OBI changelog

Behavior changes in each OBI (OpenTelemetry eBPF Instrumentation) release that may affect existing dashboards, queries, span filters, and log parsers. Review the relevant version before upgrading. For current behavior, see the OBI overview and Exported metrics.

Oct 30, 2026

OBI v1.0.0

v1.0.0 is OBI's first stable release. It promotes the v0.15.2 instrumentation unchanged (no additional metrics, attributes, or feature flags) and establishes a compatibility contract for the v1 line.

Stable surfaces. The v1 contract covers the public Go API of go.opentelemetry.io/obi, Config v2 and its schema, documented CLI flags and environment variables, container and binary behavior, and the telemetry classified as stable (its OTLP and Prometheus forms). Within v1, minor releases add only backward-compatible changes and incompatible changes require a new major version. Telemetry classified as development, alpha, beta, or release candidate stays outside the contract, and network-flow attribute names remain excluded pending semantic-convention alignment.
Configuration v2 is the v1 configuration contract. Configuration v1 remains supported but is frozen and receives no fixes or feature work. Use obi config validate and obi config migrate to move to v2, and canary before rollout.
The telemetry schema advances to 1.0.0 (from 0.15.2), keeping the existing transformation history without additional renames.
The Collector receiver stays on its own v0 release line. v1.0.0 tags only the core module; the receiver remains go.opentelemetry.io/obi/collector at collector/v0.15.2. Pair them explicitly: go get go.opentelemetry.io/obi/collector@v0.15.2 go.opentelemetry.io/obi@v1.0.0. See OBI as an OpenTelemetry Collector receiver.
Known limitations. Application metrics can intermittently stay absent after a Pod starts or restarts until OBI is restarted; child spans can occasionally start before their parents; and port-based service discovery can select a helper or port-forwarder instead of the workload, so use executable-path, container, or Kubernetes selectors when several processes share a port.

The Coralogix otel-integration Helm chart ships OBI v1.0.0 starting with chart version 0.0.375.

If you are upgrading from before v0.15.2, review the v0.15.2 changes below first: span names, rpc.method values, and several default metric attributes changed.

Full release notes for OBI v1.0.0.

Oct 16, 2026

OBI v0.15.2

v0.15.2 supersedes the unpublished v0.15.0 and v0.15.1 releases and carries their changes. It is dominated by span-name changes that align OBI with the OpenTelemetry semantic conventions, and it is the instrumentation that v1.0.0 promotes to stable.

