You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@cassandra.apache.org by Ashwin Jayaprakash <as...@gmail.com> on 2010/03/06 01:54:33 UTC
Unreliable transport layer
Hey guys! I have a simple question. I'm a casual observer, not a real
Cassandra user yet. So, excuse my ignorance.
I see that the Gossip feature uses UDP. I was curious to know if you guys
faced issues with unreliable transports in your production clusters? Like
faulty switches, dropped packets etc during heavy network loads?
If I'm not mistaken are all client reads/writes doing point-to-point over
TCP?
Thanks,
Ashwin.
Re: Unreliable transport layer
Posted by Boris Shulman <sh...@gmail.com>.
So, do you think it will be problematic on a large scale? If yes is
it configurable in 0.6?
On Sat, Mar 6, 2010 at 6:19 PM, Avinash Lakshman
<av...@gmail.com> wrote:
> We have observed this. But in practice it doesn't cause any deleterious
> effects. IMHO detecting false failures of nodes is the most dangerous thing
> that could result from this kind of behavior. But that is why we have an
> Accrual FD which reacts and adjusts to these conditions. But having said
> that moving TCP is not a bad option at all at relatively small scale.
>
> Cheers
> Avinash
>
> On Fri, Mar 5, 2010 at 4:54 PM, Ashwin Jayaprakash
> <as...@gmail.com> wrote:
>>
>> Hey guys! I have a simple question. I'm a casual observer, not a real
>> Cassandra user yet. So, excuse my ignorance.
>>
>> I see that the Gossip feature uses UDP. I was curious to know if you guys
>> faced issues with unreliable transports in your production clusters? Like
>> faulty switches, dropped packets etc during heavy network loads?
>>
>> If I'm not mistaken are all client reads/writes doing point-to-point over
>> TCP?
>>
>> Thanks,
>> Ashwin.
>>
>>
>
>
Re: Unreliable transport layer
Posted by Avinash Lakshman <av...@gmail.com>.
I think Open PGM uses IP mulitcast and this is not available in all data
centers. For instance inside FB mulicast across racks is disabled for a
multitude of reasons. However using TCP should not be a major concern as
long as some discipline is in place w.r.t how size of packets are restricted
as you scale to several hundreds of nodes. For instance you could cap the
packet size at 4K post compression and that will let you scale to perhaps a
few thousand nodes.
Cheers
Avinash
On Sun, Mar 28, 2010 at 11:40 AM, Ashwin Jayaprakash <
ashwin.jayaprakash@gmail.com> wrote:
>
> Or, use OpenPGM (http://code.google.com/p/openpgm/) as an alternative? But
> they don't have any Java bindings yet.
>
> ZeroMQ (http://www.zeromq.org/) uses this.
> --
> View this message in context:
> http://n2.nabble.com/Unreliable-transport-layer-tp4684470p4814206.html
> Sent from the cassandra-user@incubator.apache.org mailing list archive at
> Nabble.com.
>
Re: Unreliable transport layer
Posted by Ashwin Jayaprakash <as...@gmail.com>.
Or, use OpenPGM (http://code.google.com/p/openpgm/) as an alternative? But
they don't have any Java bindings yet.
ZeroMQ (http://www.zeromq.org/) uses this.
--
View this message in context: http://n2.nabble.com/Unreliable-transport-layer-tp4684470p4814206.html
Sent from the cassandra-user@incubator.apache.org mailing list archive at Nabble.com.
Re: Unreliable transport layer
Posted by Avinash Lakshman <av...@gmail.com>.
We have observed this. But in practice it doesn't cause any deleterious
effects. IMHO detecting false failures of nodes is the most dangerous thing
that could result from this kind of behavior. But that is why we have an
Accrual FD which reacts and adjusts to these conditions. But having said
that moving TCP is not a bad option at all at relatively small scale.
Cheers
Avinash
On Fri, Mar 5, 2010 at 4:54 PM, Ashwin Jayaprakash <
ashwin.jayaprakash@gmail.com> wrote:
> Hey guys! I have a simple question. I'm a casual observer, not a real
> Cassandra user yet. So, excuse my ignorance.
>
> I see that the Gossip feature uses UDP. I was curious to know if you guys
> faced issues with unreliable transports in your production clusters? Like
> faulty switches, dropped packets etc during heavy network loads?
>
> If I'm not mistaken are all client reads/writes doing point-to-point over
> TCP?
>
> Thanks,
> Ashwin.
>
>
>
Re: Unreliable transport layer
Posted by Benjamin Black <b...@b3k.us>.
Because it is extremely well-understood, handles a lot of the
reliability needs itself, and nothing more is required for the
application.
On Thu, Jul 29, 2010 at 7:02 PM, ChingShen <ch...@gmail.com> wrote:
> Why? What reasons did you choose TCP?
>
> Shen
>
> On Sat, Mar 6, 2010 at 9:15 AM, Jonathan Ellis <jb...@gmail.com> wrote:
>>
>> In 0.6 gossip is over TCP.
>>
>> On Fri, Mar 5, 2010 at 6:54 PM, Ashwin Jayaprakash
>> <as...@gmail.com> wrote:
>> > Hey guys! I have a simple question. I'm a casual observer, not a real
>> > Cassandra user yet. So, excuse my ignorance.
>> >
>> > I see that the Gossip feature uses UDP. I was curious to know if you
>> > guys
>> > faced issues with unreliable transports in your production clusters?
>> > Like
>> > faulty switches, dropped packets etc during heavy network loads?
>> >
>> > If I'm not mistaken are all client reads/writes doing point-to-point
>> > over
>> > TCP?
>> >
>> > Thanks,
>> > Ashwin.
>> >
>> >
>> >
>
>
>
>
Re: Unreliable transport layer
Posted by ChingShen <ch...@gmail.com>.
Why? What reasons did you choose TCP?
Shen
On Sat, Mar 6, 2010 at 9:15 AM, Jonathan Ellis <jb...@gmail.com> wrote:
> In 0.6 gossip is over TCP.
>
> On Fri, Mar 5, 2010 at 6:54 PM, Ashwin Jayaprakash
> <as...@gmail.com> wrote:
> > Hey guys! I have a simple question. I'm a casual observer, not a real
> > Cassandra user yet. So, excuse my ignorance.
> >
> > I see that the Gossip feature uses UDP. I was curious to know if you guys
> > faced issues with unreliable transports in your production clusters? Like
> > faulty switches, dropped packets etc during heavy network loads?
> >
> > If I'm not mistaken are all client reads/writes doing point-to-point over
> > TCP?
> >
> > Thanks,
> > Ashwin.
> >
> >
> >
>
Re: Unreliable transport layer
Posted by Jonathan Ellis <jb...@gmail.com>.
In 0.6 gossip is over TCP.
On Fri, Mar 5, 2010 at 6:54 PM, Ashwin Jayaprakash
<as...@gmail.com> wrote:
> Hey guys! I have a simple question. I'm a casual observer, not a real
> Cassandra user yet. So, excuse my ignorance.
>
> I see that the Gossip feature uses UDP. I was curious to know if you guys
> faced issues with unreliable transports in your production clusters? Like
> faulty switches, dropped packets etc during heavy network loads?
>
> If I'm not mistaken are all client reads/writes doing point-to-point over
> TCP?
>
> Thanks,
> Ashwin.
>
>
>