hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Yu <yuzhih...@gmail.com>
Subject Re: [DISCUSSION] MR jobs started by Master or RS
Date Fri, 23 Sep 2016 12:39:48 GMT
I suggest you look at Matteo's work for AssignmentManager which is to make
Master more stable.

Cheers

On Fri, Sep 23, 2016 at 5:32 AM, 张铎 <palomino219@gmail.com> wrote:

> No, not your fault, at lease, not this time:)
>
> Why I call the code ugly? Can you simply tell me the sequence of calls when
> starting up the HMaster? HMaster is also a regionserver so it extends
> HRegionServer, and the initialization of HRegionServer sometimes needs to
> make rpc calls to HMaster. A simple change would cause probabilistic dead
> lock or some strange NPEs...
>
> That's why I'm very nervous when somebody wants to add new features or add
> external dependencies to HMaster, especially add more works for the start
> up processing...
>
> Thanks.
>
> 2016-09-23 20:02 GMT+08:00 Ted Yu <yuzhihong@gmail.com>:
>
> > I read through HADOOP-13433
> > <https://issues.apache.org/jira/browse/HADOOP-13433> - the cited race
> > condition is in jdk.
> >
> > Suggest pinging the reviewer on JIRA to get it moving.
> >
> > bq. But the ugly code in HMaster is readlly a problem...
> >
> > Can you be specific as to which code is ugly ? Is it in the backup /
> > restore mega patch ?
> >
> > Cheers
> >
> > On Thu, Sep 22, 2016 at 10:44 PM, 张铎 <palomino219@gmail.com> wrote:
> >
> > > If you guys have already implemented the feature in the MR way and the
> > > patch is ready for landing on master, I'm a -0 on it as I do not want
> to
> > > block the development progress.
> > >
> > > But I strongly suggest later we need to revisit the design and see if
> we
> > > can seperated the logic from HMaster as much as possible. HA is not a
> big
> > > problem if you do not store any metada locally. But the ugly code in
> > > HMaster is readlly a problem...
> > >
> > > And for security, I have a issue pending for a long time. Can someone
> > help
> > > taking a simple look at it? This is what I mean, ugly code... logout
> and
> > > destroy the credentials in a subject when it is still being used, and
> > > declared as LimitPrivacy so I can not change the behivor and the only
> way
> > > to fix it is to write another piece of ugly code...
> > >
> > > https://issues.apache.org/jira/browse/HADOOP-13433
> > >
> > > 2016-09-23 12:53 GMT+08:00 Vladimir Rodionov <vladrodionov@gmail.com>:
> > >
> > > > >> If in the future, we find better ways of doing this without using
> > MR,
> > > we
> > > > can certainly consider that
> > > >
> > > > Our framework for distributed operations is abstract and allows
> > > > different implementations. MR is just one implementation we provide.
> > > >
> > > > -Vlad
> > > >
> > > > On Thu, Sep 22, 2016 at 9:38 PM, Devaraj Das <ddas@hortonworks.com>
> > > wrote:
> > > >
> > > > > Guys, first off apologies for bringing in the topic of MR-based
> > > > > compactions.. But I was thinking more about the SpliceMachine
> > approach
> > > of
> > > > > managing compactions in Spark where apparently they saw a lot of
> > > > benefits.
> > > > > Apologies for giving you that sore throat Andrew; I really didn't
> > mean
> > > to
> > > > > :-)
> > > > >
> > > > > So on this issue, we have these on the plate:
> > > > > 0. Somehow not use MR but something like that
> > > > > 1. Run a standalone service other than master
> > > > > 2. Shell out from the master
> > > > >
> > > > > I don't think we have a good answer to (0), and I don't think it's
> > even
> > > > > worth the effort of trying to build something when MR is already
> > there,
> > > > and
> > > > > being used by HBase already for some operations.
> > > > >
> > > > > On (1), we have to deal with a myriad of issues - HA of the server
> > not
> > > > > being the least of them all. Security (kerberos authentication,
> > another
> > > > > keytab to manage, etc. etc. etc.). IMO, that approach is DOA.
> Instead
> > > > let's
> > > > > substitute that (1) with the HBase Master. I haven't seen any good
> > > reason
> > > > > why the HBase master shouldn't launch MR jobs if needed. It's not
> > > ideal;
> > > > > agreed.