Span names across many protocols now follow the OpenTelemetry semantic conventions, which recreates span-metric series. Span names feed the span_name label on span metrics (traces_span_metrics_* and the legacy traces_spanmetrics_*), so these renames recreate the affected series. The ones most likely to affect existing queries: gRPC span names and rpc.method drop the leading slash (an unreadable path falls back to the span name grpc and method _OTHER, with the raw value kept in rpc.method_original); Redis span names no longer include the database namespace; AWS span names use S3., SQS., and SNS. prefixes; GraphQL server span names drop the GraphQL prefix; and GenAI client span names use the operation and request model, with Anthropic's message operation becoming chat. Database and ONC RPC span names also shift. Update dashboards, span filters, and queries that match old span names or rpc.method values; the gRPC and ONC RPC rpc.method changes also recreate the rpc.client.duration and rpc.server.duration series.
Three metric attributes become default, recreating their Prometheus series. Database duration metrics (client and server) now carry db.collection.name, gRPC client duration carries server.port, and failed gRPC server duration carries error.type. Each adds a Prometheus label (db_collection_name, server_port, error_type), so the affected series are recreated. For Elasticsearch and OpenSearch, db.collection.name can raise cardinality, exclude it through attributes->select if needed. See Exported metrics.
Configuration v2 protocol metric selection now covers span metrics and service graphs. Setting a protocol's enabled.metrics: false suppresses that protocol's span metrics and service graphs in both the OTLP and Prometheus outputs, not just its application metrics. The obi config migrate CLI reports any coverage it can't preserve when converting a Configuration v1 file; Configuration v1 behavior is unchanged.
HTTP client server.address drops the Host-header port. server.port and url.full keep it. Kubernetes Service hosts that include a port now resolve to the service and namespace, which can change peer and service-graph names.
Messaging attribute changes for MQTT and AMQP. MQTT SUBSCRIBE spans and process metrics use messaging.destination.subscription.name for the topic filter instead of the destination name, and shared subscriptions also emit messaging.consumer.group.name. AMQP OTLP metrics omit empty destinations and add a default messaging.operation.type. These add or shift messaging metric labels.
GraphQL, MCP, and JSON-RPC server spans drop their HTTP attributes (method, status code, route, URL, body size, user agent, and obi.http.response.observed), keeping the protocol, address, network, and error attributes. Captured HTTP headers are now limited to plain HTTP and GraphQL spans. AWS SNS spans drop rpc.system.name and rpc.method, keeping their SNS. span names and cloud.region.
Added metrics and instrumentation. .NET CPU (dotnet.process.cpu.count, dotnet.process.cpu.time) and last-collection heap size (dotnet.gc.last_collection.heap.size) runtime metrics; expanded .NET and Rails route harvesting; an opt-in url.template attribute on HTTP client spans; opt-in AWS Route53 resolution of peer names from hosted-zone records; and internal ring-buffer write counters (obi.bpf.ringbuf.writes, obi.bpf.ringbuf.write.failures) in the opt-in eBPF metrics family. See Exported metrics.
Reliability and security fixes. OBI no longer injects trace context into sockets misclassified as HTTP/2, OTLP traces are encoded as JSON when the protocol is http/json, distinct metric label tuples no longer collide, duplicate target.info counts are fixed when a process moves to an existing service, and the Collector receiver keeps running after its start context ends. Security updates bump go.opentelemetry.io/otel/sdk to v1.45.0, golang.org/x/crypto to v0.52.0, golang.org/x/net to v0.55.0, and go-git to v5.19.2.
Go Trace API auto-instrumentation gains an embedded Auto SDK path. OBI can now activate the Auto SDK embedded in go.opentelemetry.io/otel/trace without the separate go.opentelemetry.io/auto/sdk module. See Go Trace API auto-instrumentation.
The Collector receiver ships as a separate Go module. It is now go.opentelemetry.io/obi/collector, with its own release line, and keeps the existing import path. Collector build manifests must require that module. See OBI as an OpenTelemetry Collector receiver.

Full release notes for OBI v0.15.2.

Oct 2, 2026

OBI v0.14.0

Several span attributes are now opt-in. service.peer.name, http.request.body.size, http.response.body.size, and obi.http.response.observed are no longer emitted by default. Select the ones you need through attributes->select->traces->include. Service-graph generation that relies on service.peer.name must select it explicitly. See Exported metrics.
HTTP metrics carry error.type on failed requests. HTTP duration and body-size metrics now add error.type by default when a request fails, so the corresponding Prometheus series gain an error_type label (empty for successful requests). This recreates the existing series, so update dashboards, alerts, and recording rules that aggregate them.
HTTP body-size histograms are now opt-in behind application_sizes. The application metrics feature still includes them, but application_red does not. Add application_sizes (or keep application) if you need http.*.body.size.
host.id leaves the span-metric data points. The value is unchanged on the resource, so target_info still carries host_id. Consumers that flatten OTLP span metrics to Prometheus and key on a per-series host_id should read it from target_info instead; OBI's own Prometheus exporter is unaffected.
Empty-valued span attributes are omitted instead of emitted as empty strings. Update span filters and queries that match empty-string values. Manual span attributes, resource attributes, and metric labels keep their existing behavior.
External trace-to-profile correlation is now opt-in. The pinned traces_ctx_v1 map is no longer populated by default. Set ebpf.populate_trace_context: true (env OTEL_EBPF_BPF_POPULATE_TRACE_CONTEXT=true) if you correlate traces with external profiles or relied on the Go channel span-link fallback. The log enricher and the Node.js manual-span bridge enable it automatically, so trace-log correlation is unaffected.
Elastic Cloud database metrics split by cluster. Elastic Cloud client spans and database-operation metrics now carry db.namespace from the X-Found-Handling-Cluster header, so existing series split per cluster. AWS SQS spans no longer carry aws.extended_request_id.
Added runtime and protocol metrics. A full .NET runtime metrics family (requires .NET 8+), a JVM jvm.gc.duration histogram, the Node.js v8js.resource.active gauge, MCP operation and session duration histograms, GPU/CUDA metrics, and an obi.stat.tcp.successful.connections counter. See Exported metrics.
Broader instrumentation coverage. Adds AWS SNS instrumentation, HTTP route harvesting for Java (now enabled by default on Linux), Django, FastAPI, Flask, Rails, .NET, Symfony, Laravel, and Slim, generic Python asyncio server context propagation, and uprobe/kprobe attachment without CAP_SYS_ADMIN on kernels between 5.14 and 6.6 (memory-level Go context propagation still requires CAP_SYS_ADMIN).

