airavata-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Suresh Marru <sma...@apache.org>
Subject Re: Too Many Leaf Modules.
Date Tue, 02 Jun 2015 14:53:49 GMT
Let me try to put it in the terminology you are using:

GFac component level goal is to submit a job. In the RPC approach (which is defunct now) the
component level interface looks like - https://github.com/apache/airavata/blob/master/modules/gfac/gfac-thrift-descriptions/gfac.cpi.service.thrift#L47
<https://github.com/apache/airavata/blob/master/modules/gfac/gfac-thrift-descriptions/gfac.cpi.service.thrift#L47>

This is currently implemented as one such component implementation. The component implementation
is programmed internally against an internal interface - GFac which has implementations like
BetterGfacImpl

There is a Orchestrator component level interface methods to validate a experiment - https://github.com/apache/airavata/blob/master/modules/orchestrator/orchestrator-thrift-descriptions/orchestrator.cpi.service.thrift#L67
<https://github.com/apache/airavata/blob/master/modules/orchestrator/orchestrator-thrift-descriptions/orchestrator.cpi.service.thrift#L67>

You can implement this in any number of ways ultimately giving a boolean. There is one such
implementation defined by  AbstractOrchestrator interfaces and implemented by SimpleOrchestratorImpl
classes and so forth. 

So in your proposed maven modules, I am not seeing you pulling together component interfaces.
You are pulling together component implementations (defined as Java interfaces). Do Orchestrator
component level implementation GFac component level details fit in one place? 

What if you want to have a new component which can do submit Job with a totally new implementation
(which will have its own GFac interface and corresponding one or more GfacImpl’s)? Isn’t
this component level details which Airavata as a whole should not be concerned about?  

Suresh

> On Jun 2, 2015, at 10:28 AM, Shameera Rathnayaka <shameerainfo@gmail.com> wrote:
> 
> Hi Suresh, 
> 
> As you have fist hand experience of how airavata architecture evolve, let's do as you
suggest. But this is not correct way IMO. "Airavata Components will need to be loosely coupled
and should be developed at a different pace upgradable and replaceable independently" , we
can achieve these goals without any issue with the Maven modules that I am suggesting, may
be I am not descriptive enough to make you understand why this maven module refactoring is
so important for us. It would be great to know how you suggest to have project structure after
the refactoring.
> 
> Thanks, 
> Shameera. 
> 
> On Tue, Jun 2, 2015 at 9:49 AM, Suresh Marru <smarru@apache.org <mailto:smarru@apache.org>>
wrote:
> Hi Shameera,
> 
> We are getting close to be on same page but not quite yet. What I see missing is a reference
to a succinct Airavata architecture vision document. We have many papers on high level goals,
but what we need is a concise one pager on Architecture goals. I will work on it, but will
need some time. In the mean time I will suggest we proceed without yet merging the core (even
though it is only interfaces). It is not the additional 1MB I am worried about, I am worried
about going against the architectural principles laid out (which I will extract from papers
onto the website).
> 
> Very briefly, Airavata is a high level framework close to the business functionality
and assembles together multiple usecases. To make this challenge a conceivable effort, we
layer over rich lower level tools and framework. Initial struggle was to come up with a unified
API. But embracing thrift we addressed this by not unifying all usecases into one abstraction,
but multiple blocks of abstractions. This will require assembling together multiple components
at the implementation. Hypothetically you can assemble multiple cores, but that counters the
principle that components can be re-usable across recipes. There is no good description of
Airavata technical recipes but similar community effort in science gateways is at [1] [2],
there are about 33 recopies which need to be consolidated. 
> 
> I appreciate your enthusiasm, but need to slow down on drastic changes. We sure need
to get a stable one capability which is single job execution, but we need to do so without
jeopardizing larger goals of the project. I will very soon start discussions on these larger
goals. These will infer architectural principles which will need to be slowly build into 2
or 3 major versions. Understandably some of these will not make sense for 1.0. For instance
Airavata Components will need to be loosely coupled and should be developed at a different
pace upgradable and replaceable independently. This is requirement for making Airavata platform
ready. Just because we cannot do that yet, we do not want to go in other direction. Similarly,
there should be a minimal shared understanding of Airavata (data models, registry catalogs
and messaging contexts), but otherwise each component in the system will require minimal understanding
of remaining constituents of the airavata system. All of this goes against merging all interfaces
into a common core. 
> 
> For 0.16 release, lets please proceed with inconvenience of having multi-module parents.
Before next version we will need to do sufficient discussion to alter this. 
> 
> Suresh
> 
> [1] - http://dx.doi.org/10.1109/CLUSTER.2013.6702702 <http://dx.doi.org/10.1109/CLUSTER.2013.6702702>
> [2] - https://www.xsede.org/web/gateways/gateways-cookbook <https://www.xsede.org/web/gateways/gateways-cookbook>

