hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Eric Owhadi <eric.owh...@esgyn.com>
Subject RE: Question on hbase.client.scanner.timeout.period
Date Thu, 10 Sep 2015 17:21:04 GMT
Thanks for pointing me to HBase-13333, it is indeed supposed to address the
very same problem. With the drawback of requiring client side involvement,
of asynchronous nature. I still have not discovered any reason why just
doing it the way I proposed would lead to any negative side effect. Must
admit I feel uncomfortable since the patch is just about removing code that
usually is added with a purpose :-).
We have not yet run full QA, but at least 100% of trafodion regression test
pass.
As for when the patch will make it to trafodion, given that I did it only
for a CDH build of Trafodion with HBase 1.0 support, I still cannot check it
in (trafodion is still on .98 and builds OK for Cloudera,Hortonworks,Mapr
and Apache). Trafodion would first need to have full support for HBase 1.0
for all Hadoop distro we support, then I will need to redo the patch that is
distro specific, and make sure the build process deals with this... It is my
plan to do so... Hoping that I do not discover any issue with other distro
(like private attributes or functions that I cannot circumvent... but that
would just mean that the patch would not be available for a specific distro)
Eric


-----Original Message-----
From: Jerry He [mailto:jerryjch@gmail.com]
Sent: Saturday, September 5, 2015 1:47 PM
To: dev <dev@hbase.apache.org>
Subject: Re: Question on hbase.client.scanner.timeout.period

You can take a look at HBASE-13333: Renew Scanner Lease without advancing
the RegionScanner, which may be helpful in this kind of case  Your proposal
sounds like a good alternative approach as well.
We should add that JIRA to the blog link Stack mentioned.

Jerry

On Sat, Sep 5, 2015 at 9:07 AM, Stack <stack@duboce.net> wrote:

