Make change-queue handle continues steaks of failed patches together -------------------------------------------------------------------- Key: OVIRT-1852 URL: https://ovirt-jira.atlassian.net/browse/OVIRT-1852 Project: oVirt - virtualization made easy Issue Type: Improvement Components: Change Queue Reporter: Barak Korren Assignee: Barak Korren Lets assume we have a change queue testing projects 'a' and 'b' where patches had been submitted in the following order (from left to right): a1 a2 a3 a4 b1 a5 Lets further assume that a1 and b2 both introduced regressions to their respective projects, and a5 fixed the regression for project 'a'. When running with all changes the CQ test will fail, and it will then run the bisection process testing the following sets of changes (where all tests will fail): a1 a2 a3 a1 <- a1 will be removed from the queue following this a2 a3 a4 b1 a5 a2 a3 a2 <- a2 will be removed from the queue following this a3 a4 b1 a5 a3 a4 a3 <- a3 will be removed from the queue following this a4 b1 a5 a4 <- a4 will be removed from the queue following this b1 a5 b1 <- b1 will be removed from the queue following this a5 <- This run will finally succeed letting the build of 'a' go the 'tested'. As we can see for the case described we will need 14 test runs to clear the queue and find a stable state. We can optimize this using the fact that change queue can track individual projects as change streams, and furthermore, the queue knows in every point in time which change streams exhibited failure so far. What we can do is when calculating a bisection, check if the 1st change in the queue belongs to a know failing project, at this point test it individually rather then performing a bisection of the full queue. With the starting conditions as described above, the sets of changes being tested would be: a1 a2 a3 a4 b1 a5 a1 a2 a3 <- We are starting a normal bisection here because at this point there is not yet any knowledge about the failures in 'a' a1 <- a1 will be removed and the failure in 'a' will be recorded a2 a3 a4 b1 a5 a2 <- Now the mechanism described above kicks in and a3 is tested alone and removed a3 a4 b1 a5 a3 <- Same mechanism now causes a2 to be tested alone a4 b1 a5 a4 <- a4 will be removed from the queue following this b1 a5 b1 <- b1 will be removed from the queue following this a5 <- This run will finally succeed letting the build of 'a' go the 'tested'. So a total of 12 test runs - we save 2 test runs,. This can be significant in terms of time saving and more runs an be saved the more failing patches we have.