> 
>> On Jun 1, 2015, at 11:13 PM, Shameera Rathnayaka <shameerainfo@gmail.com <mailto:shameerainfo@gmail.com>>
wrote:
>> 
>> Hi All, 
>> 
>> While reading through this thread again, I understood that we might be talking two
different things here. So I thought to explain the difference between Maven modules and Modular(
Separation of Concern) in Component based architecture. What I am suggesting is restructure
maven modules not to change existing Modular concept. If anyone say making one maven module
call "core" by putting all basic interfaces together will break Modular concept in Component
based architecture, that is not true. These are two different concepts. While component based
architecture is software engineering architectural design, Maven is used to project management.
Component based architecture may or may not have one maven module "core" in project which
has all basic interfaces. If this core maven module is so complex then it is worth to break
it to few modules. But for Airavata this doesn't relate, as suggested core maven module is
not that complex.  
>> Thanks, 
>> Shameera.
>> 
>> On Mon, Jun 1, 2015 at 9:09 PM, Shameera Rathnayaka <shameerainfo@gmail.com <mailto:shameerainfo@gmail.com>>
wrote:
>> Hi Raminder, 
>> 
>> Different implementations can depend on different versions of the same jar, as we
are using native java class loading, it doesn't support different versions to class loaded
and live in one runtime, unless we come up with OSGi like modular system or customize class
loader behavior which is tricky. If we are getting this issue with basic Airavata components
then we should fix this and use one version through all the components. If different plugable
implementations use different versions then both can't work together only one version will
be class loaded, this is a known restriction. Anyway I am not suggesting to have one bulky
Maven module, we should have some level of categorization. If  any interface has different
implementations and those are distinct enough, then having different modules make sense. 
>> 
>> Definitely we must remove implementation classes from current core modules. With
multiplex thrift support which Suresh is working on, for me it make sense to have all services
in one maven module call "service". Anyway let's think other alternatives too. 
>> 
>> Thanks, 
>> Shameera.
>> 
>> On Mon, Jun 1, 2015 at 1:25 PM, Raminderjeet Singh <raminderjsingh@gmail.com <mailto:raminderjsingh@gmail.com>>
wrote:
>> Hi Shameera,
>> 
>> I think, we are tying to deal with 2 problems. Code manageability/usage and distribution.
 
