Distributed traces with OBI

Introduction

OBI supports distributed traces for applications with some limitations and kernel version restrictions.

Distributed tracing is implemented through the propagation of the W3C traceparent header value. OBI reads incoming context automatically. Outgoing network-level context propagation is disabled by default and must be enabled as described below.

With the applicable propagation mode enabled, OBI reads incoming trace context, tracks program execution, and adds traceparent to outgoing HTTP or gRPC requests. If an application already added traceparent, OBI uses that value instead of its own generated context. If OBI can’t find incoming context, it generates one according to the W3C specification.

For details about how OBI chooses the parent request when work moves across threads, goroutines, tasks, or event loops, see trace context association.

Compatibility

OBI supports distributed tracing and context propagation in the following configurations:

The versions listed here are the versions OBI explicitly supports for distributed tracing features. Other versions might also work, but they are not part of the documented support scope unless stated otherwise. In particular, the Go 1.18+ requirement here applies to distributed tracing and context propagation; other OBI Go library-level instrumentation has a lower minimum version.

Implementation

The trace context propagation is implemented in two different ways:

1. By writing the outgoing header information at network level
2. By writing the header information at library level for Go

Depending on the programming language your service is written in, OBI uses one or both approaches of context propagation. We use these multiple approaches to implement context propagation, because writing memory with eBPF depends on the kernel configuration and the Linux system capabilities granted to OBI. For more details on this topic, see our KubeCon NA 2024 talk So You Want to Write Memory with eBPF?.

The context propagation at network level is disabled by default and can be enabled by setting the environment variable OTEL_EBPF_BPF_CONTEXT_PROPAGATION=all or by modifying the OBI configuration file:

ebpf:
  context_propagation: 'all'

Context propagation at network level

The context propagation at network level is implemented by writing the trace context information in the outgoing HTTP headers as well at the TCP/IP packet level. HTTP context propagation is fully compatible with any other OpenTelemetry based tracing library. This means that OBI instrumented services correctly propagate the trace information, when sending to and receiving from services instrumented with the OpenTelemetry SDKs. We use Linux Traffic Control (TC) to perform the adjustment of the network packets, which requires that other eBPF programs that use Linux Traffic Control chain properly with OBI. For special considerations regarding Cilium CNI, consult our Cilium Compatibility guide.

For TLS encrypted traffic (HTTPS), OBI is unable to inject the trace information in the outgoing HTTP headers and instead it injects the information at TCP/IP packet level. Because of this limitation, OBI is only able to send the trace information to other OBI instrumented services. L7 proxies and load balancers disrupt the TCP/IP context propagation, because the original packets are discarded and replayed downstream. Parsing incoming trace context information from OpenTelemetry SDK instrumented services still works.

For gRPC, OBI injects a per-stream traceparent HPACK header. This works across programming languages and preserves distinct trace contexts for concurrent HTTP/2 streams. OBI does not use TCP options for gRPC because TCP options are connection-scoped and can’t represent multiple multiplexed streams. Generic non-gRPC HTTP/2 context propagation remains limited to Go library instrumentation.

If you need finer control, context_propagation also accepts headers, tcp, and headers,tcp. The former http alias has been removed. The deprecated ip value has no effect.

This type of context propagation works for any programming language and doesn’t require that OBI runs in privileged mode or has CAP_SYS_ADMIN granted. For more details, see the Distributed traces and context propagation configuration section.

Kubernetes Configuration

The recommended way to deploy OBI on Kubernetes with distributed tracing support at network level is as DaemonSet.

The following Kubernetes configuration must be used:

• OBI must be deployed as a DaemonSet with host network access (hostNetwork: true).
• The /sys/fs/cgroup path from the host must be volume mounted as local /sys/fs/cgroup path.
• The CAP_NET_ADMIN capability must be granted to the OBI container.

