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 // BuildRun appends messages, builds their tree, and stops the run by the rules
  26 // of buildStop.
  27 type BuildRun struct {
  28         memory    *Memory
  29         client    *ModelClient
  30         usagePath string
  31 }
  32
  33 // Run appends the entries. It waits for the builds only before a user message,
  34 // as a turn does, so up to buildMax builds run at once in between, and it waits
  35 // for all builds at the end. It returns the number of messages appended and why
  36 // the run stopped early, or "" when it did not. After a stop it waits for the
  37 // calls in flight, so no sent call is cancelled.
  38 func (r *BuildRun) Run(ctx context.Context, entries []DatedMessage) (int, string) {
  39         appended, reason := r.append(ctx, entries)
  40         if reason == "" {
  41                 return appended, ""
  42         }
  43
  44         err := r.memory.Wait(ctx)
  45         if err != nil {
  46                 reason += "; " + err.Error()
  47         }
  48
  49         return appended, reason
  50 }
  51
  52 // append appends the entries and stops at the first stop rule.
  53 func (r *BuildRun) append(ctx context.Context, entries []DatedMessage) (int, string) {
  54         for n, entry := range entries {
  55                 if entry.Kind == "user" {
  56                         reason := r.wait(ctx)
  57                         if reason != "" {
  58                                 return n, reason
  59                         }
  60                 }
  61
  62                 if ctx.Err() != nil {
  63                         return n, ctx.Err().Error()
  64                 }
  65
  66                 _, err := r.memory.AppendDated(entry.Kind, entry.Text, entry.Date)
  67                 if err != nil {
  68                         return n, err.Error()
  69                 }
  70
  71                 reason := r.check()
  72                 if reason != "" {
  73                         return n + 1, reason
  74                 }
  75         }
  76
  77         return len(entries), r.wait(ctx)
  78 }
  79
  80 // wait returns when no build runs, or when a stop rule ends the run. It checks
  81 // the rules each second, so a run stops soon also during a long wait.
  82 func (r *BuildRun) wait(ctx context.Context) string {
  83         for {
  84                 reason := r.check()
  85                 if reason != "" {
  86                         return reason
  87                 }
  88
  89                 err := r.waitBuilds(ctx)
  90                 if err == nil {
  91                         return r.check()
  92                 }
  93
  94                 if ctx.Err() != nil {
  95                         return ctx.Err().Error()
  96                 }
  97         }
  98 }
  99
 100 // check returns why the run must stop, or "" when it can go on.
 101 func (r *BuildRun) check() string {
 102         if r.client.LimitReached() {
 103                 return fmt.Sprintf("call limit of %d reached", r.client.CallLimit())
 104         }
 105
 106         records, err := buildReadUsage(r.usagePath)
 107         if err != nil {
 108                 return err.Error()
 109         }
 110
 111         return buildStop(records)
 112 }
 113
 114 // waitBuilds waits up to one second without leaving a child context open.
 115 func (r *BuildRun) waitBuilds(ctx context.Context) error {
 116         ctx, cancel := context.WithTimeout(ctx, time.Second)
 117         defer cancel()
 118
 119         return r.memory.Wait(ctx)
 120 }
 121
 122 // DatedMessage is one message with an optional RFC 3339 date.
 123 type DatedMessage struct {
 124         Kind string `json:"kind"`
 125         Text string `json:"text"`
 126         Date string `json:"date,omitempty"`
 127 }
 128
 129 // buildReadUsage reads the records of usage.jsonl at path, in the order of the
 130 // start of their calls. The file has them in the order of the end of their
 131 // calls, which differs when calls run at once.
 132 func buildReadUsage(path string) (records []ModelUsageRecord, err error) {
 133         file, err := os.Open(path)
 134         if errors.Is(err, os.ErrNotExist) {
 135                 return nil, nil
 136         }
 137
 138         if err != nil {
 139                 return nil, fmt.Errorf("open usage file: %w", err)
 140         }
 141
 142         defer func() {
 143                 closeErr := file.Close()
 144                 if closeErr != nil {
 145                         err = errors.Join(err, fmt.Errorf("close build usage: %w", closeErr))
 146                 }
 147         }()
 148
 149         scanner := bufio.NewScanner(file)
 150         scanner.Buffer(make([]byte, 64*1024), buildLineBytes)
 151
 152         for scanner.Scan() {
 153                 var record ModelUsageRecord
 154
 155                 err = json.Unmarshal(scanner.Bytes(), &record)
 156                 if err != nil {
 157                         return nil, fmt.Errorf("decode usage: %w", err)
 158                 }
 159
 160                 records = append(records, record)
 161         }
 162
 163         err = scanner.Err()
 164         if err != nil {
 165                 return nil, fmt.Errorf("read usage file: %w", err)
 166         }
 167
 168         slices.SortStableFunc(records, func(a, b ModelUsageRecord) int {
 169                 return buildCallStart(a).Compare(buildCallStart(b))
 170         })
 171
 172         return records, nil
 173 }
 174
 175 // buildWriteReport writes the result of a build run.
