[
https://issues.apache.org/jira/browse/HADOOP-2185?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12542640
]
Allen Wittenauer commented on HADOOP-2185:
------------------------------------------
The primary issue is that hadoop isn't predictable. The randomization of ports
(including picking purely ephemeral ports) is that it is impossible to:
1) Make sure Hadoop doesn't sit on a port that another application may be using
unless that application is already running.
2) Create firewall rules that prevents connection to Hadoop services.
3) Create QoS (Quality of Service) settings such that HDFS has a higher/lower
priority vs. some other service.
I really want predictability. I want to be able to say that my JobTracker is
always using port a,b,c, my namenode is always using ports d,e,f, my datanode
is always using port g,h,i, etc. If that port is in use, then it should be
perfectly accepting to have that process fail. Predictability == management
when we're talking about large scale administration.
As Arun pointed out, if someone really doesn't care what ports these services
run on, then using 0 should be a reliable equivalent to using port rolling.
As to the names, I'd prefer 'hdfs' over 'dfs' if only because a lot of the
people I talk to always follow up with "Why are you using Microsoft's DFS?".
*sigh* I prefer 'http' over 'info' if only because most people are more
likely to recognize that a web interface is sitting on that port and that it
might require extra care. [The (ab)use of that port by Hadoop is another
issue... ;) ]
One concern I have is what happens if I have multiple interfaces (nics). How
does it work if I only want to bind to one interface or if I want to bind to
all of them or I want to bind to different ports on those different interfaces?
> Server ports: to roll or not to roll.
> -------------------------------------
>
> Key: HADOOP-2185
> URL: https://issues.apache.org/jira/browse/HADOOP-2185
> Project: Hadoop
> Issue Type: Improvement
> Components: conf, dfs, mapred
> Affects Versions: 0.15.0
> Reporter: Konstantin Shvachko
> Fix For: 0.16.0
>
>
> Looked at the issues related to port rolling. My impression is that port
> rolling is required only for the unit tests to run.
> Even the name-node port should roll there, which we don't have now, in order
> to be able to start 2 cluster for testing say dist cp.
> For real clusters on the contrary port rolling is not desired and some times
> even prohibited.
> So we should have a way of to ban port rolling. My proposition is to
> # use ephemeral port 0 if port rolling is desired
> # if a specific port is specified then port rolling should not happen at all,
> meaning that a
> server is either able or not able to start on that particular port.
> The desired port is specified via configuration parameters.
> - Name-node: fs.default.name = host:port
> - Data-node: dfs.datanode.port
> - Job-tracker: mapred.job.tracker = host:port
> - Task-tracker: mapred.task.tracker.report.bindAddress = host
> Task-tracker currently does not have an option to specify port, it always
> uses the ephemeral port 0,
> and therefore I propose to add one.
> - Secondary node does not need a port to listen on.
> For info servers we have two sets of config variables *.info.bindAddress and
> *.info.port
> except for the task tracker, which calls them *.http.bindAddress and
> *.http.port instead of "info".
> With respect to the info servers I propose to completely eliminate the port
> parameters, and form
> *.info.bindAddress = host:port
> Info servers should do the same thing, namely start or fail on the specified
> port if it is not 0,
> and start on any free port if it is ephemeral.
> For the task-tracker I would rename tasktracker.http.bindAddress to
> mapred.task.tracker.info.bindAddress
> For the data-node the info dfs.datanode.info.bindAddress should be included
> into the default config.
> Is there a reason why it is not there?
> This is the summary of proposed changes:
> || Server || current name = value || proposed name = value ||
> | NameNode | fs.default.name = host:port | same |
> | | dfs.info.bindAddress = host | dfs.info.bindAddress = host:port |
> | DataNode | dfs.datanode.port = port | same |
> | | dfs.datanode.info.bindAddress = host | dfs.datanode.info.bindAddress =
> host:port |
> | | dfs.datanode.info.port = port | eliminate |
> | JobTracker| mapred.job.tracker = host:port | same |
> | | mapred.task.tracker.info.bindAddress = host |
> mapred.task.tracker.info.bindAddress = host:port |
> | | mapred.task.tracker.info.port = port | eliminate |
> | TaskTracker| mapred.task.tracker.report.bindAddress = host |
> mapred.task.tracker.report.bindAddress = host:port |
> | | tasktracker.http.bindAddress = host |
> mapred.task.tracker.info.bindAddress = host:port |
> | | tasktracker.http.port = port | eliminate |
> | SecondaryNameNode | dfs.secondary.info.bindAddress = host |
> dfs.secondary.info.bindAddress = host:port |
> | | dfs.secondary.info.port = port | eliminate |
> Do we also want to set some uniform naming convention for the configuration
> variables?
> Like having hdfs instead of dfs, or info instead of http, or systematically
> using either datanode
> or data.node would make that look better in my opinion.
> So these are all +*api*+ changes. I would +*really*+ like some feedback on
> this, especially from
> people who deal with configuration issues on practice.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
