[ https://issues.apache.org/jira/browse/ZOOKEEPER-4681?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
krystal he updated ZOOKEEPER-4681: ---------------------------------- Description: Using a [tool|https://github.com/kry4tall/CC-ZOO358] that I modifyed from [Filip Niksic's zootester|https://github.com/fniksic/zootester] for testing ZooKeeper, I discovered the following scenario which causes uncommitted requests to be executed. Zab protocol has three rounds: PROPOSE, ACK, and COMMIT. My tool can isolate the PROPOSAL, ACK and COMMIT messages and collect the values of some variables of each server at the end of each round. Initially, start an ensemble with 3 servers called A, B, and C, and initialize 2 znodes called /key0 and /key1, and set them to 0 and 1 respectively. # Request to set /key0 to 1000 on 3 servers. # Isolate the ack messages of 2 servers. # Request to set /key0 to 1001 on 3 servers. # Stop all servers and restart them. # Check the contents of /key0 and /key1. Surprisingly, both requests worked,reading /key0 == 1000, /key1 == 1001. If we skip step 4, we can read "/key0 == null, /key1 == null" from A and B, and read "/key0 == 0, /key1 == 1001" from C. Value null is because the "ConnectionLoss" with znodes. However, 1000 and 1001 should not appear in any znode, because the first proposal cannot obtain enough ack, it cannot be committed. According to the source code of Zookeeper, the server will not commit the proposal of the second request, because there is a proposal that has not been committed before. was: Using a [tool|https://github.com/kry4tall/CC-ZOO358] that I modifyed from [Filip Niksic's zootester|https://github.com/fniksic/zootester] for testing ZooKeeper, I discovered the following scenario which causes uncommitted requests to be executed. Zab protocol has three rounds: PROPOSE, ACK, and COMMIT. My tool can isolate the PROPOSAL, ACK and COMMIT messages and collect the values of some variables of each server at the end of each round. Initially, start an ensemble with 3 servers called A, B, and C, and initialize 2 znodes called /key0 and /key1, and set them to 0 and 1 respectively. # Request to set /key0 to 1000 on 3 servers. # Isolate the ack messages of 2 servers. # Request to set /key0 to 1001 on 3 servers. # Stop all servers and restart them. # Check the contents of /key0 and /key1. Surprisingly, both requests worked,reading /key0 == 1000, /key1 == 1001. "ConnectionLoss" for Znode > Uncommitted requests have been executed > ---------------------------------------- > > Key: ZOOKEEPER-4681 > URL: https://issues.apache.org/jira/browse/ZOOKEEPER-4681 > Project: ZooKeeper > Issue Type: Bug > Components: quorum > Affects Versions: 3.5.8 > Reporter: krystal he > Priority: Major > Attachments: no-restart.patch, restart-all-server.patch > > > Using a [tool|https://github.com/kry4tall/CC-ZOO358] that I modifyed from > [Filip Niksic's zootester|https://github.com/fniksic/zootester] for testing > ZooKeeper, I discovered the following scenario which causes uncommitted > requests to be executed. > Zab protocol has three rounds: PROPOSE, ACK, and COMMIT. My tool can isolate > the PROPOSAL, ACK and COMMIT messages and collect the values of some > variables of each server at the end of each round. > Initially, start an ensemble with 3 servers called A, B, and C, and > initialize 2 znodes called /key0 and /key1, and set them to 0 and 1 > respectively. > # Request to set /key0 to 1000 on 3 servers. > # Isolate the ack messages of 2 servers. > # Request to set /key0 to 1001 on 3 servers. > # Stop all servers and restart them. > # Check the contents of /key0 and /key1. Surprisingly, both requests > worked,reading /key0 == 1000, /key1 == 1001. > If we skip step 4, we can read "/key0 == null, /key1 == null" from A and B, > and read "/key0 == 0, /key1 == 1001" from C. Value null is because the > "ConnectionLoss" with znodes. > However, 1000 and 1001 should not appear in any znode, because the first > proposal cannot obtain enough ack, it cannot be committed. According to the > source code of Zookeeper, the server will not commit the proposal of the > second request, because there is a proposal that has not been committed > before. -- This message was sent by Atlassian Jira (v8.20.10#820010)