Please note that all files on unix file system are basically hard links referring specific inode. If you make a hard link to a file, it means the inode has two referring names. When the SSTable is compacted and GCed, Cassandra "delete" the old SSTable but keep snapshot. Now the reference count to the inode become one. Big advantage of hard link is that you don't need copy the data. So snapshot completes very fast. If you need separate copy of the snapshot in different volume, you can write a script to copy them.
From iPhone On 2011/05/19, at 10:17, Sameer Farooqui <cassandral...@gmail.com> wrote: > Ahh.. yeah. And during a compaction a new SSTable is created with the merged > data. > > So, if I take a snapshot before compaction, the old SSTables won't be deleted > (b/c the snapshot hard links still have a reference to the files). > > But if I hadn't taken a snapshot before compaction, does compaction also > automatically delete the old SSTables? > > FYI - The O'Reilly Cassandra book has a really misleading definition of > snapshotting. It says that a snapshot makes a copy of the keyspace and saves > it to a separate database file. > > Is there a way to do a read from a snapshot? So, using a client like Hector, > can I request a read from a snapshot that I took like 2 weeks ago? > > It sounds like the main benefit of a snapshot is that it makes a bunch of > nicely organized hard links in a separate folder from a specific point in > time. > > So, taking snapshots to a different volume doesn't make sense since the hard > links can't span file systems. But it would be nice to have a feature where > the entire point-in-time copy of the SSTables can be copied to a different > volume. Currently if the data volume gets corrupted, the snapshots on it can > also get corrupted. > > On Wed, May 18, 2011 at 5:44 PM, Watanabe Maki <watanabe.m...@gmail.com> > wrote: > SSTables are immutable. Those won't changed once written to disk. > > From iPhone > > > On 2011/05/19, at 9:37, Sameer Farooqui <cassandral...@gmail.com> wrote: > >> As of 0.8.0, is it possible to take a Cassandra snapshot to a different >> volume (like a EBS volume dedicated for backups)? >> >> About a year ago, Jonathan Ellis said that this won't be implemented b/c >> snapshots are basically hard links: >> http://mail-archives.apache.org/mod_mbox/cassandra-commits/201002.mbox/%3c821546961.221891265939548009.javamail.j...@brutus.apache.org%3E >> >> But I don't fully understand that. If a snapshot is just a hardlink, won't >> the snapshot also change as new data is written to the SSTables? >> >> - Sameer >