> On Fri, Sep 4, 2015 at 5:06 PM, Eric Owhadi <eric.owhadi@esgyn.com> wrote:
>
> > OK so to answer the "is it easy to insert the patched scanner for
> > trafodion", the answer is no.
> >
>
> I suspected this.
>
>
>
> > Was easier on .98, but on 1.0 it was quite a challenge. All about
> > dealing with private attributes instead of protected that are not
> > visible to the PatchClentScanner class that extends ClientScanner.
> > Currently running the regression tests to see if there is no side
> effect...
> > Was able to demonstrate with breakpoint on next() waiting more than
> > 1 mn (the default lease timeout value) that with the patch things
> > gracefully reset and all is good, no row skipped or duplicated,
> > while without, I get the Scanner time out exception. Patch can be
> > turn on or off with a new
> key
> > in hbase-site.xml...
> > I will feel better when this will be deprecated :-).
> >
>
> Smile.
>
> Excellent. You have a patch for us then Eric?  Sounds like the
> interjection of your new Scanner would be for pre-2.0. For 2.0 we
> should just turn on this behavior as the default.
>
> Thanks,
> St.Ack
>
>
>
> > Eric Owhadi
> >
> > -----Original Message-----
> > From: saint.ack@gmail.com [mailto:saint.ack@gmail.com] On Behalf Of
> Stack
> > Sent: Friday, August 28, 2015 6:35 PM
> > To: HBase Dev List <dev@hbase.apache.org>
> > Subject: Re: Question on hbase.client.scanner.timeout.period
> >
> > On Fri, Aug 28, 2015 at 11:31 AM, Eric Owhadi
> > <eric.owhadi@esgyn.com>
> > wrote:
> >
> > > That sounds good, but given trafodion needs to work on current and
> > > future released version of HBase, unpatched, I will first
> > > implement a ClientScannerTrafodion (to be deprecated), inheriting
> > > from ClientScanner that will just overload the loadCache(),and
> > > make sure that the code that is picking up the right scanner based
> > > on scan object is bypassed to force getting the
> > > ClientScannerTrafodion when appropriate.
> > > Not very elegant, but need to take into consideration trafodion
> > > deployment requirements.
> > > Then, if we do not discover any side effect during our QA related
> > > to this code I will port the fix on HBase to deprecate the custom
> > > scanner (probably first on HBase 2.0, then will let the community
> > > decide if this fix is worth it for back porting...). It will be a
> > > first for me, but that's great, I'll take your offer to help ;-)...
> > >
> >
> > Sweet. Suggest opening an umbrellas issue in hbase to implement this
> > feature. Reference HBASE-2161 (it is closed now). Link trafodion
> > issue to it. A subtask could have implementation in hbase 2.0,
> > another could be backport.
> >
> > Is is easy to insert your T*ClientScanner?
> > St.Ack
> >
> >
> >
> > > Regards,
> > > Eric
> > >
> > > -----Original Message-----
> > > From: saint.ack@gmail.com [mailto:saint.ack@gmail.com] On Behalf
> > > Of Stack
> > > Sent: Thursday, August 27, 2015 3:55 PM
> > > To: HBase Dev List <dev@hbase.apache.org>
> > > Subject: Re: Question on hbase.client.scanner.timeout.period
> > >
> > > On Thu, Aug 27, 2015 at 1:39 PM, Eric Owhadi
> > > <eric.owhadi@esgyn.com>
> > > wrote:
> > >
> > > > Oops, my bad, the related JIRA was :
> > > > https://issues.apache.org/jira/browse/HBASE-2161
> > > >
> > > > I am suggesting that the special code client side in loadCache()
> > > > of ClientScanner that is trapping the UnknownScannerException,
> > > > then on purpose check if it is coming from a lease timeout (and
> > > > not by a region move) to decide that it would throw a
> > > > ScannerTimeoutException instead of letting the code go and just
> > > > reset the scanner and start from last successful retrieve (the
> > > > way it works for an unknowScannerException due to a region moving).
> > > > By just removing the special handling that tries to
> > > > differentiate from unkownScannerException due to lease timeout,
> > > > we should have a resolution to JIRA 2161- And to our trafodion
> > > > issue.
> > > >
> > > > We are still protecting against dead client that would cause
> > > > resource leak at region server, since we keep the lease timeout
> > > > mechanism.
> > > >
> > > > Not sure if I have overlooked something, as usually, code is
> > > > here for a reason :-)...
> > > >
> > > >
> > > Your proposal sounds good to me.
> > >
> > > Scanner works the way it does because it has always work this way
> > (smile).
> > > A while back, one of the lads suggested we do like dynamodb and
> > > have scanner have no state on the serverside, the scan next would
> > > just supply all necessary context. It was argued against because
> > > serverside setup is so costly. Your suggestion is similar only we
> > > do it only if Scanner has timed out.
> > >
> > > Suggest we keep the current semantic in 1.x at least. We could
> > > flip to your behavior in 2.x.  Meantime, you'd have to ask for it
> > > when you set up your Scan object by setting a flag.
> > >
> > > Would that work? If you want to have a go at it, I could help out
> > > on the issue.
> > >
> > > St.Ack
> > >
> > >
> > >
> > >
> > > > Regards,
> > > > Eric
> > > >
> > > >
> > > >
> > > > -----Original Message-----
> > > > From: saint.ack@gmail.com [mailto:saint.ack@gmail.com] On Behalf
> > > > Of Stack
> > > > Sent: Thursday, August 27, 2015 3:23 PM
> > > > To: HBase Dev List <dev@hbase.apache.org>
> > > > Subject: Re: Question on hbase.client.scanner.timeout.period
> > > >