Full release notes for OBI v0.14.0.

Sep 4, 2026

OBI v0.13.0

service.name and service.namespace are no longer default metric labels. They remain resource attributes, available through target_info and the OTLP resource, but they no longer decorate metric series by default. Dashboards, queries, and alerts that group or filter metrics by service.name or service.namespace must now select them explicitly through the attributes->select configuration. See Exported metrics.
error.type values are unified across spans and metrics. The attribute name is unchanged, but its values are now consistent between spans and metrics. Queries and span filters that match specific error.type values may need updating.
Python, JVM, and Node.js runtime metrics are expanded. With the application_runtime feature enabled, Python services export CPython garbage-collection metrics, JVM services add class, thread, and CPU metrics alongside the existing memory metrics, and Node.js services add V8 garbage-collection and heap-space metrics. See Exported metrics.
Database and messaging metrics carry additional attributes. Database duration metrics gain db.namespace, db.response.status_code, and server.port; messaging spans gain messaging.operation.name.
Context-propagation and protocol reliability fixes. A group of memory-safety fixes stops the HTTP header injector from writing a traceparent into an unrelated TLS stream (which could reset the connection), bounds message-buffer reads so unrelated kernel memory cannot appear in span payloads, and invalidates stale message buffers. Application-supplied traceparent headers are preserved in Go HTTP/2, TCP client and server roles are classified by the local port (fixing swapped network.tcp.handshake.role values for same-namespace clients), established idle connections are no longer counted in obi.stat.tcp.failed.connections, and DNS spans follow the OpenTelemetry conventions with IPv6 reverse-DNS support.

Full release notes for OBI v0.13.0.

Aug 22, 2026

OBI v0.12.2

Log enrichment is limited to standard output and standard error. Earlier versions intercepted every pipe write from an instrumented process, which could blank out shell command substitutions and, in some cases, stall shells. OBI now enriches (and suppresses the original of) only writes to a tracked process's standard output or standard error; any other pipe (a command substitution, application IPC, or a pipe on another file descriptor) passes through untouched. See Trace-log correlation.
application_span and application_span_sizes span-metric features are deprecated. They keep working and still emit the Grafana-convention traces_spanmetrics_* metrics, but OBI logs a deprecation warning at startup. Move to application_span_otel, which emits the OpenTelemetry-convention traces_span_metrics_* metrics. application_span_sizes has no direct replacement. The closest metrics are http.server.request.body.size and http.server.response.body.size. Enabling a legacy span-metric feature together with application_span_otel is rejected, and the * (all) selector keeps only the OTel version. See Exported metrics.
Reliability and accuracy fixes. HTTP/2 headers are decoded in capture order, preventing HPACK dynamic-table drift from attaching the wrong header values to a span; TLS connections that use in-memory BIOs are correlated again; and Aerospike responses that span multiple TCP reads can be reassembled, if you see missed or truncated Aerospike spans, raise the capture buffer with ebpf.buffer_sizes.aerospike.

Full release notes for OBI v0.12.2.

Aug 20, 2026

OBI v0.12.1

OBI v0.12.0 was never published, v0.12.1 supersedes it and carries its changes.

