When leader becomes follower and has uncommitted entries it can't to commit entries from new leader.

NicolasT / kontiki

An implementation of the Raft consensus protocol

BSD 3-Clause "New" or "Revised" License

122 stars 15 forks source link

When leader becomes follower and has uncommitted entries it can't to commit entries from new leader. #15

Closed drchaos closed 10 years ago

drchaos commented 10 years ago

Test Case: Initial conditions: have launched 3 nodes (udp binaries) on different machines (physical, virtual or containers). There are some entries should be committed to the log.

Step 1: Stop networking on leader node.
Step 2: Wait until new leader commit some entries to log and old leader add some entries to his own log.
Step 3: Start networking on old leader node.

Expected: Old leader becomes follower and successfully replicates log. Actual: Old leader becomes follower but can't replicate log. It serves his own log with uncommitted entries.

qrilka commented 10 years ago

why are testcases in form of comments and not actual tests? :)

drchaos commented 10 years ago

Maybe, because cluster simulator isn't merged still? :) #11 And one thing: how can I simulate disconnecting network from one node?

jkozlowski commented 10 years ago

Heh that's probably because it was never finished, sorry :( I must say I am generally swamped by other things, as of a while now...

Sent from my iPhone

On 9 Sep 2014, at 14:09, "dr.Chaos" notifications@github.com wrote:

Maybe, because cluster emulator is still not merged? :) #11

— Reply to this email directly or view it on GitHub.

drchaos commented 10 years ago

I'am sorry. This is not an kontiki/raft issue. This is my error in leader handleAppendEntriesResponce code which was added during some optimization.