         if err != nil {
 142                 return nil, fmt.Errorf("open usage file: %w", err)
 143         }
 144
 145         defer func() {
 146                 closeErr := file.Close()
 147                 if closeErr != nil {
 148                         err = errors.Join(err, fmt.Errorf("close build usage: %w", closeErr))
 149                 }
 150         }()
 151
 152         scanner := bufio.NewScanner(file)
 153         scanner.Buffer(make([]byte, 64*1024), buildLineBytes)
 154         for scanner.Scan() {
 155                 var record ModelUsageRecord
 156                 err = json.Unmarshal(scanner.Bytes(), &record)
 157                 if err != nil {
 158                         return nil, fmt.Errorf("decode usage: %w", err)
 159                 }
 160
 161                 records = append(records, record)
 162         }
 163
 164         err = scanner.Err()
 165         if err != nil {
 166                 return nil, fmt.Errorf("read usage file: %w", err)
 167         }
 168
 169         slices.SortStableFunc(records, func(a, b ModelUsageRecord) int {
 170                 return buildCallStart(a).Compare(buildCallStart(b))
 171         })
 172
 173         return records, nil
 174 }
 175
 176 // buildWriteReport writes the result of a build run.
 177 func buildWriteReport(out io.Writer, appended int, stop string, records []ModelUsageRecord, sizes []int) error {
 178         var input, cached, output, failed int
 179         for _, record := range records {
 180                 input += record.InputTokens
 181                 cached += record.CachedTokens
 182                 output += record.OutputTokens
 183                 if record.Error != "" {
 184                         failed++
 185                 }
 186         }
 187
 188         var lines []string
 189         if stop != "" {
 190                 lines = append(lines, "stopped early: "+stop)
 191         }
 192
 193         lines = append(lines,
 194                 fmt.Sprintf("messages: %d", appended),
 195                 fmt.Sprintf("calls: %d (%d failed), %.2f per message", len(records), failed, buildRatio(len(records), appended)),
 196                 fmt.Sprintf("model lines: %d, retries: %d, %.2f calls per model line", len(sizes), len(records)-failed-len(sizes), buildRatio(len(records), len(sizes))),
 197                 fmt.Sprintf("tokens: %d input, %d cached (%.1f%%), %d output", input, cached, 100*buildRatio(cached, input), output),
 198         )
 199
 200         if len(sizes) > 0 {
 201                 lines = append(lines, buildSizeText(sizes))
 202         }
 203
 204         if len(records) > 0 {
 205                 lines = append(lines, buildTimeText(records))
 206         }
 207
 208         if len(records) > 0 {
 209                 lines = append(lines,
 210                         "quota at the first call: "+buildQuotaText(records[0]),
 211                         "quota at the last call: "+buildQuotaText(records[len(records)-1]),
 212                 )
 213         }
 214
 215         _, err := io.WriteString(out, strings.Join(lines, "\n")+"\n")
 216         if err != nil {
 217                 return fmt.Errorf("write build report: %w", err)
 218         }
 219
 220         return nil
 221 }
 222
 223 // buildStop returns why a build run must stop: a failed call, or a quota window
 224 // whose used percent grew by more than buildQuotaLimit points since the first
 225 // call. It returns "" when the run can go on.
 226 func buildStop(records []ModelUsageRecord) string {
 227         for _, record := range records {
 228                 if record.Error != "" {
 229                         return "call failed: " + record.Error
 230                 }
 231         }
 232
 233         if len(records) == 0 {
 234                 return ""
 235         }
 236
 237         first := buildUsedPercents(records[0])
 238         for name, used := range buildUsedPercents(records[len(records)-1]) {
 239                 start, found := first[name]
 240                 if found && used-start > buildQuotaLimit {
 241                         return fmt.Sprintf("%s grew from %g to %g", name, start, used)
 242                 }
 243         }
 244
 245         return ""
 246 }
 247
 248 // buildCallStart returns the start of the call of record, or the zero time when
 249 // its date is not valid.