> > > > >
> > > > > Now before going to (2), let's see what are the benefits of running
> > the
> > > > > backup/restore jobs from the master. I think Ted has summarized
> some
> > of
> > > > the
> > > > > issues that we need to take care of - basically, the master can
> keep
> > > > track
> > > > > of running jobs, and should it fail, the backup master can continue
> > > > keeping
> > > > > track of it (since the jobId would have been recorded in the proc
> > WAL).
> > > > The
> > > > > master can also do cleanup, etc. of failed backup/restore
> processes.
> > > > > Security is another issue - the job needs to run as 'hbase' since
> it
> > > owns
> > > > > the data. Having the master launch the job makes it get that
> > privilege.
> > > > In
> > > > > the (2) approach, it's hard to do some of the above management.
> > > > >
> > > > > Guys, just to reiterate, the patch as such is ready from the
> overall
> > > > > design/arch point of view (maybe code review is still pending from
> > > > Matteo).
> > > > > If in the future, we find better ways of doing this without using
> MR,
> > > we
> > > > > can certainly consider that. But IMO don't think we should block
> this
> > > > patch
> > > > > from getting merged.
> > > > >
> > > > > ________________________________________
> > > > > From: 张铎 <palomino219@gmail.com>
> > > > > Sent: Thursday, September 22, 2016 8:32 PM
> > > > > To: dev@hbase.apache.org
> > > > > Subject: Re: [DISCUSSION] MR jobs started by Master or RS
> > > > >
> > > > > So what about a standalone service other than master? You can use
> > your
> > > > own
> > > > > procedure store in that service?
> > > > >
> > > > > 2016-09-23 11:28 GMT+08:00 Ted Yu <yuzhihong@gmail.com>:
> > > > >
> > > > > > An earlier implementation was client driven.
> > > > > >
> > > > > > But with that approach, it is hard to resume if there is error
> > > midway.
> > > > > > Using Procedure V2 makes the backup / restore more robust.
> > > > > >
> > > > > > Another consideration is for security. It is hard to enforce
> > security
> > > > (to
> > > > > > be implemented) for client driven actions.
> > > > > >
> > > > > > Cheers
> > > > > >
> > > > > > > On Sep 22, 2016, at 8:15 PM, Andrew Purtell <
> > > > andrew.purtell@gmail.com>
> > > > > > wrote:
> > > > > > >
> > > > > > > No, this misses Matteo's finer point, which is "shelling
out"
> > from
> > > > the
> > > > > > master directly to run MR is a first. Why not drive this with
a
> > > utility
> > > > > > derived from Tool?
> > > > > > >
> > > > > > > On Sep 22, 2016, at 7:57 PM, Vladimir Rodionov <
> > > > vladrodionov@gmail.com
> > > > > >
> > > > > > wrote:
> > > > > > >
> > > > > > >>>> In our production cluster,  it is a common
case we just have
> > > HDFS
> > > > > and
> > > > > > >>>> HBase deployed.
> > > > > > >>>> If our Master/RS depend on MR framework (especially
some
> > > features
> > > > we
> > > > > > >>>> have not used at all),  it introduced another
cost for
> > maintain.
> > > > I
> > > > > > >>>> don't think it is a good idea.
> > > > > > >>
> > > > > > >> So , you are not backup users in this case. Many our
customers
> > > have
> > > > > full
> > > > > > >> stack deployed and
> > > > > > >> want see backup to be a standard feature. Besides this,
> nothing
> > > will
> > > > > > happen
> > > > > > >> in your cluster
> > > > > > >> if you won't be doing backups.
> > > > > > >>
> > > > > > >> This discussion (we do not want see M/R dependency)
goes to
> > > nowhere.
> > > > > We
> > > > > > >> asked already, at least twice, to suggest another framework
> > (other
> > > > > than
> > > > > > M/R)
> > > > > > >> for bulk data copy with *conversion*. Still waiting
for
> > > suggestions.
> > > > > > >>
> > > > > > >> -Vlad
> > > > > > >>
> > > > > > >>
> > > > > > >>
> > > > > > >>
> > > > > > >>> On Thu, Sep 22, 2016 at 7:49 PM, Ted Yu <yuzhihong@gmail.com
> >
> > > > wrote:
> > > > > > >>>
> > > > > > >>> If MR framework is not deployed in the cluster,
hbase still
> > > > functions
> > > > > > >>> normally (post merge).
> > > > > > >>>
> > > > > > >>> In terms of build time dependency, we have long
been
> depending
> > on
> > > > > > >>> mapreduce. Take a look at ExportSnapshot.
