INC-198232 · Issue 685700
Tombstone cleanup modified to handle very large volume environments
Resolved in Pega Version 8.7
Cassandra nodes were periodically suffering from an out of memory condition. This was traced to tumbling windows creating too many tombstones for Cassandra to clean up effectively when running in a very high volume environment with a short tumbling time. This has been resolved by modifying the procedure away from using individual tombstones to instead use range tombstones which will continuously supersede the previous tombstone so there is never an opportunity for them to build up.
INC-200152 · Issue 689865
Performance improvements for action check-in
Resolved in Pega Version 8.7
The process of 'Saving As' an action to a higher ruleset, Saving, and then running Check-in was taking an excessive amount of time. This was traced to redundant page copy activities and performance has been improved by updating the implementation.
INC-201364 · Issue 690806
Prediction outcome response timing updated
Resolved in Pega Version 8.7
Predictions using a response timeout were not emitting a negative response ('NoResponse') when the specified waiting time expired. This was traced to the the outcome and response timeout values being overridden while triggering responses for multi stage predictions along with chained predictions. This has been resolved by modifying the flow to emit each outcome as it is received and by adding the dataflow trigger in the function so that it does not override the values in case of chained predictions.
INC-201366 · Issue 691508
Performance improvements for stale thread warnings
Resolved in Pega Version 8.7
Stale thread warnings were causing performance issues during dataflow run execution. Stale thread/slow component warnings are added as part of the dataflow execution when a processing thread takes more than 5 minutes to process a single dataflow record. The stacktrace of the dataflow thread is added as part of the warning for debugging purposes, but in some scenarios the stacktrace can become very large. This has been resolved by removing the stacktraces from the warning, improving the query logic, and adding the run ID to the exception method to assist if there is an error.
INC-179727 · Issue 704545
Modified batch requestor handling to ensure cleanup
Resolved in Pega Version 8.6.4
A large buildup of batch requestors was seen, and restarting the node did not clear it. Investigation showed that in the case of one class reading a DSS value, a Pega requestor was being created when it was unable to retrieve any Pega context, and this requestor was not cleaned up afterwards. To resolve this, an update has been made to the way the requestor and Pega context is being created along with ensuring it will be cleaned up properly after use.
INC-180246 · Issue 664947
Support for apostrophe added to keyword tokenization
Resolved in Pega Version 8.6.4
A keyword containing an apostrophe was not detected properly in Text extraction model. This has been resolved by updating the annotator used in the tokenization.
INC-192826 · Issue 685797
PMML updated for nested 'if'
Resolved in Pega Version 8.6.4
Running a model on the openscoring API and the Pega PMML model gave different results. This has been resolved by updating the handling for nested 'if' in PMML.
INC-194382 · Issue 697615
ADM Model management page updates
Resolved in Pega Version 8.6.4
When accessing the ADM model management landing page in an environment with a large number of nodes, an "UnresponsiveNodeException" error was thrown from the pylaststrategyresults activity by various nodes on different days. Investigation showed the ADM factory table was truncated, leading to inconsistent ADM data in the Last Responses Section. This has been resolved, and the ADM model management page has been made more resilient to node failures.
INC-196561 · Issue 688846
Natural Language Processing (NLP) topic categorization working consistently for German
Resolved in Pega Version 8.6.4
A few German words were not recognized by the Natural Language Processing (NLP) topic categorization if they were given in capital letters. This was traced to upper case keywords being lowercased during model creation time, but not during prediction time. To resolve this, all words will be lower-cased before the model analyzes the text.
INC-206109 · Issue 706238
Delayed Learning works correctly with Volume Constraints
Resolved in Pega Version 8.6.4
When running an outbound campaign using NBA Designers predictions and delayed learning in conjunction with volume constraints, the adaptive models ended up with incorrect feedback if the prediction had an Alternative label configured with a defined waiting window. This has been resolved with the addition of a new flag 'isVolumeConstraintEnabled' during dataflow creation which will check for volume constraint, and a filter in HandleResponses for that flag.