Marko Mäkelä added a comment - 2017-08-17 10:28 I believe I observed a similar failure when testing MDEV-13536 . I did touch this test in my work-around for MDEV-13542 , maybe in vain, because this test is abruptly killing the server, not doing a clean shutdown and restart. Looking at the server error log http://buildbot.askmonty.org/buildbot/builders/kvm-bintar-quantal-amd64/builds/6902/steps/test/logs/mysqld.1.err.4 I determined that this crash occurs on the first restart, after no_checkpoint_end.inc killed the server. This server startup is expected to fail, and it did fail: --source ../../suite/innodb/include/no_checkpoint_end.inc   --echo # restart --error 1 -- source include/start_mysqld.inc The only problem here is that we got an unexpected message about decompression failure. There is only a suppression about a decryption failure: call mtr.add_suppression("InnoDB: The page \[page id: space=[1-9][0-9]*, page number=[0-9]+\] in file test/t[1-4] cannot be decrypted"); The proper solution would seem to be to add the missing suppression. The reason why this message is being issued should be that with MDEV-13536 (and MDEV-12288 ), purge now has more work to do, and thus more redo log records can be emitted before the crash. Purge would clear the DB_TRX_ID columns of each INSERT. Before MDEV-12288 , purge did not have any access to any undo log generated by INSERT to an empty table. I was not able to repeat the failure locally, even with some small tricks.

Marko Mäkelä added a comment - 2017-08-18 12:06 I hope bb-10.2-marko will run fine with the added suppressions. The test is injecting faults, and it is not completely unplausible to get decompression or read errors in addition to decryption errors.