 250 func buildCallStart(record ModelUsageRecord) time.Time {
 251         start, err := time.Parse(time.RFC3339Nano, record.Date)
 252         if err != nil {
 253                 return time.Time{}
 254         }
 255
 256         return start
 257 }
 258
 259 // buildRatio returns a / b, or 0 when b is 0.
 260 func buildRatio(a, b int) float64 {
 261         if b == 0 {
 262                 return 0
 263         }
 264
 265         return float64(a) / float64(b)
 266 }
 267
 268 // buildSizeText describes the sizes of the model lines.
 269 func buildSizeText(sizes []int) string {
 270         sorted := slices.Sorted(slices.Values(sizes))
 271         total, over := 0, 0
 272         for _, size := range sorted {
 273                 total += size
 274                 if size > nodeBytes {
 275                         over++
 276                 }
 277         }
 278
 279         return fmt.Sprintf("line bytes: min %d, median %d, mean %.0f, max %d, %d over %d", sorted[0], sorted[len(sorted)/2], buildRatio(total, len(sorted)), sorted[len(sorted)-1], over, nodeBytes)
 280 }
 281
 282 // buildTimeText describes the time per call and how many calls ran at once. A
 283 // call runs from its date for its seconds.
 284 func buildTimeText(records []ModelUsageRecord) string {
 285         type event struct {
 286                 at    time.Time
 287                 delta int
 288         }
 289
 290         var events []event
 291         seconds := make([]float64, 0, len(records))
 292         total := 0.0
 293         for _, record := range records {
 294                 start, err := time.Parse(time.RFC3339Nano, record.Date)
 295                 if err != nil {
 296                         continue
 297                 }
 298
 299                 length := time.Duration(record.Seconds * float64(time.Second))
 300                 events = append(events, event{
 301                         at:    start,
 302                         delta: 1,
 303                 }, event{
 304                         at:    start.Add(length),
 305                         delta: -1,
 306                 })
 307
 308                 seconds = append(seconds, record.Seconds)
 309                 total += record.Seconds
 310         }
 311
 312         if len(events) == 0 {
 313                 return "time: no dated calls"
 314         }
 315
 316         slices.SortFunc(events, func(a, b event) int {
 317                 return cmp.Or(a.at.Compare(b.at), cmp.Compare(a.delta, b.delta))
 318         })
 319
 320         running, most := 0, 0
 321         for _, e := range events {
 322                 running += e.delta
 323                 most = max(most, running)
 324         }
 325
 326         wall := events[len(events)-1].at.Sub(events[0].at).Seconds()
 327         slices.Sort(seconds)
 328
 329         return fmt.Sprintf("time: %.0f s in all, %.1f calls per minute, %.1f s per call (median %.1f, max %.1f), calls at once: at most %d, %.1f on average", wall, 60*float64(len(seconds))/max(wall, 1e-9), total/float64(len(seconds)), seconds[len(seconds)/2], seconds[len(seconds)-1], most, total/max(wall, 1e-9))
 330 }
 331
 332 // buildQuotaText returns the quota headers of record, sorted by name.
 333 func buildQuotaText(record ModelUsageRecord) string {
 334         names := slices.Sorted(maps.Keys(record.Quota))
 335         parts := make([]string, len(names))
 336         for n, name := range names {
 337                 parts[n] = strings.TrimPrefix(name, "x-codex-") + "=" + record.Quota[name]
 338         }
 339
 340         return strings.Join(parts, " ")
 341 }
 342
 343 // buildUsedPercents returns the used percent of each quota window of record, by
 344 // header name.
 345 func buildUsedPercents(record ModelUsageRecord) map[string]float64 {
 346         used := make(map[string]float64)
 347         for name, value := range record.Quota {
 348                 if !strings.HasSuffix(name, "-used-percent") {
 349                         continue
 350                 }
 351
 352                 percent, err := strconv.ParseFloat(value, 64)
 353                 if err == nil {
 354                         used[name] = percent
 355                 }
 356         }
 357
 358         return used
 359 }