Thanks for the info Mike, we ran in to a race condition which was killing table snap, I want to share the problem and the solution/ work around and may be someone can throw some light on the effects of the solution.

tablesnap was getting killed with this error message:

Failed uploading %s. Aborting.\n%s

Looking at the code it took me to the following:

def worker(self):
        bucket = self.get_bucket()

        while True:
            f = self.fileq.get()
            keyname = self.build_keyname(f)
                self.upload_sstable(bucket, keyname, f)
                self.log.critical("Failed uploading %s. Aborting.\n%s" %
                             (f, format_exc()))
                # Brute force kill self
                os.kill(os.getpid(), signal.SIGKILL)


It builds the filename and then before it could upload it, the file disappears (which is possible), I simply commented out the line which kills tablesnap if the file is not found, it fixes the issue we were having but I would appreciate if some one has any insights on any ill effects this might have on backup or restoration process.


On Jul 11, 2013, at 7:03 AM, Mike Heffner <> wrote:

We've also noticed very good read and write latencies with the hi1.4xls compared to our previous instance classes. We actually ran a mixed cluster of hi1.4xls and m2.4xls to watch side-by-side comparison.

Despite the significant improvement in underlying hardware, we've noticed that streaming performance with 1.2.6+vnodes is a lot slower than we would expect. Bootstrapping a node into a ring with large storage loads can take 6+ hours. We have a JIRA open that describes our current config:

Aiman: We also use tablesnap for our backups. We're using a slightly modified version [1]. We currently backup every sst as soon as they hit disk (tablesnap's inotify), but we're considering moving to a periodic snapshot approach as the sst churn after going from 24 nodes -> 6 nodes is quite high.



On Thu, Jul 11, 2013 at 7:33 AM, Aiman Parvaiz <> wrote:
We also recently migrated to 3 hi.4xlarge boxes(Raid0 SSD) and the disk IO performance is definitely better than the earlier non SSD servers, we are serving up to 14k reads/s with a latency of 3-3.5 ms/op.
I wanted to share our config options and ask about the data back up strategy for Raid0.

We are using C* 1.2.6 with

key_chache and row_cache of 300MB
I have not changed/ modified any other parameter except for going with multithreaded GC. I will be playing around with other factors and update everyone if I find something interesting.

Also, just wanted to share backup strategy and see if I can get something useful from how others are taking backup of their raid0. I am using tablesnap to upload SSTables to s3 and I have attached a separate EBS volume to every box and have set up rsync to mirror Cassandra data from Raid0 to EBS. I would really appreciate if you guys can share how you taking backups.


On Jul 9, 2013, at 7:11 AM, Alain RODRIGUEZ <> wrote:

> Hi,
> Using C*1.2.2.
> We recently dropped our 18 m1.xLarge (4CPU, 15GB RAM, 4 Raid-0 Disks) servers to get 3 hi1.4xLarge (16CPU, 60GB RAM, 2 Raid-0 SSD) servers instead, for about the same price.
> We tried it after reading some benchmark published by Netflix.
> It is awesome and I recommend it to anyone who is using more than 18 xLarge server or can afford these high cost / high performance EC2 instances. SSD gives a very good throughput with an awesome latency.
> Yet, we had about 200 GB data per server and now about 1 TB.
> To alleviate memory pressure inside the heap I had to reduce the index sampling. I changed the index_interval value from 128 to 512, with no visible impact on latency, but a great improvement inside the heap which doesn't complain about any pressure anymore.
> Is there some more tuning I could use, more tricks that could be useful while using big servers, with a lot of data per node and relatively high throughput ?
> SSD are at 20-40 % of their throughput capacity (according to OpsCenter), CPU almost never reach a bigger load than 5 or 6 (with 16 CPU), 15 GB RAM used out of 60GB.
> At this point I have kept my previous configuration, which is almost the default one from the Datastax community AMI. There is a part of it, you can consider that any property that is not in here is configured as default :
> cassandra.yaml
> key_cache_size_in_mb: (empty) - so default - 100MB (hit rate between 88 % and 92 %, good enough ?)
> row_cache_size_in_mb: 0 (not usable in our use case, a lot of different and random reads)
> flush_largest_memtables_at: 0.80
> reduce_cache_sizes_at: 0.90
> concurrent_reads: 32 (I am thinking to increase this to 64 or more since I have just a few servers to handle more concurrence)
> concurrent_writes: 32 (I am thinking to increase this to 64 or more too)
> memtable_total_space_in_mb: 1024 (to avoid having a full heap, shoul I use bigger value, why for ?)
> rpc_server_type: sync (I tried hsha and had the "ERROR 12:02:18,971 Read an invalid frame size of 0. Are you using TFramedTransport on the client side?" error). No idea how to fix this, and I use 5 different clients for different purpose  (Hector, Cassie, phpCassa, Astyanax, Helenus)...
> multithreaded_compaction: false (Should I try enabling this since I now use SSD ?)
> compaction_throughput_mb_per_sec: 16 (I will definitely up this to 32 or even more)
> cross_node_timeout: true
> endpoint_snitch: Ec2MultiRegionSnitch
> index_interval: 512
> I am not sure about how to tune the heap, so I mainly use defaults
> HEAP_NEWSIZE="400M" (I tried with higher values, and it produced bigger GC times (1600 ms instead of < 200 ms now with 400M)
> -XX:+UseParNewGC
> -XX:+UseConcMarkSweepGC
> -XX:+CMSParallelRemarkEnabled
> -XX:SurvivorRatio=8
> -XX:MaxTenuringThreshold=1
> -XX:CMSInitiatingOccupancyFraction=70
> -XX:+UseCMSInitiatingOccupancyOnly
> Does this configuration seems coherent ? Right now, performance are correct, latency < 5ms almost all the time. What can I do to handle more data per node and keep these performances or get even better once ?
> I know this is a long message but if you have any comment or insight even on part of it, don't hesitate to share it. I guess this kind of comment on configuration is usable by the entire community.
> Alain


  Mike Heffner <>
  Librato, Inc.