Product Updates

• Improved Static LP Configuration Support
  Advanced users can now define static link prediction modes via ModelPlan, enabling better control over leakage and modeling strategies.

• S3 to BigQuery Export Optimization
  Batch prediction results are now streamed from S3 to BigQuery to improve performance and prevent memory-related job failures.

​

#ingestion & connectors

• Databricks Connector Guardrails
  The system now prevents creation of Databricks connectors if backend support isn’t enabled, ensuring early validation and fewer runtime surprises.
• Connector Credential Retention on Update
  Updating a Databricks connector now preserves existing credentials, fixing a bug that previously caused silent prediction failures.
• Table Name Clean-Up and Collision Prevention
  Table names with repeated underscores are automatically normalized, and manual entry with multiple underscores is now blocked.
• Graph Link Suggestions After Table Edit
  Editing a table node within a graph now immediately triggers updated graph link suggestions, reducing friction in graph setup workflows. 
• Clear Feedback for Table Graph Errors
  Improved graph editing behaviors now include cleaner warnings, better name overflow handling, and consistent layout.

​

#training & evaluation

• Trainer Stability for Large Embedding Tables
  Trainer now exits gracefully when encountering large right-hand-side (RHS) tables that would previously cause CUDA out-of-memory crashes.
• PQ Validation Robustness with Multi-word Keywords
  Autocomplete logic in the predictive query editor has been upgraded to correctly suggest multi-word keywords like FOR EACH, avoiding syntax corruption.
• Support for Multi-Class Targets with Empty Strings
  Empty-string rows are now excluded from multi-class classification tasks, resolving CUDA indexing errors and ensuring class label consistency.
• Improved Error Messaging for Static Node Prediction
  Predictive query errors now provide helpful suggestions (e.g., use LAST() or aggregate values) when multiple matches are detected per entity.
• Better Split Statistics in TimeSplit
  The split stats logic now aligns with offset-based boundaries and non-strict modes, producing accurate distribution metrics across partitions.

​

#predictivequery

• Validation Skipped for Legacy Trained Queries
  Trained models with deprecated or renamed connectors can now be reviewed in the UI without triggering connector re-validation.

​

#batch prediction

• Prediction Anchor Time Consistency
  The Prediction Overview now accurately displays the backend-provided anchor time value, removing timezone-related confusion during batch prediction reviews. 

Improved Prediction on Non-Empty Columns
Users now receive actionable guidance when a prediction job fails because the prediction column is already populated.

• Predictive Query Validation Latency Reduction: Internal PQ validator now runs significantly faster under concurrent load, with latency reduced up to 25% across 20-100 parallel jobs.
• Connector API Call Reductions: Eliminated redundant connector API calls during Prediction creation and Graph View loads to improve page responsiveness.

#ingestion & connectors

• Databricks Connector Creation Guardrails: UI now blocks creation of multiple Databricks connectors to avoid backend stability issues.
• S3 Path Copy Button: Users can now easily copy S3 source paths directly from connector detail panels.

#training & evaluation

• Primary Key Auto-Suggestion: The system suggests likely primary keys based on column heuristics when users omit this field during table creation.
• Adaptive Sampling Neighbor Validation: Improved validation logic estimates safe node counts based on NumNeighbors and sampling config to prevent GPU OOM failures during training. Users receive clearer warnings when configurations exceed safe thresholds.
• Negative Weight Validation: Training jobs are blocked if weight columns contain negative values, preventing invalid model configurations.
• Improved  Metrics Table Display: The evaluation metrics table now uses the new UI Catalog Datatable for improved readability, displays baseline model scores for easier comparison, and clearly shows macro-averaged metrics like F1.
• Prediction Parallel Worker Limit Removal: UI no longer enforces hardcoded parallel worker limits; backend config fully controls allowable values.

