Details
Major Description
It's pretty clear that it's impossible to fix atomics with "Read Repair" atomically since it's impossible to lock entries during the repair process.
Even get from backups has no guarantee to return consistent values under load.
But to fix we must also perform an additional step - cache put.
So, value can be changed between gets, can be changed after gets but before put, but it still seems to be possible to automize the fix.
Idea is to decide what entry won on the last check attempt and put this value using the entry processor.
During the entry processor execution, we should check the current node's value, and if the value is as it was during the check we must replace it with the consistent value.
Attachments
Issue Links
- links to
