Product Updates

Online Evaluations is now Generally Available for AI Configs, enabling automated, model-based evaluation of LLM completions in both production and pre-production environments. This GA release also introduces Customizable Judges, allowing you to define evaluation prompts and scoring criteria aligned with your domain-specific requirements.

Online Evaluations attaches AI Judges to AI Config variations and emits structured evaluation metrics (for example, accuracy, relevance, or toxicity) for each completion. Judge results are surfaced in real time on the AI Config Monitoring page and can be used to detect regressions when prompts, models, or parameters change.

Customizable Judges give you fine-grained control over evaluation logic while maintaining consistent metric output for monitoring and analysis.

Docs

Metric Data Sources are now Generally Available for Warehouse Native Experimentation, making it easier to connect your existing warehouse tables to LaunchDarkly without forcing everything into a single fixed-schema table.

What’s new

With Metric Data Sources, you can:

• Connect one or more warehouse tables as a data source for experimentation
• Map your existing schema to the required experimentation fields (no rigid event table format)
• Use SQL queries to define and select the event data you want to measure—giving you more flexibility as your program scales across teams and products

Why it matters

Previously, Warehouse Native Experimentation required centralizing metric events into one table with a fixed schema, which limited how teams modeled and created metrics. Metric Data Sources remove that constraint, so you can bring your warehouse structure to LaunchDarkly instead of reshaping your data to fit LaunchDarkly. (This also unlocks key workflows for teams with mature, multi-table setups.)

Availability

It’s Generally Available for Warehouse Native Experimentation customers in the US region.

Learn more | Docs

Flag Delivery v2 (FDv2) is now available in Early Access for the LaunchDarkly Java Server-side SDK v7.11.0, bringing faster initialization, smarter data handling, and improved resiliency to feature flag delivery.

FDv2 introduces several key enhancements:

• Two-phase initialization: Start quickly from a polling data source, then synchronize from a long-lived streaming connection.
• Data-saving mode: Resume from a previously saved state during reconnects to reduce bandwidth and processing time.
• File-based initialization: Initialize from a local file first, then connect to LaunchDarkly APIs for updates.
• Automatic failover: Seamlessly switch to polling if the streaming connection fails.
• Custom configurations: Combine initializers and synchronizers to meet advanced redundancy or behavior requirements.
• Improved flag caching: When using a persistent store in non-daemon mode, flag data is no longer limited by cache TTL.

Together, these improvements make flag delivery faster, more reliable, and more adaptable across production environments.

Availability
Early Access: Data-saving mode is only available to members of LaunchDarkly’s Early Access Program (EAP).

Learn more
Data-saving mode documentation

Flag Delivery v2 (FDv2) is now available in Early Access for the LaunchDarkly Ruby Server-side SDK v8.12.0, bringing faster initialization, smarter data handling, and improved resiliency to feature flag delivery.

FDv2 introduces several key enhancements:

• Two-phase initialization: Start quickly from a polling data source, then synchronize from a long-lived streaming connection.
• Data-saving mode: Resume from a previously saved state during reconnects to reduce bandwidth and processing time.
• File-based initialization: Initialize from a local file first, then connect to LaunchDarkly APIs for updates.
• Automatic failover: Seamlessly switch to polling if the streaming connection fails.
• Improved flag caching: When using a persistent store in non-daemon mode, flag data is no longer limited by cache TTL.

Together, these improvements make flag delivery faster, more reliable, and more efficient across all environments.

Availability
Early Access: Data-saving mode is only available to members of LaunchDarkly’s Early Access Program (EAP). If you want access to this feature, join the EAP.

Learn more
Data-saving mode documentation

We’ve replaced the statistical model used under the hood for Guarded Releases. We changed from using Bayesian statistics to a new Frequentist sequential testing approach. You’ll get the same release experience, but fewer false alarms when we check for regressions repeatedly throughout the rollout.

What’s new

• New statistical model for Guardian: moved from Bayesian to Frequentist sequential testing plus multiple comparisons correction.
• Reduced false positive rate: Dramatic improvement in false positives, especially for percentile metrics.

Why it matters

As customers scale Guardian to run longer and guard more metrics, repeated checks can trigger false alarms, telling you there’s a regression even when things are actually fine.

With sequential testing and multiple comparisons correction, we intentionally make detection less sensitive so that false alarms become rare. The tradeoff in some cases is slower detection, but the outcome is consistency and much higher confidence of real regressions.

Learn more
Docs