airavata-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Shenoy, Gourav Ganesh" <>
Subject Re: Linked Container Services for Apache Airavata Components - Phase 1 - Requirement identification
Date Thu, 02 Nov 2017 03:40:28 GMT
Hi Dimuthu,

Sorry for catching up late on your emails and thanks again for summarizing your findings.
I believe this is really helpful for anyone who wishes to understand the intricacies involved
in re-designing the architecture and at the same time understanding the technologies. I have
a few comments though; please find my feedback inline in blue.

Thanks and Regards,
Gourav Shenoy

From: DImuthu Upeksha <>
Reply-To: "" <>
Date: Monday, October 16, 2017 at 1:48 AM
To: "" <>
Subject: Re: Linked Container Services for Apache Airavata Components - Phase 1 - Requirement

Hi All,

Thanks for all the valuable feedback that you have provided so far and please find attached
document that contains the evaluation of Kubernetes, DC/OS and Helix as candidate platforms
to deploy Airavata mircoservices. Use the google doc [1] to provide your suggestions and comments.

Summary of the document:

Considering all the facts, I believe that Kuberentes is more suitable for our use cases.

Advantages of Kubernetes over DC/OS

1. DC/OS uses Marathon framework to perform container orchestration. Marathon framework should
be deployed on Mesos framework. So, from the architectural point of view, there are two framework
levels which one installed on top of another. Reason for this is because Mesos is a generic
framework to deploy any application and Marathon is the one which adds the value by providing
container orchestration. Although this is a good design from design point of view, it requires
more resources and involves a lot of complexities when it comes to an production deployment.
From the other hand, Kubernetes has built from the scratch to support container orchestration
so it can do the same work which DC/OS performs, with less resources and less complexity.

DC/OS is an operating system which uses Mesos as the underlying resource manager (to add abstraction
if your infrastructure spans multiple clouds OR if you have an on-premise infra), whereas
Marathon is “generally” the scheduler used by DC/OS to orchestrate applications/services
as containers over the infrastructure managed by Mesos. I say “generally”, because DC/OS
also allows you to run one-off jobs (services are long running, jobs are one-time run), and
uses Aurora as the scheduler for it. The reason I am reiterating this is because of the following:

  1.  You mentioned DC/OS using more resources, which I believe might not be accurate (I might
be wrong). DC/OS has evolved out of an ecosystem which comprises of resource management and
application management.
  2.  The complexity involved in getting a working “production grade” Kubernetes cluster
setup is far more complicated than getting DC/OS bootstrapped.
  3.  It really doesn’t matter “what scheduler” we use to orchestrate the containers,
because in the end the battle is between Kubernetes and Marathon, and not Kubernetes vs DC/OS.
In fact, DC/OS (as an operating system) supports the use of multiple orchestration schedulers
(like running multiple browsers over a Mac/Windows machine) – so as a devops engineer, DC/OS
allows you to decide whether to use Marathon or Kubernetes. You will be pleasantly surprised
to know that Kubernetes is a first-class package on DC/OS.

2. Kubernetes has fewer components compared to DC/OS and comparatively lighter than DC/OS
framework deployment. Lesser components makes it easy to maintain and monitor the framework.

As I mentioned in my comments above, I do not think this might be accurate. Yes, maintainability
is a very important aspect but both these frameworks are proven to be easily maintainable.

3. When it comes to high availability deployments, DC/OS has more components (Mesos masters,
Marathon masters) to make available than Kubernetes. This makes the production deployment
and management process complex and tiresome.

Same as points (1, 2) above, the real battle is between Kubernetes and Marathon. Mesos is
just the resource layer which allows you to easily manage applications across heterogeneous/hybrid

4. Having the capability of deploying non container based applications in the platform is
not one of our requirements. So that feature of DC/OS will rarely be beneficial for us.

Partially agreed. Yes, the discussion is towards containerizing Airavata components. But knowing
that DC/OS also supports running distributed services is beneficial.

5. Kubernetes has a huge community and a lot of exposure to opensource arena throughout the
inception. DC/OS is mainly managed by Mesosphere even though it has been made opensource in
2016. Most of the feature designs and issue discussions are well documented in Kubernetes
github repository which makes it really easy to track and solve when come to an issue of the
framework, where as in DC/OS that ecosystem is still at the very early age.

I agree. Kubernetes has a larger adoption and good community support.

6. There are lot of vendors working on Kubernetes and currently there are significant amount
of tools developed around Kubernetes to deploy, monitor and manage Kubernetes clusters whereas
in DC/OS we can not see that amount of traction.

