Sentry reworks log opt-in while Datadog battles Python memory leaks

Sentry redraws the Python opt-in line

The headline change this week is Sentry Python 2.68.0, which turns enable_logs and enable_metrics into no-ops ahead of their removal in the next major. The shift is more than deprecation: the logging and Loguru integrations previously gated automatic log collection behind the global enable_logs flag, and that behavior is gone. Teams now pass capture_sentry_logs=True at the integration level — off by default, so nothing auto-collects anymore. If you were relying on the old behavior, this release requires a deliberate action before anything ships. That's clearly a correctness-over-convenience call from Sentry, and it pairs naturally with the earlier 2.67.1 global span-count flush trigger preventing unbounded memory growth in the batcher.

The week's other Sentry Python release, v2.67.0, rounds out the AI story with reasoning-token tracking for Langchain, OpenAI tool definitions for Chat Completions and Responses APIs, and a promoted before_send_span option. It also fixes double-counting in Langchain multi-candidate responses and duplicate Pydantic AI spans — a good sign Sentry is treating AI observability as a first-class surface rather than an add-on.

Datadog's Python APM stability sprint

Datadog shipped an unusual run of five Python APM releases in the span of a week, and they tell one coherent story: memory and crash fixes. The recurring thread is a crash on Python 3.9–3.12.3 triggered when background worker threads clean up thread-local data, first addressed in dd-trace-py v4.11.5 (which also fixed an OpenFeature integration break with SDK 0.10.0) and carried through v4.10.12 and v4.12.3. The other recurring fix is memory growth in stack profiling for asyncio and gevent workloads, addressed in two threads including v4.11.6. The newest release, v4.14.0rc2, fixes traced Snowflake cursors when command is passed as a keyword argument — a subtle break that only surfaces with certain call patterns.

Beyond stability, v4.13.0 deprecates ml_app in favor of agent_service and moves AI Guard to a top-level ddtrace.aiguard package — worth flagging if you're on that path since the old import will die in the next major. And v4.14.0rc1 makes Feature Flagging load agentless over HTTPS by default and adds list_experiments()/submit_feedback() for LLM Observability. Also worth noting: Datadog Agent 7.82.1 fixes a kernel panic on multi-GPU nodes with Hopper/Blackwell hardware.

Mobile and platform polish

On the mobile side, SentrySQLiteDriver is now stable in the Android SDK 8.53.0, giving Room 2.7+/3.0+ apps SQL query spans — either through the Gradle plugin auto-wrap or by wrapping your driver directly. Meanwhile Sentry's Cocoa SDK 9.22.0+ captures the last 100 feature flag evaluations per error and flags suspects when a flag correlates with a spike, with webhook integrations for LaunchDarkly, Statsig, Unleash, and Flagsmith. That's a genuinely useful debugging surface for flag-related incidents.

Datadog also announced secretless authentication for its Azure integration, replacing client-secret rotation with federated identity trust — available during onboarding and via migration paths through Azure CLI, Terraform, or the portal. Simpler credential management is easy to undervalue until rotation bites.

SaaS quick hits

Sentry's AI Query Assistant UX now shows status indicators and a regenerate button, and dashboards get pre-starred templates plus per-user Recently Viewed. The JavaScript SDK released 10.70.0 with MCP SDK v2 support and Cloudflare Durable Object fixes, alongside an early 11.0.0-alpha.1 migration preview. Axiom's Go SDK added IsPartial/IsEstimate to query status and UpdatedAt fields plus a URL fix. Dash0's Agent0 write-action approvals give an org-wide pause before any destructive action.