The following YAML snippet shows an example OBI deployment configuration:

spec:
  serviceAccount: obi
  hostPID: true # <-- Important. Required in DaemonSet mode so OBI can discover all monitored processes
  hostNetwork: true # <-- Important. Required in DaemonSet mode so OBI can see all network packets
  dnsPolicy: ClusterFirstWithHostNet
  containers:
    - name: obi
      resources:
        limits:
          memory: 120Mi
      terminationMessagePolicy: FallbackToLogsOnError
      image: 'docker.io/otel/ebpf-instrument:main'
      imagePullPolicy: 'Always'
      env:
        - name: OTEL_EXPORTER_OTLP_ENDPOINT
          value: 'http://otelcol:4318'
        - name: OTEL_EBPF_KUBE_METADATA_ENABLE
          value: 'autodetect'
        - name: OTEL_EBPF_CONFIG_PATH
          value: '/config/obi-config.yml'
      securityContext:
        runAsUser: 0
        readOnlyRootFilesystem: true
        capabilities:
          add:
            - BPF # <-- Important. Required for most eBPF probes to function correctly.
            - SYS_PTRACE # <-- Important. Allows OBI to access the container namespaces and inspect executables.
            - NET_RAW # <-- Important. Allows OBI to use socket filters for http requests.
            - CHECKPOINT_RESTORE # <-- Important. Allows OBI to open ELF files.
            - DAC_READ_SEARCH # <-- Important. Allows OBI to open ELF files.
            - PERFMON # <-- Important. Allows OBI to load BPF programs.
            - NET_ADMIN # <-- Important. Allows OBI to inject HTTP and TCP context propagation information.
      volumeMounts:
        - name: cgroup
          mountPath: /sys/fs/cgroup # <-- Important. Allows OBI to monitor all newly sockets to track outgoing requests.
        - mountPath: /config
          name: obi-config
  tolerations:
    - effect: NoSchedule
      operator: Exists
    - effect: NoExecute
      operator: Exists
  volumes:
    - name: obi-config
      configMap:
        name: obi-config
    - name: cgroup
      hostPath:
        path: /sys/fs/cgroup

If /sys/fs/cgroup is not mounted as a local volume path for the OBI DaemonSet some requests may not have their context propagated. We use this volume path to listen to newly created sockets.

Kernel version limitations

The network level context propagation incoming headers parsing generally requires kernel 5.17 or newer for the addition and use of BPF loops.

Some patched kernels, such as RHEL 9.2, may have this functionality ported back. Setting OTEL_EBPF_OVERRIDE_BPF_LOOP_ENABLED skips kernel checks in the case your kernel includes the functionality but is lower than 5.17.

Go context propagation by instrumenting at library level

This type of context propagation is only supported for Go applications and uses eBPF user memory write support (bpf_probe_write_user). The advantage of this approach is that it works for HTTP and HTTPS. For HTTP/2 and gRPC, OBI can inject context on new and reused HTTP/2 and gRPC connections when HTTPS isn’t used. Using bpf_probe_write_user requires granting OBI CAP_SYS_ADMIN or running it as a privileged container.

Instrument applications that use the Go Trace API

Starting with OBI v0.11.0, OBI can instrument applications that use the OpenTelemetry Go Trace API without registering an SDK. When this integration is active, OBI detects Trace API calls and exports the resulting manual spans alongside its eBPF spans. Applications that already register an OpenTelemetry SDK continue to manage and export their own SDK telemetry. Registering any global TracerProvider, including by calling otel.SetTracerProvider(auto.TracerProvider()), prevents this automatic activation.

OBI activates the integration only when all of the following conditions are met:

• The application uses a supported OpenTelemetry module version and checksum combination, without a module replacement.
• The executable and host use a supported 64-bit architecture.
• OBI can resolve the required symbols and field layouts.
• OBI has permission to use bpf_probe_write_user.

If any check fails, the Auto SDK remains inactive and spans created through the global Trace API remain non-recording. OBI’s eBPF instrumentation continues to operate independently. When OBI can detect Trace API calls, it can export partial synthetic spans that contain the span name, parent relationship, status, and some primitive attributes. These spans do not include the instrumentation scope, events, or requested span kind.

In OBI v0.11.0, the encoded payload for each span exported through the Auto SDK must not exceed 16 KiB. OBI does not emit a metric or log message when it activates the integration or drops an oversized payload.