#predictivequery

• Improved Error Readability: PQ validator now displays human-readable operators (e.g., >) instead of internal enum names (e.g., RelOp.GT).
• PQ Warning UX: PQ validation warnings are now surfaced even when queries are otherwise valid, ensuring users receive early guidance without blocking execution.

#UI/UX polish

• List Tooltip Behavior: Tooltips only show on truncated list items, eliminating unnecessary hover pop-ups.
• Graph View Auto-Layout Fixes: Dragged table positions in Graph View no longer auto-reset after idle refreshes.

• COUNT(fkey) Allowed in PQuery: COUNT operations on foreign keys are now permitted in grouping clauses.

• Faster Graph Snapshot: Batched metadata retrieval significantly reduced /snapshots/graph/id latency. 
• Databricks Row Count Optimization: Now uses warehouse queries instead of downloading files, improving performance.

#ingestion 

• Local File Upload Enhancements: Improved progress, formatting, preview, and connector auto-refresh after uploads.
• Connector Quotas & Limits UI: Displays connector-specific quotas and file upload rules in the UI.
• Duplicate Table Flow Fix: Clears validation errors after duplicating a table.
• Create Time Column Message: Clarified error messaging for invalid timestamp types in new tables.
• S3 Region Validation: Display error when S3 buckets are in a different region as the instance.
• Bad Timestamp Handling in Spark: Invalid timestamps are now mapped to null during ingestion.

#training

• F1 Score Display: Added F1@0.5 to model summary and dynamic row in Threshold Metrics.
• Block fkey Prediction in PQuery: Prevents selecting entity primary key as a prediction target.
• Detailed Static Node Errors: Validation now explains missing FK path and incorrect targeting more clearly.
• Weight Column Validation: Display errors on invalid weight values.

#xai

• XAI Frontend Updates: Enhanced subgraph labeling, aggregation display, and formatting for explainability views.
• Conditional Tab Display: “Subgraph Summary” tab is now hidden when empty and deep links reroute appropriately.

#batchprediction

• Empty DataFrame Handling in Parquet: Writes valid snappy.parquet files for empty DataFrames.

• Graph creation blocked if no primary key: Graph builder now prevents creation when no primary or foreign key is detected, ensuring better data integrity during setup.
• Snowflake access control respected: Users now see only the Snowflake tables they have read access to, reducing ingestion-related permission errors.
• Automatic renaming of unnamed columns: Local uploads automatically rename empty or generic column headers (e.g., “Unnamed: 0”) to avoid ingestion failures.
• Improved SDK reliability on Snowflake session expiry: The SDK now automatically retries generate_prediction_table() when a Snowflake session expires, preventing job crashes.
• Connector errors converted into readable UI failures: Missing or renamed source files now surface clear error messages in the UI instead of returning generic server errors.
• Model plan conflict warning for channels and aggregation: Users are now alerted when conflicting legacy and scoped configuration fields are defined in the same model plan.
• Improved empty state experience across the UI: Legacy placeholder views were replaced with polished UI Catalog components across connectors, tables, models, and predictions.
• Clearer error messaging for failed jobs: Job failure messages now specify the exact failed stage (e.g., Graph Snapshot, Prediction Table), making it easier to identify issues.
• Standardized input components across product: Legacy form fields were updated to use consistent UI Catalog components such as TextInput, TextArea, and Label across multiple workflows.
• Reduced snapshot polling frequency: Snapshot polling was reduced from every 2 seconds to every 30 seconds to improve system performance and reduce backend load.
• Training job settings no longer duplicated: Training configuration details are now shown only once in the job details view, improving visual clarity.
• Graph snapshot histogram charts restored: A prior patch was reverted to restore consistent rendering of histogram charts in the Graph Snapshot view.
• Tag filtering restored in V1 UI: Tag filters in the prediction jobs list now reliably return matching results.

• Multi-file upload rolled back: The “Add Table” button in Local Upload was removed to preserve a linear and streamlined onboarding experience; multi-upload is no longer supported in the connector panel.

