XPFv3.6.0 Release Notes

Author: Luke Heberling, Created: 2025-02-17

Release Date: 17 Feb 2025

Release Overview

• Scale Enhancements

  • Various caching improvements to aid in stability as well as reduce the runtime of FOR jobs.
  • Database optimization, including but not limited to refactoring some commonly used large tables to reduce database resource use and improve query time.
  • 5% faster runtimes for all jobs. Much more stability around successful jobs, in particular rebuilds.
  • Faster button clicks when predictions and other jobs are running.
  • Project Copy runtime consistency improved.

• XAT Cleanup

  • Rename the Request Limit Permission to be the Web Service Limit permission. Clean up Error messages surrounding this permission to be more understandable for the end user.
  • You must be a User in XAT to be eligible to hit the Xperiflow API. This does not apply to the APIs that configure XAT.
  • Fix potential leftover scope after a failed project creation or project copy

• New service - LanguageModelService:

  • This submodule adds the ability to communicate with a variety of different AI models. Its split up into two main categories: Chat models, and Embedding models.
  • Embedding:
    • Added the OpenAI "text-embedding-3-small" model to generate embeddings
  • Chat:
    • Added a handful of OpenAI models that you can communicate with:
      • GPT-4o
      • GPT-4o-mini
  • New Resources
    • LLMChatResource
      • Communicate with chat models through REST api instead of service.
    • LLMEmbeddingResource
      • Communicate with embedding models through REST api instead of service

• Enhanced FOR Health Metrics

  • Revamped health metrics on a target-by-target basis to be less harsh if a target sees a level shift up
    • Changed health metric calculations from using Mean Absolute Error (MAE) to Mean Absolute Error Percentage (MAE%) to allow for models to see less harsh health drops due to high MAE levels for target level shifts.

• New Routines

  • Union Inner Join FVA
    • Reduces the complexity of creating complex SQL Join statements when preparing data for Forecast Value Add, with a the component workflow allowing for a multitude of options for how the join is performed

• Updated Routines

  • Pycaret Clustering
    • K-means clustering allows users to group their data into clusters by specifying a K value or allowing the routine to detect the optimal number of K groups based on the properties of the data. In the updated routine, users receive interactive graphs, including a 3D PCA plot with tooltips that highlight the most important attributes in each cluster, along with summary statistics and additional information.
  • Frequency Resample
    • Frequency Resampling now allows for multiple value columns to be aggregated at once, using independent functions on each value column. (ex. from Value_1, Value_2 -> Value1_sum, Value1_max, Value2_sum)
  • Target Flagging Routine
    • Displayed additional columns to allow easier understanding for Growth Rate calculations.
    • Revamped Growth Rate to be in terms of percentage.
      • Ex: beforehand if we had seen a growth rate of 10% it was displayed as 0.1 now, we will see it displayed as 10.
    • Reorganized columns to show all metrics first and model information second.
  • Anomaly Detectors and TimeSeries Cleaning
    • Improved consistency of the seasonality calculation when “auto” is chosen.

• Miscellaneous Bug Fixes/Improvements

  • Feature datasources not marked as requires_processing = True on rebuild, before 3.6 after doing a rebuild feature datasource attributes were not updated meaning to users it appeared as though the features were still in use and would not need to be processed before pipeline but this was not the case as they would not be used in the subsequent runs if they weren’t processed. This is fixed in 3.6 and now feature datasources will correctly show that requires_processing = True for all previously committed datasources.
  • Generator block fit parameter updates in utilization not used. Before 3.6 updating a generator’s fit parameters in utilization would not have any effect on the prediction (i.e., adding a new event instance wouldn’t cause the corresponding spike or dip that other instances of that event would cause). This behavior has been fixed and now new instances have an effect on the prediction.
  • Known in advance features are now automatically lagged.

Dependencies

Must be on OneStream Platform version 8.4

Known Issues

Solved in XPFv3.6.1 and XPFv3.6.2 Releases