I am afraid of bugs caused by introducing DNS-server: * if could easily get out of sync with hostname-ip correspondence on the nodes (the host name a node has could be resolved by DNS-server to IP or another node, or even worse: a node could be started with hostname like "ubuntu.com"). * two or more DNS-server (as one is the SPOF) could get out of sync. * DNS-servers could be unreachable due to overload caused by other clients (it happens quite often in dev environment). * DNS-resolving introduces additional delay which could be slow due to UDP-packet loss and retry.
Perhaps, a script maintaining /etc/hosts on all nodes or solutions like Consul.io would address some of these problems. I like your solution with --address command line switch more than HDFS's property "dfs.namenode.datanode.registration.ip-hostname-check", as it allows to use nodes with more than 1 IP. On 9/6/15, Josh Elser <josh.el...@gmail.com> wrote: > FYI, there is reasoning behind doing this in respect to the larger > Hadoop community. When you run with Kerberos, things will not work at > all without proper hostnames. There are lots of cases where dealing with > hostnames is easier than IPs (e.g. what happens if the physical host you > run the namenode on dies and you need to transition it to a different > node -- IPs make this much more difficult). > > That said, if there is something we can do that will make this easier > for your situation, let's give it some thought. I don't think I realized > that things were actually broken with IPs, just that we strongly > recommend using hostnames instead. > > Denis wrote: >> I found the answer in the sources: start the daemons with "--address >> 10.x.x.x" command line switch. >> >> I am trying to install Accumulo 1.7 (I have experience with early >> versions) and found if inconvenient the Accumulo now requires >> hostnames; >> I do not like the idea to maintain the configuration of a DNS-server >> (or hostname-IP correspondence in /etc/hosts of all the servers on all >> the servers) >> >> On 9/5/15, Josh Elser<josh.el...@gmail.com> wrote: >>> I'm not entirely sure if I understand your question. Are you running >>> Accumulo 1.5 and having problems running Accumulo w/ only IP addresses? >>> >>> It's not readily apparent to me what changes were actually made in the >>> tickets as it appears JIRA didn't pick up the commit reference. >>> >>> For as long as I can remember, the network interface (e.g. 127.0.0.1, >>> localhost, `hostname -a`, `hostname -f`) you provide in >>> ${ACCUMULO_CONF_DIR}/{masters,monitors,slaves,gc,tracers} should >>> ultimately determine what is advertised in ZooKeeper. Maybe that's the >>> change Eric made... >>> >>> Denis wrote: >>>> Hello >>>> >>>> What would be the recommended way to win the pre-ACCUMULO-1585 >>>> behavior, i.e IPs of services in ZooKeeper and no need for DNS not >>>> reverse DNS? >