Yes, this is a very good point. Just deploying Airavata over DC/OS or Kubernetes is not the
goal. Eventually having a streamlined CI/CD process is extremely essential. Yes, as Kubernetes
has a wider adoption there are a lot of tools available for CI/CD over Kubernetes. Although
I haven’t followed up or tried it myself, but looks like DC/OS has fairly good support for
elastic CI/CD pipelines. Mesosphere ecosystem boasts about it here.<>

NOTE: I am not advocating DC/OS over Kubernetes. I just wanted to clarify some of the subtle
differences between the two. Most people confuse between them as being competing technologies
(me included), but this blog<> throws
some light over this topic.

Advantages of Kubernetes over Helix

While the overall differences between Kubernetes and Helix makes total sense, however since
we are considering only the “deployment” aspect here (containers being the protagonist
in our story), Helix is completely out of the equation. We are only considering Helix as the
distributed task execution framework for managing workloads “within” Airavata. The idea
was to leverage Helix’s task execution APIs to define custom task executors, and orchestrate
the DAGs defined using Helix nomenclature.

Although Helix provides cluster management capabilities (and honestly that was the sole purpose
behind the team at LinkedIn building Helix), we are not interested in using Helix for managing
Airavata microservices. Rather, we need to identify the best way to build our micro-services
around Helix’s task execution framework. I am not saying this is the ideal way to solve
our problem, but certainly is one of the powerful candidates out there.

1. Containers (docker) are the proper way to deploy microservice due to its platform agnostic
packaging model, resource limitation capability and ease of distributability.

2. Helix doesn’t support container deployment out of the box where as Kubernetes itself
is a container orchestration framework

3. Helix has lesser components compared to Kubernetes however considering the features Kubernetes
provides over Helix is justifiable for having more components

4. In Helix, microservice application logic is tightly coupled with the Helix participant
code (more precisely to the State Model). This provides several issues in a production deployments

  *   Update process become very complex as we have to restart participant nodes for each
update. And it will affect other services as well which is not acceptable under any circumstance.
  *   We can not limit resources for each microservice as all are run inside the same JVM
of Participant node.

5. Kubernetes in contrast have clearly defined boundaries between application logic and the
runtime framework. Application logic is bundled as a docker image and they are run as separate
processes which makes the update process and resource limitation very easy

6. Kubernetes comes with the service discovery and load balancing out of the box whereas Helix
doesn’t provide such features by default.

7. Kubernetes has a well defined and scalable node affinity API but in Helix we have to write
custom Re-Balancers to achieve it and it is not scalable either.

8. It is very complex to come up with a proper CI/CD pipeline for Helix as application code
is tightly coupled to the framework. Kubernetes has a straightforward way to integrate CI/CD
pipelines to test and deploy microservices

9. Kubernetes has a comprehensive role based access model (RBAC) to authorize the resources
while Helix doesn’t have as such



On Mon, Oct 9, 2017 at 9:43 AM, Supun Nakandala <<>>
+1 for the idea.

On Sun, Oct 8, 2017 at 2:52 AM, DImuthu Upeksha <<>>
Hi Supun,

My belief also letting orchestrator to determine the worker to run particular job is complex
to implement and will make the maintainability of orchestrator code quite hard in long run.
I'm also in partially agreement with embedding a worker inside the firewall protected resource
but I guess we can improve it further to make homogenous and stateless. Have a look at following

In above design we keep all the workers outside and keep a daemon inside the protected resource
to securely communicate with workers. Then the problem is how do we make the worker homogenous
as this is still just adding another layer to the solution stated above. Trick is, we decouple
the communication between worker and resource. Communication to any resource is being done
through a well defined API. Speaking in java

public interface CommunicationInterface {
      public String sshToResource(String resourceIp, String command);
      public void transferDataTo(String resourceIp, String target, InputStream in);
      public void transferDataFrom(String resourceIp, String target, OutputStream out);
Implementation of this API might change according to the resource. We keep a separate Catalog
that will cater the libraries that have the implementation specific to each resource. For
example, if Worker 1 needs to talk to Resource 1 which acts behind a firewall and the Airavata
communication agent is placed inside, it will query the Catalog for the Resource 1 and fetch
the library that implemented CommunicationInterface to talk securely with Airavata Agent.
If it wants to talk to Resource 2, another library will be fetched from Catalog that has default
implementations. Once those SDKs are fetched, they are loaded into the JVM at runtime using
a class loader and communication will be done afterwards.

We can improve this by caching libraries inside workers and reusing them as much as possible
to limit number of queries to Catalog from workers.

Advantage of this is, we can add resources with different security levels without changing
the Worker implementations. Only thing we have to do is to come up with an agent and a library
to talk with agent. Then add them to Catalog and rest will be taken cared by the framework.
This model is analogous to the sql drivers that we use in java to connect to databases.

Please note that I came up with this design based on the limited knowledge I have in Airavata
Workers and Resources. There will be lot of corner cases that I have not identified. Your
views and ideas are highly appreciated.