 176 func buildWriteReport(out io.Writer, appended int, stop string, records []ModelUsageRecord, sizes []int) error {
 177         var input, cached, output, failed int
 178
 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
 190         if stop != "" {
 191                 lines = append(lines, "stopped early: "+stop)
 192         }
 193
 194         lines = append(lines,
 195                 fmt.Sprintf("messages: %d", appended),
 196                 fmt.Sprintf("calls: %d (%d failed), %.2f per message", len(records), failed, buildRatio(len(records), appended)),
 197                 fmt.Sprintf("model lines: %d, retries: %d, %.2f calls per model line", len(sizes), len(records)-failed-len(sizes), buildRatio(len(records), len(sizes))),
 198                 fmt.Sprintf("tokens: %d input, %d cached (%.1f%%), %d output", input, cached, 100*buildRatio(cached, input), output),
 199         )
 200
 201         if len(sizes) > 0 {
 202                 lines = append(lines, buildSizeText(sizes))
 203         }
 204
 205         if len(records) > 0 {
 206                 lines = append(lines, buildTimeText(records))
 207         }
 208
 209         if len(records) > 0 {
 210                 lines = append(lines,
 211                         "quota at the first call: "+buildQuotaText(records[0]),
 212                         "quota at the last call: "+buildQuotaText(records[len(records)-1]),
 213                 )
 214         }
 215
 216         _, err := io.WriteString(out, strings.Join(lines, "\n")+"\n")
 217         if err != nil {
 218                 return fmt.Errorf("write build report: %w", err)
 219         }
 220
 221         return nil
 222 }
 223
 224 // buildStop returns why a build run must stop: a failed call, or a quota window
 225 // whose used percent grew by more than buildQuotaLimit points since the first
 226 // call. It returns "" when the run can go on.
 227 func buildStop(records []ModelUsageRecord) string {
 228         for _, record := range records {
 229                 if record.Error != "" {
 230                         return "call failed: " + record.Error
 231                 }
 232         }
 233
 234         if len(records) == 0 {
 235                 return ""
 236         }
 237
 238         first := buildUsedPercents(records[0])
 239         for name, used := range buildUsedPercents(records[len(records)-1]) {
 240                 start, found := first[name]
 241                 if found && used-start > buildQuotaLimit {
 242                         return fmt.Sprintf("%s grew from %g to %g", name, start, used)
 243                 }
 244         }
 245
 246         return ""
 247 }
 248
 249 // buildCallStart returns the start of the call of record, or the zero time when
 250 // its date is not valid.
 251 func buildCallStart(record ModelUsageRecord) time.Time {
 252         start, err := time.Parse(time.RFC3339Nano, record.Date)
 253         if err != nil {
 254                 return time.Time{}
 255         }
 256
 257         return start
 258 }
 259
 260 // buildRatio returns a / b, or 0 when b is 0.
 261 func buildRatio(a, b int) float64 {
 262         if b == 0 {
 263                 return 0
 264         }
 265
 266         return float64(a) / float64(b)
 267 }
 268
 269 // buildSizeText describes the sizes of the model lines.
 270 func buildSizeText(sizes []int) string {
 271         sorted := slices.Sorted(slices.Values(sizes))
 272         total, over := 0, 0
 273         for _, size := range sorted {
 274                 total += size
 275                 if size > nodeBytes {
 276                         over++
 277                 }
 278         }
 279
 280         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)
 281 }
 282
 283 // buildTimeText describes the time per call and how many calls ran at once. A
 284 // call runs from its date for its seconds.
 285 func buildTimeText(records []ModelUsageRecord) string {
 286         type event struct {
 287                 at    time.Time
 288                 delta int
 289         }
 290
 291         var events []event
 292
 293         seconds := make([]float64, 0, len(records))
 294         total := 0.0
 295         for _, record := range records {
 296                 start, err := time.Parse(time.RFC3339Nano, record.Date)
 297                 if err != nil {
 298                         continue
 299                 }
 300
 301                 length := time.Duration(record.Seconds * float64(time.Second))
 302                 events = append(events, event{
 303                         at:    start,
 304                         delta: 1,
 305                 }, event{
 306                         at:    start.Add(length),
 307                         delta: -1,
 308                 })
 309
 310                 seconds = append(seconds, record.Seconds)
 311                 total += record.Seconds
 312         }
 313
 314         if len(events) == 0 {
 315                 return "time: no dated calls"
 316         }
 317
 318         slices.SortFunc(events, func(a, b event) int {
 319                 return cmp.Or(a.at.Compare(b.at), cmp.Compare(a.delta, b.delta))
 320         })
 321
 322         running, most := 0, 0
 323         for _, e := range events {
 324                 running += e.delta
 325                 most = max(most, running)
 326         }
 327
 328         wall := events[len(events)-1].at.Sub(events[0].at).Seconds()
 329
 330         slices.Sort(seconds)
 331
 332         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))
 333 }
 334
 335 // buildQuotaText returns the quota headers of record, sorted by name.
 336 func buildQuotaText(record ModelUsageRecord) string {
 337         names := slices.Sorted(maps.Keys(record.Quota))
 338         parts := make([]string, len(names))
 339         for n, name := range names {
 340                 parts[n] = strings.TrimPrefix(name, "x-codex-") + "=" + record.Quota[name]
 341         }
 342
 343         return strings.Join(parts, " ")
 344 }
 345
 346 // buildUsedPercents returns the used percent of each quota window of record, by
 347 // header name.
 348 func buildUsedPercents(record ModelUsageRecord) map[string]float64 {
 349         used := make(map[string]float64)
 350         for name, value := range record.Quota {
 351                 if !strings.HasSuffix(name, "-used-percent") {
 352                         continue
 353                 }
 354
 355                 percent, err := strconv.ParseFloat(value, 64)
 356                 if err == nil {
 357                         used[name] = percent
 358                 }
 359         }
 360
 361         return used
 362 }