> > > > On Tue, Aug 25, 2015 at 8:03 AM, Eric Owhadi
> > > > <eric.owhadi@esgyn.com>
> > > > wrote:
> > > >
> > > > > Hello St.Ack,
> > > > > Thanks for your pointer, but I had already investigated JIRA
> > > > > https://issues.apache.org/jira/browse/HBASE-13090
> > > > > Unfortunately, this heartbeat will protect against rpc
> > > > > timeout, not server side lease timeout that we are experiencing
> > > > > right now.
> > > > > I have not seen an active JIRA fixing our issue.
> > > > > Only https://issues.apache.org/jira/browse/HBASE6121 is
> > > > > complaining about the exact same issue, but was never resolved.
> > > > >
> > > > >
> > > > Which issue? https://issues.apache.org/jira/browse/HBASE-6121
> > > > seems unrelated.
> > > >
> > > >
> > > >
> > > > > The heartbeat JIRA in 13090 protect for situation where server
> > > > > scanner takes so long to retrieve the highly filtered
> > > > > information, that it exceeds the RPC timeout (hbase.rpc.timeout).
> > > >
> > > >
> > > >
> > > > > The timeout we are experiencing is the
> > > > > hbase.client.scanner.timeout.period,
> > > > > also deprecatedly known as hbase.regionserver.lease.period The
> > > > > mechanism is different: here, region server scanners wants to
> > > > > protect themselves against dead clients that would not perform
> > > > > "close", and allow releasing server side scanner resources. To
> > > > > do that, a lease mechanism is implemented, and if between 2
> > > > > next() call, more than hbase.regionserver.lease.period occurs,
> > > > > the server side scanner will have been forced closed by this
> > > > > lease timeout safety mechanism. On late next() call, client
> > > > > will receive a DNRIOE of type unknownScannerException, and the
> > > > > client will assess that it is coming most likely from the
> > > > > lease timeout (and not from a region move), therefore throwing
> > > > > an exception instead of reset scanner (for the region move
> > > > > scenario).
> > > > >
> > > > > Hbase 1.1 does not address, as far as I have researched, the
> > > > > hbase.client.scanner.timeout.period issue we are facing.
> > > > >
> > > > >
> > > >
> > > > Can you not have the high-level query that is being fed by a
> > > > scan do HBASE-13333? That is, tickle, the ongoing scan on
> > > > occasion just to say that I'm still alive?
> > > >
> > > > Otherwise, what would you suggest? A scan that does not timeout?
> > > > Or the client being able to set a timeout in the Scan passed to
> > > > the
> > server?
> > > >
> > > > Sorry for late reply,
> > > > St.Ack
> > > >
> > > >
> > > >
> > > > > And yes, we will move to Hbase 1.1, and 1.0 as Cloudera and
> > > > > Hortonworks are having version mismatch on the next official
> > > > > builds trafodion will support.
> > > > >
> > > > > So my question is still open?
> > > > >
> > > > > Best regards,
> > > > > Eric Owhadi
> > > > >
> > > > >
> > > > >
> > > > > -----Original Message-----
> > > > > From: saint.ack@gmail.com [mailto:saint.ack@gmail.com] On
> > > > > Behalf Of Stack
> > > > > Sent: Monday, August 24, 2015 11:07 PM
> > > > > To: HBase Dev List
> > > > > Subject: Re: Question on hbase.client.scanner.timeout.period
> > > > >
> > > > > On Mon, Aug 24, 2015 at 4:48 PM, Eric Owhadi
> > > > > <eric.owhadi@esgyn.com>
> > > > > wrote:
> > > > >
> > > > > > Hello everyone,
> > > > > > We have been facing a situation on trafodion, where we are
> > > > > > hitting the hbase.client.scanner.timeout.period scenario:
> > > > > > basically, when doing queries that require spilling to disk
> > > > > > because of high complexity of what is involved, the
> > > > > > underlying hbase scanner serving one of the operation
> > > > > > involved in the complex query cannot call the next() withing
> > > > > > the timeout specify... too busy taking care of other business.
> > > > > > This is legit scenario, and I was wondering why in the code,
> > > > > > special care is done to make sure that client side, if a
> > > > > > DNRIOE of type unknownScannerException shows up, and the
> > > > > > hbase.client.scanner.timeout.period time elapsed, we make
> > > > > > sure to throw a scannerTimeoutException, instead of just let
> > > > > > it go and reset scanner.
> > > > > >
> > > > > > Scanners were redone in hbase 1.1. Can Trafodion come up
> > > > > > onto hbase
> > > > 1.1?
> > > > > See
> > > > > https://blogs.apache.org/hbase/entry/scan_improvements_in_hbas
> > > > > e_1
> > > > > for summary.
> > > > > St.Ack
> > > > >
> > > > >
> > > > >
> > > > > > I imagine that the lease time out implementation on region
> > > > > > server side is supposed to protect from resource leak of
> > > > > > scanner object server side. But I am not sure why we would
> > > > > > make it so that client side throw this timeout exception,
> > > > > > when in fact what just happened was that client was too busy
> > > > > > to call next() on
> time.
> > > > > >
> > > > > > I am sure there is a reason, but cannot figure it out :-).
> > > > > >
> > > > > > BTW, I found this JIRA, talking about exact same thing:
> > > > > > https://issues.apache.org/jira/browse/HBASE61-21 but with no
> > > > resolution.
> > > > > >
> > > > >
> > > > >
> > > > > > Any help understanding the reason of the timeout thrwown
> > > > > > client side instead of an automatic reset would be much
> > > > > > appreciated, Best regards, Eric Owhadi
> > > > > >
> > > > >
> > > >
> > >
> >
>

Mime
View raw message