Semi-automated view-change test #246
Conversation
|
The test sometimes fails. One issue is that clients continue submitting requests when replicas are stopped, so some replicas get stopped before they have chance to deliver all the requests being processed. I have a fix for that. But the test also happen to fail from time to time because it takes too long to handle messages communicated during view change. Remember that checkpoints are not implemented yet to limit the size of the protocol history (#218). Implementation of ViewChange and NewView messages is also simple, but inefficient (#216). |
The script starts replicas, as well as clients submitting requests, and periodically suspends some of them, thus triggering view changes. It checks that all replica processes stay alive and finally compares if they produced the same sequence of blocks. Signed-off-by: Sergey Fedorov <[email protected]>
sergefdrv
force-pushed
the
view-change-test
branch
from
40fe97a
to
1a2000b
Compare
|
Summary of changes:
|
|
The test is much more stable with lower number of rounds. It may still fail because view-change messages sometimes get too big and exceed gRPC limit. We may want to increase that limit in our sample application. |
This pull request resolves #243.