OpenTelemetry Java agent 2.32.0 is out as the release candidate for 3.0, targeted for October 2026. The release lets teams preview changes to telemetry conventions, capture settings, and instrumentation defaults before they become the default behaviour.
Database and code semantic conventions become stable defaults, while messaging adopts newer conventions that remain experimental. Some behaviour may change before 3.0. Jay DeLuca, Staff Software Engineer at Grafana Labs, set out a test-deployment migration process.
Comparing database and code telemetry
DeLuca recommends capturing baseline telemetry before enabling preview settings. With OTEL_INSTRUMENTATION_COMMON_V3_PREVIEW=false, teams can set OTEL_SEMCONV_STABILITY_OPT_IN=database/dup,code/dup and OTEL_SEMCONV_STABILITY_PREVIEW=messaging/dup to emit old and new attributes together, plus metrics where dual emission is supported.
The comparison requires checking values and units, not simply replacing attribute names. In the post’s PostgreSQL JDBC example, a connection to the orders database using the public schema emits legacy db.name as orders, but stable db.namespace as orders|public. The example’s schema is not a migration requirement.
Other changes include db.system: “mssql” becoming db.system.name: “microsoft.sql_server”, while code.namespace and code.function consolidate into code.function.name. Database connection-pool duration metrics change from milliseconds to seconds, requiring corresponding threshold conversions.
Connection-pool metrics adopt new names and units even with database/dup. Each span also retains only one name and kind, with dual emission using the newer convention for both.
Setting OTEL_INSTRUMENTATION_COMMON_V3_PREVIEW=true enables the combined conventions and defaults, and disables dual emission for those domains. To isolate convention changes from default changes, DeLuca advises turning the umbrella flag off and removing /dup from the domain settings.
Database endpoint grouping also needs review. For supported clients, server.address describes the configured target, potentially a list of cluster endpoints, whereas network.peer.address identifies the endpoint actually contacted where available. Service-graph grouping can therefore change without an attribute rename.
Kafka trace relationships change
Messaging advances from semantic conventions v1.24 to v1.43. Kafka span names change from orders publish, orders receive, and orders process to send orders, poll orders, and process orders. The former messaging.operation attribute splits into operation name and type: a Kafka poll has name poll and type receive.
Parent relationships depend on consumer execution. For single-message processing with producer context propagated, an active application span becomes the process span’s parent in the preview; the process span links to the producer rather than taking it as its parent.
With receive spans enabled and no active consumer-side span, the process span instead becomes a child of the send span in the same trace. The poll span represents a separate client operation, changes kind from CONSUMER to CLIENT, and links to the send span. Batch processing can link to multiple messages and does not follow this single-message model.
Receive spans remain opt-in. In Kafka 2.32.0, enabling receive telemetry also enables poll-duration metrics. Kafka did not emit the listed legacy messaging instruments, so messaging/dup does not provide both metric sets for comparison.
Capture and exporter compatibility changes
The umbrella preview can export existing SLF4J structured key-value fields without separately enabling capture. MDC remains separately configured and opt-in. Common .included and .excluded selectors replace source-specific structured-field settings, which preview mode ignores. Their glob support means copying * from a literal-name capture list can broaden collection.
User identity capture remains opt-in but requires new settings. enduser.id becomes user.name, enduser.role changes from a comma-separated string to the user.roles string array, and enduser.scope has no replacement.
Hibernate, Hystrix, and Twilio instrumentation default to off and require explicit re-enablement where needed. Extension maintainers also face invokedynamic instrumentation becoming the default, changing instrumentation and helper-class loading.
Zipkin exporter support is removed in preview mode; configurations using it can fail agent/SDK initialisation and need to switch to OTLP.
See also: Atlassian moves 100,000-host metrics platform to OpenTelemetry

Want to learn more about cybersecurity from industry leaders? Check out Cyber Security & Cloud Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the AI & Big Data Expo. Click here for more information.
Developer is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.