clintkitson
commented
9 years ago @vaibhavkhanduja thanks for the feedback. Steve should be posting some notes from the meeting. In general, we have laid out a short term plan to implement what is possible today with Mesos and external volumes considering the usage of the Mesos containerizer. Beyond that the points you are talking about are very relevant.
vaibhavkhanduja
First, they are very well written and researched. The approach you mention in phase 1 & 2 are very feasible. Some thoughts on what I thought may be use-ful, though I am sure team must be aware of them: a) The Persistent volume implementation is done using "dynamic reservable resource" allocation. Since it is dynamic, the slave process crash results in losing information of volumes. The slave today keeps the information checkpoint at its sandbox location. Any external volume linking might be impacted if the meta-information related external volume is lost. b) The Mesos is getting popular with Docker as executor, the implementation of Docker volume plugin is still very WIP. The Mesos scheduler do not have a way specify volume plugin information (other than parameters -> https://issues.apache.org/jira/browse/MESOS-3392). I expect more development in this areas, unless there is an alternative coming from Docker. c) There seems to be some buzz around having a volume manager connection for persistent volumes.