We are running multiple test iterations of test doing "ungraceful" shutdown and scaling down a pod. The test is using 3 byteman rules to sleep for 3 seconds during transaction processing (AT EXIT prepare, AT ENTRY commit, AT EXIT commit).
The EAP is correctly started in migration pod every time and the data are consistent in the databases. The problem is that sometimes there are remaining records in pg_prepared_xacts after recovery finished successfully.

See the attached logs. In case of iteration_0_-xa-load-migration-1-g97ds there were remaining records in the table opposite to iteration_2_-xa-load-migration-1-tfdqd when it was correctly cleaned up.