helix-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sandeep Nayak <sand...@chegg.com>
Subject Re: Newbie Questions
Date Wed, 12 Feb 2014 19:30:46 GMT
Hi Kishore,

I have still to work through the algorithms on how I will distribute the
components to different nodes. I will loop back when I work through those.

As regards the asynchronous work, I will be happy to give feedback and
wherever necessary contribute if appropriate back to Helix.

Thanks,

Sandeep

From:  kishore g <g.kishore@gmail.com>
Reply-To:  "user@helix.apache.org" <user@helix.apache.org>
Date:  Wednesday, February 12, 2014 9:46 AM
To:  "user@helix.apache.org" <user@helix.apache.org>
Subject:  Re: Newbie Questions


Hi Sandeep,


Deployment: It makes sense at a high level. If I understand correctly, you
will generate the component to node mapping based on the resource usage
and I am assuming you would want to change this mapping dynamically as new
containers come up or existing
 containers fail or load changes. This fits well in the Helix
architecture. How do you plan to come up with the algorithm to do the
mapping between components and containers. We have multiple options here
but I would like to get more details before I give all
 the possible solutions. We are available on the free node IRC
#apachehelix and we can discuss the details any time on the IRC.


Regarding asynchronous work, this is definitely a great fit for Helix.We
are working on simplifying/enhancing the existing task rebalancer to
support task management as a first class citizen in Helix. Will be great
to get your feedback on that.


thanks,

Kishore G





On Wed, Feb 12, 2014 at 2:29 AM, Sandeep Nayak
<sandeep@chegg.com> wrote:

Hi Kishore,

Thank you for the detailed response, it certainly helps me understand
better.

For my case I have an home-grown 'container' deployed as a war file. The
container allows install/uninstall of components (I use the term component
to indicate a unit of functionality) at runtime. The war file runs on a
vanilla Tomcat server (node) hosted on EC2. The current deployment model
involves having a cluster of nodes each running the container and housing
'all' the components front ended by a load balancer so its a  homogeneous
install.

The container leverages registries to track artifacts used by the
components and I am looking to cluster these registries (this is where I
hope to use Helix) so that I can start moving away from a homogeneous
install. I plan for components to be installed on nodes based on their
resource usage rather than installing everything on each node i.e. move to
a heterogeneous install.

So considering this, my approach is to build intelligence in the container
(rather than the load-balancers) to route requests to nodes which have
those components. The way I plan to achieve this is by having the
registries carry a shared view of the cluster. It allows the ability to
build routing from this shared view. I already have a component deployment
model in place which tracks the install/uninstall of components through
events on each node and I think the Helix concepts map nicely to what I
already have. My events would simply propagate to the cluster and I can
get individual registries to listen to these events to build a shared
view.

I also have some other use-cases where I manage asynchronous work on
individual nodes which I would like to route to other nodes based on
capacity. In phase-II I plan to move towards installing components
on-demand on nodes which don't carry them but have capacity to handle
them, this would trigger when requests exceed capacity on existing servers.

Let me know what you think, it appears so far like Helix is a good fit for
what I am planning to use it for.

Thanks,

Sandeep


From:  kishore g <g.kishore@gmail.com>
Reply-To:  "user@helix.apache.org" <user@helix.apache.org>
Date:  Wednesday, February 12, 2014 1:44 AM
To:  "user@helix.apache.org" <user@helix.apache.org>
Subject:  Re: Newbie Questions


Hi Sandeep,



1) Difference between Apache Helix and Norbert.

While both projects originated at LinkedIn, there are some fundamental
differences between the two. Will try to explain the difference via a
simple use case of partitioned search index. In general you have 3
scenarios


* Start up --> Partition assignment: One needs to distribute the
partitions among the nodes in the cluster.


* In Norbert this process is manual. One needs to generate the mapping of
partition to node and push this configuration to each each server in the
system needs (this is generally done via a config file). At start up,
Norbert will simply write this configuration
 to zookeeper so that the clients can discover the partition to node
mapping.

* In Helix, the nodes simply join the cluster. Helix will inform the nodes
which partitions to host based on the objectives and constraints (A simple
objective would be to distribute the partitions evenly among the nodes)




* Failure: When a node fails, you have multiple options #1. Do nothing #2.
Re-assign the partitions hosted on the failed nodes to the remaining
nodes. #3 Start a new node and assign the partitions to that new node.


* Norbert simply informs you that a node left the group. You need to
program your requirement.
* Helix is capable of doing #1, #2 or even #3. #3 feature is work in
progress and is possible if the deployment system is flexible and allows
starting up process dynamically, If you are deploying in EC2, this should
be possible.




* Scalling: If you add more nodes to handle work load, you would want to
redistribute some of the work to new nodes.


* Norbert: This is again manual, you would have to change the
configuration and re-start all the nodes


* Helix: Helix would detect new nodes and fire appropriate transitions.




As you already mentioned Helix treats partitions, replicas, state,
transitions as first class citizens. What this means is you can not only
say how many partitions you have but also mention the number of replicas
for each partition. For example, for redundancy
 you can say you need 3 replicas for each partition and Helix will ensure
that 3 replicas exist for each partition.
2. No we haven't used Helix on Amazon, not sure if any one has done this.
There was another thread about this and he dint seem to think it would be
a big problem. It will help if you give us more information about your
application and set up.

3. "Incubating" indicates the state of the project in Apache, its does not
reflect the quality of the code or production readiness. Apart from
LinkedIn, companies like (Box, Instagram, Jboss jBPM clustering) have used
Helix in production.
Hope this helps.

Cheers

Kishore G












On Wed, Feb 12, 2014 at 12:55 AM, Sandeep Nayak
<sandeep@chegg.com> wrote:

Hi,

I am a newbie to Apache Helix and am evaluating technologies to build
clustered services. I have read through the documentation on the
http://helix.incubator.apache.org/ site but did not find answers to the
questions below so decided to ask them here.

(1) What is the difference between Apache Helix and LinkedIn Norbert? I
believe Norbert does not support state-machine transitions like Helix but
is there a document/summary on what are the differences so someone like me
can use that in my evaluation?

(2) Has Apache Helix been used on Amazon? Is there any documentation on
how to get this working?

(3) The website as 0.6.2-incubating-stable and 0.7.0-incubating-alpha.
Does incubating indicate the state of the project in Apache or is it
indicative of the production-readiness of the library? I imagine the
former because prior to getting to Apache the library was used at
linked-in, am I correct?


Thanks in advance,

Sandeep















Mime
View raw message