It would help if you could narrow it down to "here are the keys I
expect to see that I am not," especially if you can reproduce on a
single-node cluster.
On Sat, Feb 6, 2010 at 2:04 AM, Jack Culpepper <jackculpep...@gmail.com> wrote:
> Hi Jonathon,
>
> I am seeing a dramatic difference in the number of keys I can scan
> when I use these two methods.
>
> The former (deprecated) method seems to return the correct result.
> That is, it's on the right order of magnitude of around 500K, and if I
> continue to insert keys via a separate process as I repeatedly count
> them, the count grows. The recommended alternative, get_range_slice(),
> returns far fewer keys and if I count repeatedly as I insert using a
> separate process, the count bounces around erratically.
>
> I am using the python thrift interface against a two node setup. I am
> running the current 0.5.0 release (just upgraded from rc1 since I saw
> some other thrift bug was fixed). Here is my program (there are three
> commented lines to switch from one method to the other):
>
> if sys.argv[1] == "count_things":
>
> from thrift import Thrift
> from thrift.transport import TTransport
> from thrift.transport import TSocket
> from thrift.protocol.TBinaryProtocol import TBinaryProtocolAccelerated
> from cassandra import Cassandra
>
> socket = TSocket.TSocket("10.212.230.176", 9160)
> transport = TTransport.TBufferedTransport(socket)
> protocol = TBinaryProtocol.TBinaryProtocolAccelerated(transport)
> client = Cassandra.Client(protocol)
>
> transport.open()
>
> column_parent = ColumnParent(column_family="thing")
> slice_range = SliceRange(start="key", finish="key")
> predicate = SlicePredicate(slice_range=slice_range)
>
> done = False
> seg = 1000
> start = ""
>
> while not done:
> #result = client.get_key_range("gg", "thing", start, "", seg,
> ConsistencyLevel.ONE)
> result = client.get_range_slice("gg", column_parent,
> predicate, start, "", seg, ConsistencyLevel.ONE)
>
> if len(result) < seg: done = True
> #else: start = result[seg-1]
> else: start = result[seg-1].key
>
> record_count += len(result)
>
> t = now()
> dt = t - startTime
> record_per_sec = record_count / dt
> #print "\rstart %d now %d dt %d rec/s %.4f rec %d s %s f
> %s"%(startTime,t,dt,record_per_sec,record_count,result[0],result[-1]),
> print "\rstart %d now %d dt %d rec/s %.4f rec %d s %s f
> %s"%(startTime,t,dt,record_per_sec,record_count,result[0].key,result[-1].key),
> print
>
> An example of the output using get_range_slice(), without a concurrent
> insertion process -- it counts 133674 keys.
>
> start 1265440888 now 1265441098 dt 210 rec/s 636.1996 rec 133674 s
> 9f9dd2c0f043902f7f571942cfac3f6c28b82cec f
> 9ffff14fd361b981faea6a04c5ef5699a96a8d6d
>
> Using get_key_range() I get 459351 keys, and the throughput is less:
>
> start 1265442143 now 1265443092 dt 948 rec/s 484.2775 rec 459351 s
> ffce8099f808d10a09db471b04793315f555ccbd f
> ffffffa1b5e3aeb9ca92d4d848280093bdf49892
>
> get_range_slice() seems to skip keys in each of the segments.
>
> The "thing" column family is a super column. There are no errors
> reported to the log. The keys I am inserting are python generated
> UUIDs:
>
> import uuid
> key = uuid.uuid4().hex
>
> I'm not posting the program that inserts the data, but I can if that
> would be help. Thanks very much,
>
> Jack
>