On Sun, Oct 8, 2017 at 10:51 AM, Supun Nakandala <<>>
Hi Dimuthu,

Thank you for the very good summary. I think you have covered almost all the things.

I would also like to mention one other futuristic requirements that I think will be important
in this discussion.

In my opinion going forward, Airavata will get the requirement of working with firewall protected
resources. In such cases, workers which are residing outside will not be able to communicate
with the protected resources. What we initially thought was to deploy a special type of worker
which will be placed inside the firewall-protected network and will coordinate with Airavata
orchestrator to execute actions. One such tool which is used by ServiceNow in enterprise settings
is the MidServer ( The downside
of this approach is that it breaks our assumption of all workers being homogenous and therefore
require orchestrator to be worker aware. Perhaps, instead of workers picking work we can design
such that orchestrator will grant work to the corresponding work. But this incorporates a
lot of complexity on the orchestrator's side.

On Oct 5, 2017 10:47 AM, "DImuthu Upeksha" <<>>
Hi Gaurav,

Thanks a lot for the detailed description about DC/OS and how it can be utilized in Airavata.
Seems like it is an interesting project and I'll add it to the technology list that are to
be evaluated.

When selecting a technology, in addition to the features it provides, we might have to take
some non-functional features like the community participation (committers, commits and forks),
number of customers  who are  running it  in production environments, maturity of the project
and the complexity it brings in to the total system into the consideration. So I'll first
try to go through the resources (documentation and source) and try to grab concepts of DC/OS
and hopefully I can work with you to dig deeper to understand more about DC/OS


On Thu, Oct 5, 2017 at 8:50 PM, Shenoy, Gourav Ganesh <<>>
Sorry, missed the attachment in my previous email.

PS: DC/OS is just a recommendation for performing containerized deployment and application
management for Airavata. I would be happy to consider alternative frameworks such as Kubernetes.

Thanks and Regards,
Gourav Shenoy

From: "Shenoy, Gourav Ganesh" <<>>
Reply-To: "<>" <<>>
Date: Thursday, October 5, 2017 at 11:16 AM

To: "<>" <<>>
Subject: Re: Linked Container Services for Apache Airavata Components - Phase 1 - Requirement

Hi Dimuthu,

Very good summary! I am not sure if you have, but DC/OS (DataCenter Operating System) is a
container orchestration platform based on Apache Mesos. The beauty of DC/OS is the ease and
simplicity of development/deployment; yet being extremely powerful in most of the parameters
– multi-datacenter, multi-cloud, scalability, high availability, fault tolerance, load balancing,
and more importantly the community support is fantastic.

DC/OS has an exhaustive service catalog, it’s more like a PAAS for containers (not just
restricted to containers though) – you can run services like Spark, Kafka, RabbitMQ, etc
out of the box with a single click install. And Apache Mesos as the underlying resource manager
makes it seamless to deploy applications across different datacenters. There is a concept
of SERVICE vs JOB – service is considered long running and DC/OS will make sure it keeps
it running (if a service fails, it spins up a new one), whereas jobs are one time executors.
This comes handy for using DC/OS as a target runtime for Airavata.

We used DC/OS for our class project to run the distributed task execution prototype we built
(which uses RabbitMQ messaging). Here’s a link to the blog I have explaining the process: . I have
also attached a PDF paper we wrote as part of the class explaining the task execution process
and one solution using rabbitmq messaging.

I had also started with the work of containerizing Airavata and a unified build + deployment
mechanism with CI CD on DC/OS. Unfortunately, I couldn’t complete it due to time constraints,
but I would be more than happy to work with you on this. Let me know and we can coordinate.

Thanks and Regards,
Gourav Shenoy

From: DImuthu Upeksha <<>>
Reply-To: "<>" <<>>
Date: Thursday, October 5, 2017 at 9:52 AM
To: "<>" <<>>
Subject: Re: Linked Container Services for Apache Airavata Components - Phase 1 - Requirement

Hi Marlon,

Thanks for the input. I got your idea of availability mode and will keep in mind while designing
the PoC. CI/CD is the one I have missed and thanks for pointing it out.


On Thu, Oct 5, 2017 at 7:04 PM, Pierce, Marlon <<>>
Thanks, Dimuthu, this is a good summary. Others may comment about Kafka, stateful versus stateless
parts of Airavata, etc.  You may also find some of this discussion on the mailing list archives.

Active-active vs. active-passive is a good question, and we have typically thought of this
in terms of individual Airavata components rather than the whole system.  Some components
can be active-active (like a stateless application manager), while others (like the orchestrator
example you give below) are stafefull and may be better as active-passive.

There is also the issue of system updates and continuous deployments, which could be added
to your list.