> > > > > > >>>
> > > > > > >>> Cheers
> > > > > > >>>
> > > > > > >>> On Thu, Sep 22, 2016 at 7:42 PM, Heng Chen <
> > > > heng.chen.1986@gmail.com
> > > > > >
> > > > > > >>> wrote:
> > > > > > >>>
> > > > > > >>>> In our production cluster,  it is a common
case we just have
> > > HDFS
> > > > > and
> > > > > > >>>> HBase deployed.
> > > > > > >>>> If our Master/RS depend on MR framework (especially
some
> > > features
> > > > we
> > > > > > >>>> have not used at all),  it introduced another
cost for
> > maintain.
> > > > I
> > > > > > >>>> don't think it is a good idea.
> > > > > > >>>>
> > > > > > >>>> 2016-09-23 10:28 GMT+08:00 张铎 <palomino219@gmail.com>:
> > > > > > >>>>> To be specific, for example, our nice Backup/Restore
> feature,
> > > if
> > > > we
> > > > > > >>> think
> > > > > > >>>>> this is not a core feature of HBase, then
we could make it
> > > depend
> > > > > on
> > > > > > >>> MR,
> > > > > > >>>>> and start a standalone BackupManager instance
that submits
> MR
> > > > jobs
> > > > > to
> > > > > > >>> do
> > > > > > >>>>> periodical maintenance job. And if we think
this is a core
> > > > feature
> > > > > > that
> > > > > > >>>>> everyone should use it, then we'd better
implement it
> without
> > > MR
> > > > > > >>>>> dependency, like DLS.
> > > > > > >>>>>
> > > > > > >>>>> Thanks.
> > > > > > >>>>>
> > > > > > >>>>> 2016-09-23 10:11 GMT+08:00 张铎 <palomino219@gmail.com>:
> > > > > > >>>>>
> > > > > > >>>>>> I‘m -1 on let master or rs launch
MR jobs. It is OK that
> > some
> > > of
> > > > > our
> > > > > > >>>>>> features depend on MR but I think the
bottom line is that
> we
> > > > > should
> > > > > > >>>> launch
> > > > > > >>>>>> the jobs from outside manually or by
other services.
> > > > > > >>>>>>
> > > > > > >>>>>> 2016-09-23 9:47 GMT+08:00 Andrew Purtell
<
> > > > > andrew.purtell@gmail.com
> > > > > > >:
> > > > > > >>>>>>
> > > > > > >>>>>>> Ok, got it. Well "shelling out"
is on the line I think,
> so
> > a
> > > > fair
> > > > > > >>>>>>> question.
> > > > > > >>>>>>>
> > > > > > >>>>>>> Can this be driven by a utility
derived from Tool like
> our
> > > > other
> > > > > MR
> > > > > > >>>> apps?
> > > > > > >>>>>>> The issue is needing the AccessController
to decide if
> > > allowed?
> > > > > But
> > > > > > >>>> nothing
> > > > > > >>>>>>> prevents the user from running
the job
> > > manually/independently,
> > > > > > right?
> > > > > > >>>>>>>
> > > > > > >>>>>>>> On Sep 22, 2016, at 3:44 PM,
Matteo Bertozzi <
> > > > > > >>>> theo.bertozzi@gmail.com>
> > > > > > >>>>>>> wrote:
> > > > > > >>>>>>>>
> > > > > > >>>>>>>> just a remark. my query was
not about tools using MR
> > > > (everyone i
> > > > > > >>>> think
> > > > > > >>>>>>> is
> > > > > > >>>>>>>> ok with those).
> > > > > > >>>>>>>> the topic was about: "are we
ok with running MR jobs
> from
> > > > Master
> > > > > > >>> and
> > > > > > >>>> RSs
> > > > > > >>>>>>>> code?" since this will be the
first time we do this
> > > > > > >>>>>>>>
> > > > > > >>>>>>>> Matteo
> > > > > > >>>>>>>>
> > > > > > >>>>>>>>
> > > > > > >>>>>>>>> On Thu, Sep 22, 2016 at
2:49 PM, Devaraj Das <
> > > > > > >>> ddas@hortonworks.com>
> > > > > > >>>>>>> wrote:
> > > > > > >>>>>>>>>
> > > > > > >>>>>>>>> Very much agree; for tools
like ExportSnapshot /
> Backup /
> > > > > > Restore,
> > > > > > >>>> it's
> > > > > > >>>>>>>>> fine to be dependent on
MR. MR is the right framework
> for
> > > > such.
