Sentry and Datadog rework how telemetry leaves the app

Getting telemetry past the blockers

The week's most consequential change came from Sentry JavaScript, which shipped a handful of updates led by a managed tunnel the browser SDK derives from the DSN. Setting tunnel: true now builds a tunnel path automatically, and serveTunnelIntegration() answers those requests out of the box in Astro, Nuxt, Nitro, React Router, Remix 3, SolidStart, SvelteKit and TanStack Start — the ad-blocker workaround that used to need hand-written plumbing. The same line also instrumented Cloudflare Clairic calls as gen_ai.extract spans, with an Extraction tab that highlights extracted values in the trace view. The Next.js tunnel route handler completes the picture, forwarding only the event payload without request headers, though it comes with a breaking cap on @opentelemetry/api below 1.10.0 and the OTel SDK below 3.0.0. Earlier groundwork dropped sampled-out errors before they claim a promise buffer slot and stopped sending client reports for breadcrumbs. Treat the span-origin rename as a migration item: rules, queries and alerts keyed to the old origins need updating.

Over on the Python side, span streaming became the default tracing model in the first beta of the next major, retiring transactions and replacing the flat send_default_pii with a finer-grained data_collection option. With the stable release planned for November, the 2.x migration guide is worth reading now.

Matching monitoring depth to the workload

Datadog's answer to over-monitoring is Infrastructure Basic, a lightweight tier that collects core CPU, memory, disk, network, uptime and load metrics plus selected processes — and nothing more. Basic hosts coexist with Pro and Enterprise hosts in the same org, sharing dashboards and 15 months of full-resolution data, so a back-office box no longer has to carry enterprise-grade collection.

The rest of Datadog's week was dogfooding and AI tracing. Its Agent Observability audio capture records per-turn input and output audio alongside transcripts, token usage and tool calls for OpenAI Realtime API apps, converting PCM16 and G.711 to playable WAV right in the trace, with Amazon Nova 2 Sonic now traced in Python and Node.js. Datadog also explained how its own Insights tool cut agent costs and diagnosed tool-call failures — 80,000-token search results, cache-breaking context ordering, malformed retries — and showed AI Guard flagging self-replicating prompt injections planted in email and CSV inputs.

Tracers under the hood

dd-trace-js shipped twice, no longer crashing under Vite bundling and instrumenting hooked built-ins in bundled builds while adding a global exposure-logging switch and manual agent versioning, after an earlier update added OpenAI cache-write tokens split by TTL, the OpenAI Decisions API, and Postgres pool wait time. dd-trace-py fixed a nastier problem: forked child processes that hung, leaked memory and stalled their parent when several threads forked at once, plus dynamic instrumentation stalls. Its 4.17 release candidate removed the legacy pytest plugin, added Python 3.15, and brought aiohttp SSRF prevention. The Agent itself moved to Oracle 19c as the minimum external client.

Converting, scraping, and sharing

Dash0 pushed telemetry in new directions: SignalControl rules can now record a sum, gauge or exponential histogram from a span or log attribute, counted before sampling, and CloudWatch users gained interval scraping with tag filtering as a Metric Streams alternative under the same OTel names. Its Web SDK added opt-in session replay correlated with backend traces, masked by default, and AI SDLC Insights connected GitLab merge requests to linked coding sessions.

Axiom, meanwhile, made agent output portable: MCP links now open the exact result an agent returned, free and kept six hours, with new ingestion skills and Amp and ChatGPT notifiers. Traces there gained a Service Map and Statistics view that charts every hop and groups spans by service, name, kind or status. Smaller items round out the week: Sentry published its search syntax grammar as a standalone package and added new UI components with conversation extraction spans, while Datadog RUM deduplicated Next.js views across discarded mounts.