Out of interest i've done some more digging. Not sure how much more I've contributed but here goes...

Ran this against an clean v 0.6.12 and it works (I expected it to fail on the first read)

    client = pycassa.connect()
    standard1 = pycassa.ColumnFamily(client, 'Keyspace1', 'Standard1')

    uni_str = u"数時間"
    uni_str = uni_str.encode("utf-8")
    
    print "Insert row", uni_str
    print uni_str, standard1.insert(uni_str, {"bar" : "baz"})

    print "Read rows"
    print "???", standard1.get("???")
    print uni_str, standard1.get(uni_str)

Ran that against the current 0.6 head from the command line and it works. Run against the code running in intelli J and the code fails as expected. Code also fails as expected on 0.7.5

At one stage I grabbed the buffer created by fastbinary.encode_binary in the python generated batch_mutate_args.write() and it looked like the key was correctly utf-8 encoded (matching bytes to the previous utf-8 encoding of that string).

I've updated the git project https://github.com/amorton/cassandra-unicode-bug 

Am going to leave it there unless there is interest to keep looking into it. 
-----------------
Aaron Morton
Freelance Cassandra Developer
@aaronmorton

On 8 May 2011, at 13:31, Jonathan Ellis wrote:

Right, that's sort of a half-repair: it will repair differences in
replies it got, but it won't doublecheck md5s on the rest in the
background. So if you're doing CL.ONE reads this is a no-op.

On Sat, May 7, 2011 at 4:25 PM, aaron morton <aaron@thelastpickle.com> wrote:
I remembered something like that so had a look at RangeSliceResponseResolver.resolve()  in 0.6.12 and it looks like it schedules the repairs...

           protected Row getReduced()
           {
               ColumnFamily resolved = ReadResponseResolver.resolveSuperset(versions);
               ReadResponseResolver.maybeScheduleRepairs(resolved, table, key, versions, versionSources);
               versions.clear();
               versionSources.clear();
               return new Row(key, resolved);
           }


Is that right?


-----------------
Aaron Morton
Freelance Cassandra Developer
@aaronmorton
http://www.thelastpickle.com

On 8 May 2011, at 00:48, Jonathan Ellis wrote:

range_slices respects consistencylevel, but only single-row reads and
multiget do the *repair* part of RR.

On Sat, May 7, 2011 at 1:44 AM, aaron morton <aaron@thelastpickle.com> wrote:
get_range_slices() does read repair if enabled (checked DoConsistencyChecksBoolean in the config, it's on by default) so you should be getting good reads. If you want belt-and-braces run nodetool repair first.

Hope that helps.


On 7 May 2011, at 11:46, Jeremy Hanna wrote:

Great!  I just wanted to make sure you were getting the information you needed.

On May 6, 2011, at 6:42 PM, Henrik Schröder wrote:

Well, I already completed the migration program. Using get_range_slices I could migrate a few thousand rows per second, which means that migrating all of our data would take a few minutes, and we'll end up with pristine datafiles for the new cluster. Problem solved!

I'll see if I can create datafiles in 0.6 that are uncleanable in 0.7 so that you all can repeat this and hopefully fix it.


/Henrik Schröder

On Sat, May 7, 2011 at 00:35, Jeremy Hanna <jeremy.hanna1234@gmail.com> wrote:
If you're able, go into the #cassandra channel on freenode (IRC) and talk to driftx or jbellis or aaron_morton about your problem.  It could be that you don't have to do all of this based on a conversation there.

On May 6, 2011, at 5:04 AM, Henrik Schröder wrote:

I'll see if I can make some example broken files this weekend.


/Henrik Schröder

On Fri, May 6, 2011 at 02:10, aaron morton <aaron@thelastpickle.com> wrote:
The difficulty is the different thrift clients between 0.6 and 0.7.

If you want to roll your own solution I would consider:
- write an app to talk to 0.6 and pull out the data using keys from the other system (so you know can check referential integrity while you are at it). Dump the data to flat file.
- write an app to talk to 0.7 to load the data back in.

I've not given up digging on your migration problem, having to manually dump and reload if you've done nothing wrong is not the best solution. I'll try to find some time this weekend to test with:

- 0.6 server, random paritioner, standard CF's, byte column
- load with python or the cli on osx or ubuntu (dont have a window machine any more)
- migrate and see whats going on.

If you can spare some sample data to load please send it over in the user group or my email address.

Cheers

-----------------
Aaron Morton
Freelance Cassandra Developer
@aaronmorton
http://www.thelastpickle.com

On 6 May 2011, at 05:52, Henrik Schröder wrote:

We can't do a straight upgrade from 0.6.13 to 0.7.5 because we have rows stored that have unicode keys, and Cassandra 0.7.5 thinks those rows in the sstables are corrupt, and it seems impossible to clean it up without losing data.

However, we can still read all rows perfectly via thrift so we are now looking at building a simple tool that will copy all rows from our 0.6.3 cluster to a parallell 0.7.5 cluster. Our question is now how to do that and ensure that we actually get all rows migrated? It's a pretty small cluster, 3 machines, a single keyspace, a singke columnfamily, ~2 million rows, a few GB of data, and a replication factor of 3.

So what's the best way? Call get_range_slices and move through the entire token space? We also have all row keys in a secondary system, would it be better to use that and make calls to get_multi or get_multi_slices instead? Are we correct in assuming that if we use the consistencylevel ALL we'll get all rows?


/Henrik Schröder










--
Jonathan Ellis
Project Chair, Apache Cassandra
co-founder of DataStax, the source for professional Cassandra support
http://www.datastax.com





--
Jonathan Ellis
Project Chair, Apache Cassandra
co-founder of DataStax, the source for professional Cassandra support
http://www.datastax.com