Server-side database operations report db.server.operation.duration. Redis, Memcached, and SQL measured on the server side now record db.server.operation.duration (db_server_operation_duration_seconds in Prometheus) instead of db.client.operation.duration. Update dashboards and alerts that read server-side database latency from the client metric. See Exported metrics.
Four internal eBPF metrics dropped an incorrect counter suffix. In OTLP, obi.bpf.map.entries_total, obi.bpf.map.max_entries_total, obi.bpf.network.packets.total, and obi.bpf.network.ignored.packets.total become obi.bpf.map.entries, obi.bpf.map.max_entries, obi.bpf.network.packets, and obi.bpf.network.ignored.packets. In Prometheus, the two map gauges lose the _total suffix (obi_bpf_map_entries, obi_bpf_map_max_entries); the packet counters keep it. These belong to the opt-in eBPF metrics family. See Exported metrics.
Node.js event-loop runtime metrics. With the application_runtime feature enabled, Node.js services export event-loop utilization, delay statistics (min, max, mean, standard deviation, and percentiles), and busy/idle time. See Exported metrics.
Node.js manual spans can be captured without an SDK. When a Node.js process creates spans through @opentelemetry/api but registers no OpenTelemetry SDK, set nodejs.manual_spans: true (env OTEL_EBPF_NODEJS_MANUAL_SPANS) to export them. It stays inactive when an SDK is already present.
Java service-name discovery follows the OpenTelemetry Java conventions. For a Java process without an explicit service name, OBI derives one from the Spring Boot application name, then the JAR manifest Implementation-Title, then the JAR file name, before falling back to java.

Full release notes for OBI v0.12.1.

Aug 17, 2026

OBI v0.11.0

Plain-text log enrichment is enabled by default. Every service selected by the log enricher now receives space-separated trace_id=… span_id=… fields on non-JSON writes, which previously passed through unchanged. This can break downstream parsers for structured non-JSON log formats. To keep the previous behavior, set ebpf.log_enricher.plain_text.enabled: false before upgrading. JSON enrichment is unaffected. In Configuration v2 the same setting is extensions.obi.correlation.log_trace_annotation.plain_text.enabled. See Trace-log correlation.
Direct OTLP consumers receive OpenTelemetry dot-notation metric names. target_info is now target.info, traces_target_info is now traces.target.info, traces_host_info is now traces.host.info, traces_span_metrics_calls_total is now traces.span.metrics.calls, and traces_span_metrics_duration is now traces.span.metrics.duration. Consumers that scrape through a collector's Prometheus exporter are unaffected, because it maps the dots back to underscores.
Prometheus metric names are now derived from their OTLP definitions, which corrected several of them. obi_stat_tcp_retransmits is now obi_stat_tcp_retransmits_total and obi_stat_tcp_failed_connections is now obi_stat_tcp_failed_connections_total; the CUDA GPU metrics lost an incorrect suffix, so gpu_cuda_kernel_grid_size_total is now gpu_cuda_kernel_grid_size, gpu_cuda_kernel_block_size_total is now gpu_cuda_kernel_block_size, and gpu_cuda_memory_copies_bytes_total is now gpu_cuda_memory_copies_bytes. Update Prometheus queries, dashboards, and alerts that key on the old names. See Exported metrics.
JVM runtime metrics moved to the shared application_runtime feature. The deprecated application_jvm feature alias and the jvm_runtime_metrics.enabled flag were both removed. Enable JVM and Go runtime metrics through application_runtime instead. The OBI-specific obi.jvm.heap.used metric (obi_jvm_heap_used_bytes in Prometheus) was dropped; read heap usage from jvm.memory.used filtered on jvm.memory.type=heap. See Exported metrics.
Configuration v2 is available for standalone OBI and the OBI Collector receiver, and Configuration v1 remains supported. obi config validate and obi config migrate check and convert existing configurations, failing closed on settings they can't preserve. Native Configuration v2 authoring uses a default-on capture policy: omitting capture.policy.default_action, or setting it to include, captures every process that no rule excludes. Canary the generated configuration before rolling it out, and see the Config v1 to v2 migration guide. The Helm chart values shown throughout these pages use Configuration v1.
Span names and span counts changed for SQL and Kafka. SQL span names follow the OpenTelemetry semantic conventions more closely, SQL type is detected on Go SQL spans, and a multi-topic Kafka Produce or Fetch request now emits one span per topic instead of one span per request. Aggregations that count Kafka spans or group on SQL span names may shift.
Go Trace API auto-instrumentation is activated. Go applications that use the global otel.Tracer API without registering an SDK are instrumented automatically. Go runtime telemetry also expands with CPU, memory, scheduler, GC, and histogram metrics. See Go Trace API auto-instrumentation.
gRPC and HTTP/2 context propagation covers connections that were already open when OBI started. Huffman-encoded traceparent values are decoded on ingress (which needs Linux kernel 5.17+), HTTP server context is preserved through response bodies, and Go net/http HTTP/1 clients no longer inject a duplicate traceparent. See Distributed tracing.
Protocol, runtime, and AI/LLM coverage expanded. Aerospike client calls are instrumented; Deno processes are identified as a JavaScript runtime and emit metrics and spans (trace-context propagation for Deno isn't available yet); and Ollama's native /api/chat and /api/generate endpoints and OpenAI-compatible gateways (LiteLLM, vLLM, LocalAI, OpenRouter, Ollama /v1/) are instrumented, with Gemini SSE streaming responses captured.

