The test performs a number of tiny updates on a single table, and after they are done, runs mariabackup --backup. On earlier versions, it works fine, but on 10.8 it fails almost reliably with
The value of innodb_log_file_size is indeed small, in order to minimize the test, but the issue is scalable. I started from the default 96M, with several tables with more columns, more keys, and bigger number of UPDATEs, but otherwise very similar load. I suppose it scales further up the same way, I didn't try.
It fails for me most of the times, both in shm and on SSD, but it's not deterministic and the result can vary on different machines and builds.
If it doesn't fail for you, try to re-run. If it continues not to fail, try to increase the number of cycles. It also seems to fail more willingly on faster (non-debug) builds.
Possibly it can be made more stable by more precise handling of the InnoDB log.
The failure started happening on 10.8 after this commit: