I would like to make a slight adjustment to this generality.
That is in Cloud Computing, Amazon AWS/EC2 in particular.
Amazon has "regions" which are composed of 3 or more "zones" ... Each "zone" is
as nearly a seperate datacenter as physically possible except for connectivity.
That is they are generally built on seperate geographic topology (dont share
flood plains, earthquake fault lines etc) and are generally fed by issolated
infrastructure ( power, internet etc).
Yet ... they are close enough together that the latency between zones is faster
than the latency to an SSD drive write cycle. Most zones are within 2ms of
each other. Thats fast.
While this is not quite as fast as talking to nodes sharing the same rack and
bridge on a FIOS channel ... its still *damn fast*.
Creating a cluster across amazon zones within the same region can achieve both
HA and DR to a large degree. It's not perfect as its *possible* (I belive its
happened once in the last 10 years) for
multiple zones in a region to fail simultaneously ... but its vastly more
resilient than putting all your eggs in the same datacenter yet the latancy
between zones in the same region is so small that
creating clusters within a zone are extremely fast. Maybe not *quite* as fast
as on the same rack but it's a good compromise.
Especially if you make a hybrid. That is several nodes in each zone co-located
but cross-replicating ("local disk failover") across zones.
If you want real DR you should create a replica cluster in a different region
and do foreign replication to it.
But even without that ... its pretty amazing.
The result is a very fault resistant HA system that performs very well. A
good compromise.
Look forward to a coming announcement with will make such an architecture
easier to build.
But you dont *have* to wait. MarkLogic 6.0 runs great on Amazon EC2 out of the
box.
But we have made things simpler still to create a fault tolerant system that is
also HA and performing.
Nothing is perfect, everything is a compromise, but this combination of multi
zone single region clusters is as near as perfect as I can imagine, or anyone
else could build on their own.
-David
-----------------------------------------------------------------------------
David Lee
Lead Engineer
MarkLogic Corporation
[email protected]
Phone: +1 812-482-5224
Cell: +1 812-630-7622
www.marklogic.com
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Ron Hitchens
Sent: Tuesday, October 29, 2013 3:40 PM
To: MarkLogic Developer Discussion
Subject: Re: [MarkLogic Dev General] Reg: E-Node and D-Node configuration
You can think of a MarkLogic cluster as a single virtual
server. A cluster is made up of nodes (E, D or E/D) but the
cluster should be thought of as an indivisible unit.
D (data) nodes are MarkLogic processes that have forests
attached. E (evaluator) nodes are those nodes which run
XQuery/XSLT requests on an appserver. In a cluster, all nodes
share the same appserver configuration, so any node can be an
E node. Typically, when configuring dedicated E and D nodes,
you configure things to send requests to only those nodes that
you want to act as E's, allowing the others to act only as D's.
Communication between nodes in a cluster is basically this:
For queries (read-only) no locks are needed (read up on MVCC).
Each search operation is fired in parallel to every D node
in the cluster (this is the "map" phase). When the last D node
has responded, the E node can then merge the results (the "reduce").
So, the lower the latency in communication between nodes, the
better the overall throughput. You really don't want any slow
links between nodes in the cluster because it can slow down all
the E nodes.
For update (write), cluster-wide locks must be obtained for
documents that are, or might be, updated. All nodes in the cluster
must acknowledge the lock(s) before the update(s) can proceed. This
basically means that updates can't happen faster than the slowest
responding node in the cluster. Oh, and the locks need to be
released as well, via inter-node communication.
Again, bad for overall performance when communication links
between nodes slow down, even with super-fast, beefy hardware.
As Mike pointed out, clusters are not database replication.
You cluster to improve performance by spreading the immediate
work across multiple CPU and disks co-located together. You
can add synchronous replication between nodes in a cluster to
provide for HA failover in the event a node fails. This has a
latency cost, but makes the cluster more robust. You replicate
databases asynchronously between clusters to provide for disaster
recovery if an entire cluster is lost or becomes unreachable.
Hope that helps.
---
Ron Hitchens {[email protected]} +44 7879 358212
On Oct 28, 2013, at 10:03 PM, Arindam3 B <[email protected]> wrote:
>
> Thanks Mike for the great walkthrough. Just trying to understand more on the
> xqdp protocol. Can you throw some light on how it operates between enodes n
> dnodes?
>
> Thanks & Regards
> Arindam
>
> -----Michael Blakeley <[email protected]> wrote: -----
>
> =======================
> To: MarkLogic Developer Discussion <[email protected]>
> From: Michael Blakeley <[email protected]>
> Date: 10/28/2013 10:31PM
> Subject: Re: [MarkLogic Dev General] Reg: E-Node and D-Node configuration
> =======================
> Hosts within a cluster should have low-latency communications: gigabit
> ethernet or better. Ideally they should all be on the same switch and/or
> VLAN, with no router hops between hosts. If you try to set up a cluster
> across a WAN link you are likely to see poor performance and poor
> reliability. You might be trying to handle high availability (HA) and
> disaster recovery (DR) with a single cluster: that would be a mistake.
>
> For high availability, use a single cluster with low-latency communications.
> Configure forest replication and host failover to provide the desired degree
> of protection against host failures. The docs at
> http://docs.marklogic.com/guide/cluster/failover talk about this as
> "local-disk failover".
>
> For disaster recovery - scenarios where an entire data center goes offline -
> use database replication to a different cluster. This can use higher-latency
> communications, such as a WAN link. The docs at
> http://docs.marklogic.com/guide/database-replication describe this. The DR
> replica cluster can also implement local-disk failover to provide its own HA.
>
> -- Mike
>
> On 28 Oct 2013, at 06:41 , Arindam3 B <[email protected]> wrote:
>
>> Hi,
>>
>> I had a query regarding the E-Node and D-Node setup in Marklogic.
>>
>> In a distributed environment, if I plan to keep the Enodes and DNodes
>> separately in different physical locations over the LAN or WAN (across
>> geographies), what is the potential risk?
>> How does failover work in that scenario?
>> I have read that ENodes and DNodes communicate through XQDP protocol, so in
>> this case will there be performance issues?
>>
>> Does Marklogic recommend having ENode and DNode cluster in the same physical
>> box?
>> If so, then across the network if we have a set of E-D-Nodes, how is the
>> network latency reduced while synching the data during replication?
>>
>> If you can provide me with some information about XQDP protocol it would be
>> great!!
>>
>> Thanks & Regards
>> Arindam Bose
>> =====-----=====-----=====
>> Notice: The information contained in this e-mail
>> message and/or attachments to it may contain
>> confidential or privileged information. If you are
>> not the intended recipient, any dissemination, use,
>> review, distribution, printing or copying of the
>> information contained in this e-mail message
>> and/or attachments to it are strictly prohibited. If
>> you have received this communication in error,
>> please notify us by reply e-mail or telephone and
>> immediately and permanently delete the message
>> and any attachments. Thank you
>>
>>
>> _______________________________________________
>> General mailing list
>> [email protected]
>> http://developer.marklogic.com/mailman/listinfo/general
>
> _______________________________________________
> General mailing list
> [email protected]
> http://developer.marklogic.com/mailman/listinfo/general
>
> _______________________________________________
> General mailing list
> [email protected]
> http://developer.marklogic.com/mailman/listinfo/general
_______________________________________________
General mailing list
[email protected]
http://developer.marklogic.com/mailman/listinfo/general
_______________________________________________
General mailing list
[email protected]
http://developer.marklogic.com/mailman/listinfo/general
