为 LangSmith 遥测配置 collector

LangSmith 部署中的各项服务会以日志、指标和追踪的形式发出遥测数据。你的 Kubernetes 集群中可能已经设置了遥测 collector，或者你想部署一个来监控应用。

本页介绍如何配置 OTel Collector 来收集 LangSmith 的遥测数据。注意，下面讨论的所有概念也可以迁移到其他 collector，例如 Fluentd 或 FluentBit。

WARNING

本节仅适用于 Kubernetes 部署。

接收器

日志

这是一个 Sidecar collector 从自身 pod 读取日志的示例，会排除非领域特定容器的日志。这里使用 Sidecar 配置很有用，因为我们需要访问每个容器的文件系统。也可以使用 DaemonSet。

filelog:
  exclude:
    - "**/otc-container/*.log"
  include:
    - /var/log/pods/${POD_NAMESPACE}_${POD_NAME}_${POD_UID}/*/*.log
  include_file_name: false
  include_file_path: true
  operators:
    - id: container-parser
      type: container
  retry_on_failure:
    enabled: true
  start_at: end
env:
  - name: POD_NAME
    valueFrom:
      fieldRef:
        fieldPath: metadata.name
  - name: POD_NAMESPACE
    valueFrom:
      fieldRef:
        fieldPath: metadata.namespace
  - name: POD_UID
    valueFrom:
      fieldRef:
        fieldPath: metadata.uid
volumes:
  - name: varlogpods
    hostPath:
      path: /var/log/pods
volumeMounts:
  - name: varlogpods
    mountPath: /var/log/pods
    readOnly: true

INFO

此配置需要对给定 namespace 中 pod 的 'get'、'list' 和 'watch' 权限。

指标

指标可以通过 Prometheus 端点抓取。使用单个 Gateway collector 实例可以避免抓取指标时的重复查询。以下配置会抓取所有默认命名的 LangSmith 服务：

prometheus:
  config:
    scrape_configs:
      - job_name: langsmith-services
        metrics_path: /metrics
        scrape_interval: 15s
        # Only scrape endpoints in the LangSmith namespace
        kubernetes_sd_configs:
          - role: endpoints
            namespaces:
              names: [<langsmith-namespace>]
        relabel_configs:
          # Only scrape services with the name langsmith-.*
          - source_labels: [__meta_kubernetes_service_name]
            regex: "langsmith-.*"
            action: keep
          # Only scrape ports with the following names
          - source_labels: [__meta_kubernetes_endpoint_port_name]
            regex: "(backend|platform|playground|redis-metrics|postgres-metrics|metrics)"
            action: keep
          # Promote useful metadata into regular labels
          - source_labels: [__meta_kubernetes_service_name]
            target_label: k8s_service
          - source_labels: [__meta_kubernetes_pod_name]
            target_label: k8s_pod
          # Replace the default "host:port" as Prom's instance label
          - source_labels: [__address__]
            target_label: instance

INFO

此配置需要对给定 namespace 中 pod、service 和 endpoint 的 'get'、'list' 和 'watch' 权限。

追踪

对于追踪，你需要启用 OTLP receiver。以下配置可用于监听 4318 端口的 HTTP 追踪，以及 4317 端口的 GRPC：

otlp:
  protocols:
    grpc:
      endpoint: 0.0.0.0:4317
    http:
      endpoint: 0.0.0.0:4318

处理器

推荐的 OTEL 处理器

使用 OTel collector 时推荐以下处理器：

• Batch Processor：将数据分组为批次后再发送给导出器。
• Memory Limiter：防止 collector 占用过多内存而崩溃。越过软限制后，collector 会停止接受新数据。
• Kubernetes Attributes Processor：向遥测数据中添加 Kubernetes 元数据，例如 pod 名称。

导出器

导出器只需指向你想要的外部端点。以下配置允许你为日志、指标和追踪分别配置端点：

otlphttp/logs:
  endpoint: <your_logs_endpoint>
otlphttp/metrics:
  endpoint: <your_metrics_endpoint>
otlphttp/traces:
  endpoint: <your_traces_endpoint>

NOTE

OTel Collector 也支持直接导出到 Datadog 端点。

示例 collector 配置：日志 sidecar

mode: sidecar
image: otel/opentelemetry-collector-contrib
config:
  receivers:
    filelog:
      exclude:
        - "**/otc-container/*.log"
      include:
        - /var/log/pods/${POD_NAMESPACE}_${POD_NAME}_${POD_UID}/*/*.log
      include_file_name: false
      include_file_path: true
      operators:
        - id: container-parser
          type: container
      retry_on_failure:
        enabled: true
      start_at: end
  processors:
    batch:
      send_batch_size: 8192
      timeout: 10s
    memory_limiter:
      check_interval: 1m
      limit_percentage: 90
      spike_limit_percentage: 80
  exporters:
    otlphttp/logs:
      endpoint: <your-endpoint>
  service:
    pipelines:
      logs/langsmith:
        receivers: [filelog]
        processors: [batch, memory_limiter]
        exporters: [otlphttp/logs]
env:
  - name: POD_NAME
    valueFrom:
      fieldRef:
        fieldPath: metadata.name
  - name: POD_NAMESPACE
    valueFrom:
      fieldRef:
        fieldPath: metadata.namespace
  - name: POD_UID
    valueFrom:
      fieldRef:
        fieldPath: metadata.uid
volumes:
  - name: varlogpods
    hostPath:
      path: /var/log/pods
volumeMounts:
  - name: varlogpods
    mountPath: /var/log/pods
    readOnly: true

示例 collector 配置：指标和追踪 Gateway

mode: deployment
image: otel/opentelemetry-collector-contrib
config:
  receivers:
    prometheus:
      config:
        scrape_configs:
          - job_name: langsmith-services
            metrics_path: /metrics
            scrape_interval: 15s
            # Only scrape endpoints in the LangSmith namespace
            kubernetes_sd_configs:
              - role: endpoints
                namespaces:
                  names: [<langsmith-namespace>]
            relabel_configs:
              # Only scrape services with the name langsmith-.*
              - source_labels: [__meta_kubernetes_service_name]
                regex: "langsmith-.*"
                action: keep
              # Only scrape ports with the following names
              - source_labels: [__meta_kubernetes_endpoint_port_name]
                regex: "(backend|platform|playground|redis-metrics|postgres-metrics|metrics)"
                action: keep
              # Promote useful metadata into regular labels
              - source_labels: [__meta_kubernetes_service_name]
                target_label: k8s_service
              - source_labels: [__meta_kubernetes_pod_name]
                target_label: k8s_pod
              # Replace the default "host:port" as Prom's instance label
              - source_labels: [__address__]
                target_label: instance
    otlp:
      protocols:
        grpc:
          endpoint: 0.0.0.0:4317
        http:
          endpoint: 0.0.0.0:4318
  processors:
    batch:
      send_batch_size: 8192
      timeout: 10s
    memory_limiter:
      check_interval: 1m
      limit_percentage: 90
      spike_limit_percentage: 80
  exporters:
    otlphttp/metrics:
      endpoint: <metrics_endpoint>
    otlphttp/traces:
      endpoint: <traces_endpoint>
  service:
    pipelines:
      metrics/langsmith:
        receivers: [prometheus]
        processors: [batch, memory_limiter]
        exporters: [otlphttp/metrics]
      traces/langsmith:
        receivers: [otlp]
        processors: [batch, memory_limiter]
        exporters: [otlphttp/traces]

---