Full release notes for OBI v0.11.0.

Jun 30, 2026

OBI v0.10.0

gRPC/HTTP2 network-level context propagation is now supported. OBI propagates the W3C traceparent header on gRPC and HTTP/2 requests via HPACK header injection, without requiring memory-level writes. See Distributed tracing.
GenAI attributes aligned with OpenTelemetry semantic conventions. Message shapes for OpenAI, Anthropic, Google Gemini, Qwen, and AWS Bedrock have been normalized. Dashboards or queries keyed on prior attribute names may need updating.
url.query is now emitted by default on HTTP spans when present, and client url.full includes the query string. Sensitive query parameters are automatically redacted. The built-in redaction list covers AWS SigV4 (X-Amz-Signature, X-Amz-Credential, X-Amz-Security-Token), legacy AWS signed URLs (AWSAccessKeyId, Signature, SecurityToken), Google (X-Goog-Signature, sig), and common token names (token, access_token, refresh_token, id_token, jwt). Extend the list with OTEL_EBPF_SENSITIVE_QUERY_PARAMS_ADD, or trim it with OTEL_EBPF_SENSITIVE_QUERY_PARAMS_REMOVE.
GraphQL documents are no longer exported by default. Enable them explicitly with payload_extraction.http.graphql.enabled: true (env: OTEL_EBPF_HTTP_GRAPHQL_ENABLED).
Java route harvesting is disabled by default while JVM-blocking edge cases are investigated. The http.route attribute on Java spans falls back to the raw URL path until harvesting is re-enabled upstream.
ip context-propagation value is deprecated; the undocumented http alias has been removed. Valid values for OTEL_EBPF_BPF_CONTEXT_PROPAGATION are disabled, headers, tcp, and all. If you set ip, OBI warns and does not enable propagation. Use headers instead.
AI/LLM coverage expanded. Tool-call extraction is added for OpenAI, Anthropic, Gemini, and Qwen; MCP tool-call arguments and results are extracted with independent attribute-selection controls; SSE streaming is captured for OpenAI-compatible, Anthropic, and Qwen responses; and vector-database retrieval spans are emitted for Pinecone, Qdrant, Milvus, Zilliz, Chroma, and Weaviate.
New protocol and platform support. SunRPC/ONC RPC detection is added; automatic OpenShift cluster-name detection is added; JDK 26 and 27 are supported.

Full release notes for OBI v0.10.0.

May 11, 2026

OBI v0.9.0

Database error response reporting is now opt-in. OBI no longer captures database error response bodies by default. Enable it explicitly if you previously relied on those fields appearing on SQL spans.
Generic Go TLS support. TLS visibility extends beyond OpenSSL to custom Go protocols. See Distributed tracing.

Full release notes for OBI v0.9.0.

Apr 3, 2026

OBI v0.7.0+

SQL spans now use Server kind. Database spans for SQL instrumentations (MySQL, PostgreSQL, MariaDB) are emitted as SERVER spans instead of CLIENT. Update any trace or span filters that key on span kind for SQL operations.
DNS excluded from span and service-graph metrics by default. DNS calls no longer appear in span metrics or in the service graph. If you relied on DNS visibility there, expect those entries to disappear after upgrading.
Ordinal port guessing disabled by default for network metrics. OBI no longer infers a service identity from ordinal source/destination ports on network flows. Existing queries built on guessed ports may return different results.
HTTP header enrichment available as an opt-in feature. You can now capture selected HTTP request and response headers as span attributes. Configure under ebpf.payload_extraction.http.enrichment (env: OTEL_EBPF_HTTP_ENRICHMENT_ENABLED), disabled by default.

Full release notes for OBI v0.7.0 and v0.7.1.

Last updated on November 5, 2026