You'd have to check ColumnFamilyStore.open to see for sure. It might work that way, but if it does it's an accident -- seems totally fair for it to assume that there are no SSTables for a newly created CF.
One more question. I'm experimenting with the chef recipes for brisk. Upon the initial installation, the brisk recipe removes the existing system Keyspace if it's present. The new cluster will not have the" Spidertracks" keyspace in it's meta data, and will be a completely vanilla installation.
Can I leave the SSTable files in place, and create the keyspace and column families via the CLI without needing to restart the whole cluster? I.E, when a CF is added to the schema, if an SSTable for that CF exists, is it merged with the commitlog in the next compaction, and used for reads or is it ignored?
On Tue, 2011-08-02 at 22:17 -0500, Jonathan Ellis wrote:
Yes, that should work. For better/worse, CF metadata is stored in the system schema, not individual sstables. On Tue, Aug 2, 2011 at 10:00 PM, Todd Nine <email@example.com> wrote: > Hi Johnathan. Are these the correct steps for the upgrade? > > 1. Copy SSTable from 0.7.4 snapshot > 2. Drop CFS in all nodes in 0.7.4 cluster > 3. Upgrade to 0.8.2 and start cluster (with no clients connected) > 4. Create CFs using cli in cluster > 5. Shutdown all nodes > 6. Move the SSTables into place on all the nodes > 7. Restart the nodes > > This would definitely be my preferred approach, however I wasn't sure if > SSTables stored the CF meta data. > > Thanks, > Todd > > > On Tue, 2011-08-02 at 21:34 -0500, Jonathan Ellis wrote: > > If the bytes are compatible why not just use the existing sstable > files after redefining the schema? > > On Tue, Aug 2, 2011 at 8:06 PM, Todd Nine <firstname.lastname@example.org> wrote: >> Hi all, >> I seem to have backed myself into a corner and I cannot easily upgrade >> from a custom 0.7.4 installation to the default 0.8.2. Any help I could >> get >> would be greatly appreciated. Below is an outline of the problem. >> >> Current installation: >> 0.7.4 with Ed Anuff's custom composite comparators. >> >> https://github.com/riptano/hector-composite/commit/74407df3ef97c3fa90eae9f68d84b50be3d9938e >> CF Comprator Type : "DynamicComposite" >> >> >> New installation: >> 0.8.2 with Casasndra's native DynamicComposite column time. >> CF Comparator Type: >> >> "DynamicComposite(a=>AsciiType,b=>BytesType,i=>IntegerType,x=>LexicalUUIDType,l=>LongType,t=>TimeUUIDType,s=>UTF8Type,u=>UUIDType,A=>AsciiType(reversed=true),B=>BytesType(reversed=true),I=>IntegerType(reversed=true),X=>LexicalUUIDType(reversed=true),L=>LongType(reversed=true),T=>TimeUUIDType(reversed=true),S=>UTF8Type(reversed=true),U=>UUIDType(reversed=true))" >> >> >> I cannot simply upgrade Cassandra, this fails because the comparator is >> incorrectly defined for version 0.8.1 and on. My issue is that the column >> family definition has changed, the bytes that are stored are compatible, I >> simply need to re-define the CF and migrate the column data over. >> >> >> Initially I was thinking I could perform a read from the rows and columns >> from a single 0.7 node, then insert them into the 0.8 cluster, however I >> cannot have 2 different versions of the thrift API running in the same >> java >> JVM due to namespace conflicts. >> >> >> Idea is to perform the following steps, if anyone has any better >> suggestions, they would be greatly appreciated. >> >> 1. Back up all my CFs that use the dynamic composite, and copy the SSTable >> out of my Keyspace's data directory to tmp >> 2. Drop all CFs that use dynamic composite >> 3. Re-create the CFs with the new comparator definition >> 4. Using an external program directly read the 0.7 SSTables (without >> starting the cassandra daemon) and insert the rows and columns into the >> 0.8 >> cluster via thrift. >> >> >> Can anyone point me at a good example for reading rows and columns >> directly >> from the SSTables without using the thrift api? >> >> Thanks, >> Todd > > > >