> > > > > > We
> > > > > > >>>>>>> should
> > > > > > >>>>>>>>> also do compactions using
MR (just saying :) )
> > > > > > >>>>>>>>> ________________________________________
> > > > > > >>>>>>>>> From: Ted Yu <yuzhihong@gmail.com>
> > > > > > >>>>>>>>> Sent: Thursday, September
22, 2016 2:00 PM
> > > > > > >>>>>>>>> To: dev@hbase.apache.org
> > > > > > >>>>>>>>> Subject: Re: [DISCUSSION]
MR jobs started by Master or
> RS
> > > > > > >>>>>>>>>
> > > > > > >>>>>>>>> I agree - backup / restore
is in the same category as
> > > import
> > > > /
> > > > > > >>>> export.
> > > > > > >>>>>>>>>
> > > > > > >>>>>>>>> On Thu, Sep 22, 2016 at
1:58 PM, Andrew Purtell <
> > > > > > >>>>>>> andrew.purtell@gmail.com>
> > > > > > >>>>>>>>> wrote:
> > > > > > >>>>>>>>>
> > > > > > >>>>>>>>>> Backup is extra tooling
around core in my opinion.
> Like
> > > > import
> > > > > > or
> > > > > > >>>>>>> export.
> > > > > > >>>>>>>>>> Or the optional MOB
tool. It's fine.
> > > > > > >>>>>>>>>>
> > > > > > >>>>>>>>>>> On Sep 22, 2016,
at 1:50 PM, Matteo Bertozzi <
> > > > > > >>>> mbertozzi@apache.org>
> > > > > > >>>>>>>>>> wrote:
> > > > > > >>>>>>>>>>>
> > > > > > >>>>>>>>>>> What's the latest
opinion around running MR jobs from
> > > hbase
> > > > > > >>>> (Master
> > > > > > >>>>>>> or
> > > > > > >>>>>>>>>> RS)?
> > > > > > >>>>>>>>>>>
> > > > > > >>>>>>>>>>> I remember in the
past that there was discussion
> about
> > > not
> > > > > > >>> having
> > > > > > >>>> MR
> > > > > > >>>>>>>>> has
> > > > > > >>>>>>>>>>> direct dependency
of hbase.
> > > > > > >>>>>>>>>>>
> > > > > > >>>>>>>>>>> I think some of
discussion where around MOB that had
> a
> > MR
> > > > job
> > > > > > to
> > > > > > >>>>>>>>> compact,
> > > > > > >>>>>>>>>>> that later was
transformed in a non-MR job to be
> > merged,
> > > I
> > > > > > think
> > > > > > >>>> we
> > > > > > >>>>>>>>> had a
> > > > > > >>>>>>>>>>> similar discussion
for log split/replay.
> > > > > > >>>>>>>>>>>
> > > > > > >>>>>>>>>>> the latest is the
new Backup feature (HBASE-7912),
> that
> > > > runs
> > > > > a
> > > > > > >>> MR
> > > > > > >>>> job
> > > > > > >>>>>>>>>> from
> > > > > > >>>>>>>>>>> the master to copy
data or restore data.
> > > > > > >>>>>>>>>>> (backup is also
"not really core" as in.. if you
> don't
> > > use
> > > > > > >>> backup
> > > > > > >>>>>>>>> you'll
> > > > > > >>>>>>>>>>> not end up running
MR jobs, but this was probably
> true
> > > for
> > > > > MOB
> > > > > > >>> as
> > > > > > >>>> in
> > > > > > >>>>>>>>> "if
> > > > > > >>>>>>>>>>> you don't enable
MOB you don't need MR")
> > > > > > >>>>>>>>>>>
> > > > > > >>>>>>>>>>> any thoughts? do
we a rule that says "we don't want
> to
> > > have
> > > > > > >>> hbase
> > > > > > >>>> run
> > > > > > >>>>>>>>> MR
> > > > > > >>>>>>>>>>> jobs, only tool
started manually by the user can do
> > > that".
> > > > or
> > > > > > >>> can
> > > > > > >>>> we
> > > > > > >>>>>>>>>> start
> > > > > > >>>>>>>>>>> adding MR calls
around without problems?
> > > > > > >>>
> > > > > >
> > > > >
> > > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message