Function endLLMRequestPerfettoSpan

src/utils/telemetry/perfettoTracing.ts:468–685

(
  spanId: string,
  metadata: {
    ttftMs?: number
    ttltMs?: number
    promptTokens?: number
    outputTokens?: number
    cacheReadTokens?: number
    cacheCreationTokens?: number
    messageId?: string
    success?: boolean
    error?: string
    /** Time spent in pre-request setup (client creation, retries) before the successful attempt */
    requestSetupMs?: number
    /** Timestamps (Date.now()) of each attempt start — used to emit retry sub-spans */
    attemptStartTimes?: number[]
  },
)

 * End an API call span with response metadata
 */
export function endLLMRequestPerfettoSpan(
  spanId: string,
  metadata: {
    ttftMs?: number
    ttltMs?: number
    promptTokens?: number
    outputTokens?: number
    cacheReadTokens?: number
    cacheCreationTokens?: number
    messageId?: string
    success?: boolean
    error?: string
    /** Time spent in pre-request setup (client creation, retries) before the successful attempt */
    requestSetupMs?: number
    /** Timestamps (Date.now()) of each attempt start — used to emit retry sub-spans */
    attemptStartTimes?: number[]
  },
): void {
  if (!isEnabled || !spanId) return

  const pending = pendingSpans.get(spanId)
  if (!pending) return

  const endTime = getTimestamp()
  const duration = endTime - pending.startTime

  const promptTokens =
    metadata.promptTokens ?? (pending.args.prompt_tokens as number | undefined)
  const ttftMs = metadata.ttftMs
  const ttltMs = metadata.ttltMs
  const outputTokens = metadata.outputTokens
  const cacheReadTokens = metadata.cacheReadTokens

  // Compute derived metrics
  // ITPS: input tokens per second (prompt processing speed)
  const itps =
    ttftMs !== undefined && promptTokens !== undefined && ttftMs > 0
      ? Math.round((promptTokens / (ttftMs / 1000)) * 100) / 100
      : undefined

  // OTPS: output tokens per second (sampling speed)
  const samplingMs =
    ttltMs !== undefined && ttftMs !== undefined ? ttltMs - ttftMs : undefined
  const otps =
    samplingMs !== undefined && outputTokens !== undefined && samplingMs > 0
      ? Math.round((outputTokens / (samplingMs / 1000)) * 100) / 100
      : undefined

  // Cache hit rate: percentage of prompt tokens from cache
  const cacheHitRate =
    cacheReadTokens !== undefined &&
    promptTokens !== undefined &&
    promptTokens > 0
      ? Math.round((cacheReadTokens / promptTokens) * 10000) / 100
      : undefined

  const requestSetupMs = metadata.requestSetupMs
  const attemptStartTimes = metadata.attemptStartTimes