diff --git a/packages/session/session-telemetry-otel/README.i18n.yaml b/packages/session/session-telemetry-otel/README.i18n.yaml index 40c194a4a5..63fd635885 100644 --- a/packages/session/session-telemetry-otel/README.i18n.yaml +++ b/packages/session/session-telemetry-otel/README.i18n.yaml @@ -27,8 +27,8 @@ en: 78d933a0f391cc58 zh: 0552999de92573b1 /deepseek-ai-dsh-session-telemetry-otel/use-this-package/failures-and-shutdown: - en: 1e5379dd884037db - zh: f9f53bcd044ecb17 + en: a66831cb9de57e04 + zh: 3ada28bb868e0932 /deepseek-ai-dsh-session-telemetry-otel/understand-the-implementation: en: ee672a6a6f0fa02d zh: c3bab8e6b60e768c diff --git a/packages/session/session-telemetry-otel/README.md b/packages/session/session-telemetry-otel/README.md index bc5a1b92f7..cfab2495a6 100644 --- a/packages/session/session-telemetry-otel/README.md +++ b/packages/session/session-telemetry-otel/README.md @@ -73,7 +73,7 @@ The backend measures each record once with the SDK OTLP JSON serializer, includi ### Failures and shutdown -Invalid byte limits, non-positive queue/timer values, `maxExportBatchSize > maxQueueSize`, invalid endpoints, and invalid shutdown deadlines fail at load. Processor defaults are 2,048 queued records, 512 records per request, a 1,000 ms scheduling delay, and a 30,000 ms request watchdog. Byte limits may split a count batch into multiple serial HTTP requests. `exportTimeoutMillis` warns per request but never frees an unsettled transport slot; the SDK transport owns network timeout/retry. Shutdown drains these requests until `shutdownTimeoutMillis`; expiry discards the remaining queue and prevents further sends, while an active request may still settle. Large prefixes can therefore remain partially unsent at CLI exit. +Invalid byte limits, non-positive queue/timer values, `maxExportBatchSize > maxQueueSize`, invalid endpoints, and invalid shutdown deadlines fail at load. Processor defaults are 2,048 queued records, 512 records per request, a 1,000 ms scheduling delay, and a 30,000 ms request watchdog. Byte limits may split a count batch into multiple serial HTTP requests. Each request waits for SDK export-queue cleanup after its callback before the next starts. `exportTimeoutMillis` warns per request but never frees an unsettled transport slot; the SDK transport owns network timeout/retry. Shutdown drains these requests until `shutdownTimeoutMillis`; expiry discards the remaining queue and prevents further sends, while an active request may still settle. Large prefixes can therefore remain partially unsent at CLI exit. ----- diff --git a/packages/session/session-telemetry-otel/README.zh.md b/packages/session/session-telemetry-otel/README.zh.md index 7ddd5dc13b..8acb57c385 100644 --- a/packages/session/session-telemetry-otel/README.zh.md +++ b/packages/session/session-telemetry-otel/README.zh.md @@ -73,7 +73,7 @@ kind: "package-reference" ### 失败与关闭 -无效字节上限、非正队列/计时值、`maxExportBatchSize > maxQueueSize`、无效 endpoint 或关闭期限在加载时失败。processor 默认队列为 2,048 条、单请求最多 512 条、调度延迟 1,000 ms、请求监测期限 30,000 ms。字节上限可能将按条数划分的批次拆成多个串行 HTTP 请求。`exportTimeoutMillis` 针对每个请求告警，但不会释放未结束的传输槽位；SDK transport 负责网络超时和重试。关闭时发送这些请求直到 `shutdownTimeoutMillis`，到期后丢弃剩余队列并禁止后续发送，在途请求仍可能结束。因此较大的授权前缀在 CLI 退出时可能只发送了一部分。 +无效字节上限、非正队列/计时值、`maxExportBatchSize > maxQueueSize`、无效 endpoint 或关闭期限在加载时失败。processor 默认队列为 2,048 条、单请求最多 512 条、调度延迟 1,000 ms、请求监测期限 30,000 ms。字节上限可能将按条数划分的批次拆成多个串行 HTTP 请求。每个请求在回调后等待 SDK 导出队列清理完成，才启动下一个请求。`exportTimeoutMillis` 针对每个请求告警，但不会释放未结束的传输槽位；SDK transport 负责网络超时和重试。关闭时发送这些请求直到 `shutdownTimeoutMillis`，到期后丢弃剩余队列并禁止后续发送，在途请求仍可能结束。因此较大的授权前缀在 CLI 退出时可能只发送了一部分。 ----- diff --git a/packages/session/session-telemetry-otel/src/session-log.ts b/packages/session/session-telemetry-otel/src/session-log.ts index d4dfca261c..6027f78dcb 100644 --- a/packages/session/session-telemetry-otel/src/session-log.ts +++ b/packages/session/session-telemetry-otel/src/session-log.ts @@ -133,6 +133,8 @@ class SessionLogProcessor implements LogRecordProcessor { const records = this.queue.splice(0, count).map(entry => entry.record) this.bytes -= bytes await this.send(records) + // The export callback precedes removal from the SDK concurrency queue. + await this.exporter.forceFlush() } } finally { this.active = undefined } } diff --git a/packages/session/session-telemetry-otel/tests/session-log.spec.ts b/packages/session/session-telemetry-otel/tests/session-log.spec.ts index 16dc53431d..8d6062a12c 100644 --- a/packages/session/session-telemetry-otel/tests/session-log.spec.ts +++ b/packages/session/session-telemetry-otel/tests/session-log.spec.ts @@ -106,13 +106,14 @@ it('preserves full nested events without truncation or ambient headers and retai it.each([CompressionAlgorithm.GZIP, CompressionAlgorithm.NONE])('bounds %s requests and preserves nested content', async (compression) => { const target = await collector() - const { sender } = reporter(target.endpoint, { exporter: { url: target.endpoint, compression } }) + const { sender, onFailure } = reporter(target.endpoint, { exporter: { url: target.endpoint, compression, concurrencyLimit: 1 } }) const records = [0,1,2].map(seq => sessionRecord('中"\\'.repeat(100000), seq)) for (const record of records) sender.reportSessionLog(record) await sender.shutdown() expect(target.captures.length).toBeGreaterThan(1) expect(target.captures.every(c => c.bytes <= SESSION_LOG_MAX_REQUEST_BYTES)).toBe(true) expect(contents(target.captures).map(parseContent)).toEqual(records.map(r => r.event)) + expect(onFailure).not.toHaveBeenCalled() }) it('admits an exact request and diagnoses an oversized record only once', async () => {