Known limitations and follow-up work include head sampling, context handoffs, external and remote parents and TraceState, larger payloads and drop observability, and log enrichment.

For the supported combinations of module version, checksum, and architecture, see the activation eligibility matrix. You can also review the upstream Go Trace API example and the Auto SDK documentation.

Kernel integrity mode limitations

In order to write the traceparent value in outgoing HTTP/gRPC request headers, OBI needs to write to the process memory using the bpf_probe_write_user eBPF helper. Since kernel 5.14 (with fixes backported to the 5.10 series) this helper is protected (and unavailable to BPF programs) if the Linux Kernel is running in integrity lockdown mode. Kernel integrity mode is typically enabled by default if the Kernel has Secure Boot enabled, but it can also be enabled manually.

OBI automatically checks if it can use the bpf_probe_write_user helper, and enables context propagation only if it’s allowed by the kernel configuration. Verify the Linux Kernel lockdown mode by running the following command:

cat /sys/kernel/security/lockdown

If that file exists and the mode is anything other than [none], OBI cannot perform context propagation and distributed tracing is disabled.

Distributed tracing for Go in containerized environments (including Kubernetes)

Because of the Kernel lockdown mode restrictions, Docker and Kubernetes configuration files should mount the /sys/kernel/security/ volume for the OBI docker container from the host system. This way OBI can correctly determine the Linux Kernel lockdown mode. Here’s an example Docker compose configuration, which ensures OBI has sufficient information to determine the lockdown mode:

services:
  ...
  obi:
    image: 'docker.io/otel/ebpf-instrument:main'
    environment:
      OTEL_EBPF_CONFIG_PATH: "/configs/obi-config.yml"
    volumes:
      - /sys/kernel/security:/sys/kernel/security
      - /sys/fs/cgroup:/sys/fs/cgroup

If the /sys/kernel/security/ volume is not mounted, OBI assumes that the Linux Kernel is not running in integrity mode.

Go channel span links

OBI emits experimental receiver-side span links for supported work handoffs through Go channels. When both the send and receive sides have active OBI-generated spans, the receiver span links to the sender span. OBI doesn’t change trace IDs, parent-child relationships, or the sender span.

This behavior is enabled automatically with Go-specific tracing when OBI can resolve the target binary’s channel runtime offsets. Direct unbuffered and buffered handoffs through runtime.chansend1, runtime.chanrecv1, and runtime.chanrecv2 are supported. Channel operations through select aren’t supported. Disable Go-specific tracers to disable these probes; there is no separate channel-link option.

OBI honors OTEL_SPAN_LINK_COUNT_LIMIT and drops invalid, duplicate, and self-referential links.

Capture Node.js manual spans

Starting with OBI v0.12.1, OBI can capture spans that a Node.js application creates through @opentelemetry/api when the application has not registered an OpenTelemetry SDK. OBI exports these manual spans through its trace pipeline and correlates them with automatically captured server spans. If the application registers an SDK, OBI leaves span creation and export to that SDK.

This feature is disabled by default. Existing Config v1 deployments can enable it with nodejs.manual_spans: true or OTEL_EBPF_NODEJS_MANUAL_SPANS=true. Config v2 does not expose an equivalent field in v0.12.1. Continue migrating deployments to Config v2 rather than retaining Config v1 solely for this feature.

OBI must be able to reach the Node.js inspector, and the process must not register its own SIGUSR1 handler. Bundled copies of @opentelemetry/api that the CommonJS loader cannot reach are not captured. Automatic client spans are currently siblings of manual spans under the same server span, rather than children of the active manual span.

Python asyncio with uvloop

Starting with v0.7.0, OBI supports context propagation for Python asyncio workloads running on uvloop. This enables distributed tracing of asynchronous Python services that use the uvloop event loop, in addition to the standard asyncio support.

The context propagation at network level applies to Python applications running on uvloop, allowing OBI to automatically instrument and propagate trace context for asynchronous operations. No additional configuration is required beyond enabling context propagation as described in the introduction.

To use OBI with Python asyncio and uvloop, ensure your Python application is configured to use uvloop as the event loop implementation.