CURRENT file pointing to missing MANIFEST file. [JIRA: RIAK-1789] about leveldb

[engelsanchez]

It is possible that we are seeing the problem described in the All filesystems are not created equal paper. Basically, doing a sync on data files and atomic renames is not enough to ensure consistency of the set of files in the directory. The directory entry itself may need a sync in Linux systems to ensure that the file operations are executed in the expected order and the new file mods survive a crash.

[matthewvon]

@angrycub ... define "restart of a node" ... riak restart, linux restart, machine died and was restarted, etc.

[angrycub]

From the ticket narrative:

"At the same time this node was discovered down, two other nodes were being reported by ring-status as being down, but their beam processes were still running. Since the cluster already believed them unavailable, I killed the beam process on each of these nodes and restarted riak successfully. Currently all but one of the cluster members are up."

So I'm going to say, killed after being in an "indeterminate" state and then Riak was restarted. No indication that the physical nodes were rebooted; however, I can contact the user in question for more details if you'd like.

[matthewvon]

let me ponder this ... however, a vnode repair would fix. You already do that?

[angrycub]

That was how we addressed that particular ticket. Dumb shell script to look where the filename in the CURRENT file did not exist in that partitions folder and echo out the partition IDs. Ran eleveldb:repair on all of them and all was well.