• TimeSplit training logic and timestamp consistency: TimeSplit models now correctly apply boundary logic, and selected timestamp units are preserved across the pipeline to ensure reliable splits, fingerprints, and cache behavior.
• Improved timestamp handling and recognition: Users can now assign Timestamp types to string fields during table registration, and anchor times are displayed consistently in both UI and logs, eliminating manual fixes and formatting mismatches.
• Clearer error messaging across workflows: Users now see precise messages for connector path issues, missing or renamed files, type mismatches in graph validation, and filtered training datasets with zero rows.
• UI enhancements to graph creation and display: Placeholder graph names have been removed in favor of manual naming with validation, and graph creation now blocks if no primary/foreign key is detected to ensure data integrity.
• Batch prediction improvements: Prediction jobs now display custom tags, clearly identify UI-triggered jobs, exclude training weights to avoid memory overload, and hide ROC curves in degenerate cases for clarity.
• Consistent and polished input and list views: Form components and list pages have been updated for consistency in spacing, styling, and layout, and dropdowns were reorganized to surface high-priority actions like “Create New Graph.”
• Stable handling of schema and data changes: The UI now gracefully handles schema edits such as column renaming or removal, and ingestion jobs properly report on deleted files or empty row counts with actionable feedback.
• Parquet and CSV handling refinements: Header logic is now correctly scoped to file type, .parquet folder names no longer cause errors, and string-based column preview widths remain stable during data load.
• S3 and connector performance improvements: S3 file discovery supports max_items to avoid full scans, and updated connector panel messaging provides better guidance during empty or misconfigured states.
• Improved subgraph and explainability display: The XAI subgraph view now offers more complete and stable rendering of model behavior, improving interpretability.
• Miscellaneous usability and editor updates: YAML editors now support tab indentation, type names are human-readable in error messages, and dropdown behaviors and job tagging are more consistent across the UI.

• Integrated UI Catalog components into Kumo: Unified the app’s visual language by integrating Sidebar, ListOverview, status cards, and table search for a more consistent user experience.
• Generate Prediction button added to key views: Enabled quick access to prediction workflows from the Prediction List and Training Job Overview pages.
• Improved job orchestration for predictions: Enhanced logic for using parent sources, waiting on child jobs, and propagating job failure states correctly.
• Reduced excessive graph snapshot refreshes: Prevented redundant refreshes to reduce load and improve responsiveness in graph workflows.
• Async validation in graph workflows: Moved table.validate() into an async graph.validate() pipeline to improve performance and prevent blocking.
• Timestamp unit handling standardized: Inferred and preserved timestamp units across ingestion, graph refresh, and execution to ensure consistency.
• Split stats UI fixed and improved: Resolved broken displays in split table stats, restoring clarity and usability for time-based training workflows.
• Enhanced XAI subgraph and score display: Fixed local score display and improved completeness of subgraph views for explainability workflows.
• Batch prediction tagging and metadata exposed: Surfaced custom tags and UI-sourced job indicators to improve traceability in experiments.
• Prediction query engine enhancements: Added AutoML pruning, adaptive sampling, RFM query flag, and simplified syntax (e.g., “IN” over “IS IN”).
• Improved error messaging throughout the product: Delivered clearer messages for missing PKs, dtype mismatches, empty files, and normalized headers.
• Robust handling of renamed/missing columns: Enabled stable behavior in workflows when table schemas change mid-pipeline.
• Snowflake and S3 integration improvements: Enhanced error logging, credential handling, and S3 directory scanning for better reliability and performance.
• Removed deprecated model training APIs: Finalized migration to kumo-ml, added Trainer(checkpoint_path) support, and dropped legacy task dependencies.
• Updated Optuna and Databricks compatibility: Upgraded to Optuna 4.3.0 and updated Databricks table versioning to align with current backend infrastructure.