Good questions lead to good books :)
From the outside, I would expect:
1.- Region tries first to contact Master. If it does not succeed, it
simply does not try to split... But from your answer, it seems that the
split is started even if master is not there, leading to 'phantom
regions'. Btw, any 'alert' system available?
2.- Region sends regular heartbeats to master + client informs master in
case of region failure. Same as previous question: any 'alert' system
available?
I will try to take time to look into the code for the 2 answers, looking
at the region state transitions. This is a good opportunity to go inside.
Thx,
Eric
On 12/02/12 18:57, Doug Meil wrote:
Hi there-
Glad to hear that the Arch chapter was helpful!
re: #1 "splits"
The split won't be complete if the master isn't there. Good question, and
the effects of the master not being up will be added to the book.
re: #2. "RS monitoring"
Good question, I'm not familiar with the specifics on that. (and more
info will be added to the book too.)
On 2/12/12 4:00 AM, "Eric Charles"<[email protected]> wrote:
Hi Doug,
I was well thinking that master was in the play in case of region split
or problems.
I searched in the paper book I have, but didn't really find anything on
that.
Luckily, it is well documented on
http://hbase.apache.org/book/regions.arch.html:
- failover: "... The Master will detect that the RegionServer has failed.
"
- split: "...and then reports the split to the Master. "
So master seems also to be involved in a regular splitting in addition
to RS lose. More questions:
1.- What occurs if the region server can not report to Master after
splitting (for example if a guy like me killed the master to see 'what
happens'...).
2.- How does the master achieve the monitoring of its regions? (is there
an hearbeat generated by the regions or does Master pol).
(searched around in books... but didn't find, maybe something to do with
CatalogJanitor?)
Thx,
Eric
On 11/02/12 20:12, Doug Meil wrote:
Regarding the master being down, just be aware that if you lose an RS
that
you'll have issues because the master is what does the reassignment.
Per
the previous comments, at steady-state HBase can run without the master
-
there's an asterisk.
On 2/11/12 11:31 AM, "Eric Charles"<[email protected]> wrote:
Funky, cause HBase is often presented as a 3-layers server model
zookeeper/master/region (root/meta in the regions bringing still more
fogs).
Maybe the commonly understanding in the community is that HBase's
HMaster is like Hadoop's NameNode, which is not the case (NameNode's
failure will bring your cluster down, HMaster's failure will not...).
Upon presentation simplification (via 'master' renaming, or at least
via
explanation emphases), having a more simple model such as the one
described in HBASE-3171 could help (this is my very short opinion, I
remember having read valid arguments against that model).
client<->zk<->(master)<->region(root|master)
or
client<->zk<->region (+admin)
the latter looks simpler, doesn't it?
Thx,
Eric
On 11/02/12 16:45, Stack wrote:
2012/2/11 Eric Charles<[email protected]>:
Hypothetical case (probably asked a number of time here, sorry...):
Can a client correcty put, get, scan (no admins tasks such as create
table,...) with a HBase cluster having its HMaster process down ?
It can Eric. Kill your master and you can do all the above.
St.Ack
--
eric | http://about.echarles.net | @echarles
--
eric | http://about.echarles.net | @echarles
--
eric | http://about.echarles.net | @echarles
