[ https://issues.apache.org/jira/browse/MESOS-6818?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15763184#comment-15763184 ]
Anindya Sinha commented on MESOS-6818: -------------------------------------- To handle this case, we can determine if disks are empty or not at agent startup and looking at the contents of the checkpoint info at that time. If there is a disk that contains old data, there are 2 approaches that can be done: a) Do not allow the agent to register with the master. It would therefore require an operator to clean/format the disk (or change the static resources pertaining to these unclean disks. b) Allow the agent to register with disks that have clean contents only. Say, the --resources flag indicates 4 disks of which 1 of the disk is unclean (i.e. old data exists), the agent advertises only 3 disks as available resources. > Ensure empty contents in persistent volume if disk not in use. > -------------------------------------------------------------- > > Key: MESOS-6818 > URL: https://issues.apache.org/jira/browse/MESOS-6818 > Project: Mesos > Issue Type: Improvement > Components: general > Reporter: Anindya Sinha > Priority: Minor > Labels: persistent-volumes > > Currently, when a persistent volume is DESTROYed, we ensure that `rmdir` of > that directory is successful before returning that disk as an offereable > resource. So, when a CREATE is received for the same path, we thereby ensure > that the contents of that path is empty (MOUNT disks will have the root > present though). This ensures that frameworks of the same role do not leak > content to other frameworks. Reference: MESOS-5448. > However, an agent might join a cluster who do not have the path corresponding > to the disks clean. One of the scenarios might be if the agents join a new > cluster after being repurposed from another cluster without their disks being > formatted. In such a case, the CREATE of a persistent volume on an existing > path might contain data left over from a previous run in the old cluster > which might lead to inconsistencies esp if the tasks running in the container > *always* expect to receive a clean persistent volume. -- This message was sent by Atlassian JIRA (v6.3.4#6332)