Changelog: Guarded rollouts for AI Configs, Observability SDKs, Trace Viewer, and more...
Guarded Rollouts for AI Configs
We introduced Guarded Rollouts for AI Configs, making it easier and safer to rollout changes to your AI-powered features in production.
LLMs are powerful, but unpredictable. A small prompt tweak or model swap can work great in staging but degrade user experience in production. That’s where Guarded Rollouts come in. You can now set up guardrails to monitor important metrics - like increases in completion errors - and automatically rollback to a previous variation if that error count spikes in production.
Positive/negative feedback
Tokens in/out
Latency
Completion error/success rates
Learn more in our AI Configs docs and Autogenerated Metrics guide
Catching AI Feature Issues with Guarded Rollouts 🚀 - Watch Video
New Server-side SDKs for Observability
We added new SDK support for Node.js, Python, and React Native in LaunchDarkly Observability, making it even easier to capture traces, logs, and sessions across your full stack.
To learn more, check out the docs:
A brand new Trace Viewer
Our new Trace Viewer, designed to give you full context in a single view, has arrived! Open multiple spans, search across traces, and instantly jump to related logs or session replays.
See our Trace Viewer in action 🚀 - Watch Video
Other improvements
- Added support for uninformed priors to Multi-Armed Bandit
- Added base SDK to install flow for Python SDK
- Added close button for Check SDK
- Addressed a wide set of rule violations for Lint Fixes
- User-selected views are now saved per project, similar to environment preferences
- Rollouts can now be stopped directly from the rule card
- Variation approval reasons are now logged as part of AI-driven changes
- Improved metric group layout and error boundary handling
Bug fixes
- Flag Deletion Notifications now correctly reflect archived and active flags
- Fixed incorrect "from" values during multi-environment edits
- Restored always-roll-back-measured-rollouts-on-srm flag following user feedback
- Resolved 503 errors and webhook issues in multi-environment scenarios for custom approvals
- Corrected analytics percentages rounding logic in breakdown displays
- Fixed regression list ordering inconsistencies
- Fixed experimentation layers filter functionality
- Fixed per-page logic in session and error tables