(All page numbers refer to locations in Hadoop, the Definitive Guide, 2nd ed..)
Page 248 discusses how to map out a custom network topology on hadoop using script based mapping. Essentially, we need to write a script that will take a variable number of IP addresses, and return the corresponding network locations. I'm not sure how we can do this effectively, with a pre-written script. Maybe we could use stevedore to generate a script based on all existing nodes in the cluster? Check the "Hadoop Definitive Guide" source code for an example script, here.
The other option would be to implement DNS to Switch mapping:
public interface DNSToSwitchMapping {
public List<String> resolve(List<String> names);
}
- Figure out whether we want to go the script route, or the interface route. (Toni, what's easiest here, based on pallet's capabilities?)
- What IP addresses are we going to receive, public or private? I'm guessing private, but it'd be nice to have an example to reference, here.
Each cluster the user creates should have a specific tag... every node that gets created should be accessed by that tag. I think that pallet can do this now, just not sure how. (samcluster should be different than tonicluster, and our commands on nodes of the same names shouldn't interfere with each other.
Some functions I'd like:
- (hadoop-property :clustertag "mapred.job.tasks") ;=> 10
Page 284. When we change the cluster size in some way, we need to run balancer on, I believe, the namenode.
- (Where do we need to run
balancer?
balancer runs until it's finished, and doesn't get in the way of much. Here's the code for bin/start-balancer.sh
bin=`dirname "$0"`
bin=`cd "$bin"; pwd`
. "$bin"/hadoop-config.sh
# Start balancer daemon.
"$bin"/hadoop-daemon.sh --config $HADOOP_CONF_DIR start balancer $@
It would be nice to have some good interface for distcp, plus a few commands like bin/hadoop fs -getmerge. This isn't so important at all for cascalog & cascading, since various custom and supplied taps take care of everything, here.
Page 251. We can use ssh-agent to get rid of the need to supply a password when logging in.
Turns out that the hostnames in the masters file are used to start up a secondary namenode. Weird!
We should provide the ability to have a class of node, with a number of different images, all sitting under the same class.
The first place we can use this is for clusters of spot and non-spot nodes... then, spot nodes of varying prices. Beyond that, we might want some machines to have different capabilities than others, for whatever reason. (One can imagine a case where a fixed number of nodes are running, backed by EBS, hosting something like ElephantDB in addition to the hadoop namenode and jobtracker processes... the cluster can scale elastically beyond those nodes, but only by non-ebs-backed instances.
See page 286. We might add support for Ganglia, or FileContext. This would require proper modification of the conf/hadoop-metrics.sh.
Page 142... thinking here about customization of conf/log4j.sh.
Page 254. The three masters, in any given cluster, will be the namenode, the jobtracker, and the secondary namenode (optional). They can run on 1-3 machines, in any combination. I don't think we'll ever want more than one of each. And, of course, the startup order's important!
NOTES:
A cluster should take in a map of arguments (ip-type, for example) and a map of node descriptions, including base nodes for each node type, and output a cluster object. We should have a layer of abstraction on top of nodes, etc.
It seems like this is an issue a number of folks are having. We need to populate etc/hosts to skip DNS resolution, if we're going to work on local machines. On EC2, I think we can get around this issue by using the public DNS address.
Some discussion here on a way to short circuit DNS -- http://www.travishegner.com/2009/06/hadoop-020-on-ubuntu-server-904-jaunty.html
But do we want that, really?
Looks like we need to do etc/hosts internally -- we could probably do this externally as well, with Amazon's public DNS names and private IP addresses.
From here: https://twiki.grid.iu.edu/bin/view/Storage/HadoopUnderstanding
For the namenode, etc to be virtualized, you must be able to access them through DNS, or etc/hosts.
From HDFS-default -- http://hadoop.apache.org/common/docs/r0.20.2/hdfs-default.html
dfs.datanode.dns.nameserver -- The host name or IP address of the
name server (DNS) which a DataNode should use to determine the host
name used by the NameNode for communication and display purposes.
More support for using external hostnames on EC2 http://getsatisfaction.com/cloudera/topics/hadoop_configuring_a_slaves_hostname
How to get hadoop running without DNS -- http://db.tmtec.biz/blogs/index.php/get-hadoop-up-and-running-without-dns
Using etc/hosts as default -- http://www.linuxquestions.org/questions/linux-server-73/how-to-setup-nslookups-queries-using-etc-hosts-as-the-default-654882/
And, most clearly:
http://www.cloudera.com/blog/2008/12/securing-a-hadoop-cluster-through-a-gateway/
One “gotcha” of Hadoop is that the HDFS instance has a canonical name associated with it, based on the DNS name of the machine — not its IP address. If you provide an IP address for the fs.default.name, it will reverse-DNS this back to a DNS name, then subsequent connections will perform a forward-DNS lookup on the canonical DNS name
OTHER NOTES
From Noll link: http://www.mail-archive.com/[email protected]/msg00170.html http://search-hadoop.com/m/PcJ6xnNrSo1/Error+reading+task+output+http/v=threaded From a note here: http://www.michael-noll.com/tutorials/running-hadoop-on-ubuntu-linux-multi-node-cluster/#confmasters-master-only
So, we can probably do this with etc/hosts.
Okay, here's the good stuff. We're trying to get a system up and running that can configure a persistent hadoop cluster.
to act as the hadoop user;
sudo su - hadoop
With jclouds 9b, I'm getting all sorts of errors. In config, we need to make sure we're using aws-ec2, not just ec2. Also, cake-pallet adds pallet as a dependency, which forces jclouds beta-8... doesn't work, if we're trying to play in 9b's world.
Either I have to go straight back to 8b, with cake-pallet and no dependencies excluded,
Compile squid from scratch;
./configure --enable-removal-policies="heap,lru"
Then give the guys my configuration file, from my macbook.
TODO -- figure out how to get the proper user permissions, for the squid user!
run squid -z the first time. squid -N runs with no daemon mode
Squid Config Basics Starting Squid Guide
link over to Toni's instructions, on how to test this bad boy.