Samer Al-Kiswany created ZOOKEEPER-2018:
-------------------------------------------
Summary: Zookeper node fails to boot if writes are reordered
Key: ZOOKEEPER-2018
URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2018
Project: ZooKeeper
Issue Type: Bug
Affects Versions: 3.4.6
Reporter: Samer Al-Kiswany
After studying the steps ZooKeeper takes to update the logs we found the
following bug. The bug may manifest in file systems with writeback buffering.
If you run the zookeeper client script (zkCli.sh) with the following commands:
VALUE=”8KB value” # 8KB in size
create /dir1 $VALUE
create /dir1/dir2 $VALUE
the strace generated at the zookeeprer node is:
mkdir(v)
create(v/log)
append(v/log)
trunk(v/log)
…
fdatasync(v/log)
write(v/log) ……. 1
write(v/log) ……. 2
write(v/log) ……. 3
fdatasync(v/log)
The last four calls are related to the second create of dir2.
If the last write (#3) goes to disk before the second write (#2) and the system
crashes before #2 reaches the disk, the zookeeper node will not boot.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