From: "<>" <<>>
Reply-To: "<>" <<>>
Date: Thursday, October 5, 2017 at 2:40 AM
To: "<>" <<>>
Subject: Linked Container Services for Apache Airavata Components - Phase 1 - Requirement

Hi All,

Within last few days, I have been going through the requirements and design of current setup
of Airavata and I identified following ares as the key focusing areas in the technology evaluation

Micorservices deployment platform (container management system)

Possible candidates: Google Kubernetes, Apache Mesos, Apache Helix
As the most of the operational units of Airavata is supposed to be moving into microservices
based deployment pattern, having a unified deployment platform to manage those microservices
will make the DevOps operations easier and faster. From the other hand, although writing and
maintaining a single micro service is a somewhat straightforward way, making multiple microservies
running, monitoring and maintaining the lifecycles manually in a production environment is
an tiresome and complex operation to perform. Using such a deployment platform, we can easily
automate lots of pain points that I have mentioned earlier.


We need a solution that can easily scalable depending on the load condition of several parts
of the system. For example, the workers in the post processing pipeline should be able scaled
up and down depending on the events come into the message queue.


We need to support solution to be deployed in multiple geographically distant data centers.
When evaluating container management systems, we should consider this is as a primary requirement.
However one thing that I am not sure is the availability mode that Airavata normally expect.
Is it a active-active mode or active-passive mode?

Service discovery

Once we move in to microservice based deployment pattern, there could be scenarios where we
want service discovery for several use cases. For example, if we are going to scale up API
Server to handle an increased load, we might have to put a load balancer in between the client
and API Server instances. In that case, service discovery is essential to instruct the load
balancer with healthy API Server endpoints which are currently running in the system.

Cluster coordination

Although micorservices are supposed to be stateless in most of the cases, we might have scenarios
to feed some state to particular micorservices. For example if we are going to implement a
microservice that perform Orchestrator's role, there could be issues if we keep multiple instances
of it in several data centers to increase the availability. According to my understanding,
there should be only one Orchestrator being running at a time as it is the one who takes decisions
of the job execution process. So, if we are going to keep multiple instances of it running
in the system, there should be an some sort of a leader election in between Orchestrator quorum.

Common messaging medium in between mocroservices

This might be out of the scope but I thought of sharing with the team to have an general idea.
Idea was raised at the hip chat discussion with Marlon and Gaourav. Using a common messaging
medium might enable microservices to communicate with in a decoupled manner which will increase
the scalability of the system. For example there is a reference architecture that we can utilize
with kafka based messaging medium [1], [2]. However I noticed in one paper that Kafka was
previously rejected as writing clients was onerous. Please share your views on this as I'm
not familiar with the existing fan out model based on AMQP and  pain points of it.

Those are the main areas that I have understood while going through Airavata current implementation
and requirements stated in some of the research papers. Please let me know whether my understanding
on above items are correct and suggestions are always welcome :)



Marru, S., Gunathilake, L., Herath, C., Tangchaisin, P., Pierce, M., Mattmann, C., Singh,
R., Gunarathne, T., Chinthaka, E., Gardler, R. and Slominski, A., 2011, November. Apache airavata:
a framework for distributed applications and computational workflows. In Proceedings of the
2011 ACM workshop on Gateway computing environments (pp. 21-28). ACM.

Nakandala, S., Pamidighantam, S., Yodage, S., Doshi, N., Abeysinghe, E., Kankanamalage, C.P.,
Marru, S. and Pierce, M., 2016, July. Anatomy of the SEAGrid Science Gateway. In Proceedings
of the XSEDE16 Conference on Diversity, Big Data, and Science at Scale (p. 40). ACM.

Pierce, Marlon E., Suresh Marru, Lahiru Gunathilake, Don Kushan Wijeratne, Raminder Singh,
Chathuri Wimalasena, Shameera Ratnayaka, and Sudhakar Pamidighantam. "Apache Airavata: design
and directions of a science gateway framework." Concurrency and Computation: Practice and
Experience 27, no. 16 (2015): 4282-4291.

Pierce, Marlon, Suresh Marru, Borries Demeler, Raminderjeet Singh, and Gary Gorbet. "The apache
airavata application programming interface: overview and evaluation with the UltraScan science
gateway." In Proceedings of the 9th Gateway Computing Environments Workshop, pp. 25-29. IEEE
Press, 2014.

Marru, Suresh, Marlon Pierce, Sudhakar Pamidighantam, and Chathuri Wimalasena. "Apache Airavata
as a laboratory: architecture and case study for component- based gateway middleware." In
Proceedings of the 1st Workshop on The Science of Cyberinfrastructure: Research, Experience,
Applications and Models, pp. 19-26. ACM, 2015.


View raw message