>> 
>> In GFAC, I noticed that there is a dependencies on modules, which should not exist
like GFAC's GSISSH module dependent on SSH module and both the modules depending on GSISSH
library. Reason of such depend is duplication of utility methods (see GFACSSHUtils and GFACGSISSHUtils
for duplicate code) in both of these modules and we need to fix it. I think there are similar
examples in different modules. According to me, if 60% code is duplicate, we need to merge
the module as one or come. Other way of saying is we should only create an module when its
needed. Local,SSH,GSISSH should not be separate modules as they have lot in common. Just to
give you some background on GFAC modules, they were created for the purpose of having different
flavors of GFAC. This was mainly done to fix the problem of jar dependencies (different security
dependencies) of GFAC modules in a JVM. It was effecting modules to work together. An example
in the past was, Unicore, GRAM and GSISSH modules did not work together so we had to spend
time to inspect distribution to find which runtime dependency is causing problem and fix it.
With the current design we can create individual version of GFAC. We still need to enhance
GFAC service to have a flavor of GFAC registered with type and route the jobs. So we need
some of this flavoring support in GFAC.
>> 
>> I agree with you that we have too many core modules and in core modules like GFAC
core, we have implementations (e.g. BetterGFACImpl). We should move implementations to Gfac
service. Lets talk about pros and cons about having a single core and then we can decide how
to proceed. Currently airavata does not provide a single view to all of its functionality.
API server was designed to do that but its also overloaded with lot of implementation details.
If I am following your advice right, having a common airavata core will definitely help developer
to think about airavata as a system and design new components with system prospective so +1
for something like this. Only drawback can be, taking away some flexibly from a component
developer, which is anyhow good for airavata system point of view. Before we just into conclusion,
We need to just evaluate how it will work with our thrift services design.
>> 
>> Thanks
>> Raminder
>> 
>>  
>> 
>> On Sat, May 30, 2015 at 11:55 PM, Shameera Rathnayaka <shameerainfo@gmail.com
<mailto:shameerainfo@gmail.com>> wrote:
>> Hi Suresh, 
>> 
>> You are thinking about deployment perspective while I am thinking about dependency
issue. With my suggestion for each component distribution will be increased by less than 1MB,
 because only the interfaces are in the core. And in runtime those interfaces will not be
loaded. Thinking the trouble we are getting at development time and code maintain issues.
I think we can bare with that 1MB. 
>> 
>> Thanks, 
>> Shameera.
>> 
>> On Sat, May 30, 2015 at 10:45 PM, Suresh Marru <smarru@apache.org <mailto:smarru@apache.org>>
wrote:
>> Shameera,
>> 
>> Every component has in its own thrift service interface (registry and messaging have
exceptions). Every component will need to have a dependency to airavata data models (which
includes util classes) and probably registry and messaging. if a component A needs to invoke
component B via RPC call, then it just needs to include its component A’s thrift client.
If the communication is through work queue’s then there is no dependency between them. Can
you describe what you want to propose in this context? 
>> 
>> Suresh
>> 
>>> On May 30, 2015, at 8:36 PM, Shameera Rathnayaka <shameerainfo@gmail.com <mailto:shameerainfo@gmail.com>>
wrote:
>>> 
>>> Hi Suresh, 
>>> 
>>> Spark is not the right comparison for this discussion. I have been a spark incubation
mentor and have been following the code organization since its early days. All of the spark
components you mentions rely on the core. 
>>> 
>>> ​Yes, that is what I highlight , each components doesn't have their own core
modules which has the interfaces. Spark SQL module has core submodule but all interfaces reside
in main core module.
>>> 
>>> 
>>> Let me step back and ask, what is the problem you are trying to solve? We sure
need to cleanup modules and it is time to re-look at the component organization. But what
do you want to really achieve by combining all the core into monolithic components. It took
an effort to cleanly separate functionality so they can evolve and can be improved independently.

>>> 
>>> 
>>> I am not suggesting to go back to the monolithic core which has all implementations
and interfaces bundle with together. What i am saying is have core interfaces all together,
this will give us clear module dependency graph. This is something like having one root dependency
graph instead of multiple roots. As a developer it is cumbersome and hard to deal with module
dependency issues. Having too fine gran core modules introduce wrong dependency graph eventually,
and it will prevent us to follow proper design patterns in our code base. If we avoid design
patterns, it will require more time to find bugs and maintain. This is what I am trying to
resolve. I have first hand experiences with current airavata code. If you see the current
module dependency graphs, then you will understand why I am making such noise to resolve this.
>>> 
>>> 
>>>> So why do we make our repository bulky with modules unless it doesn't provide
any considerable advantage. 
>>>> ​If we really w​ant to ​separate distribution bundle for each component
( apiServer , Orchestrator and Gfac ) let​'​s use different bin.xml file​s​ to do
it instead of using different modules. But reality is we only use ​all in one distribution.
>>> 
>>> Once the monitoring is fixed to use messaging, we really need to decouple the
component deployments. Yes there is a considerable advantage. Each components has different
quality of service requirements. A production platform has to load balance and scale horizontally.
And thats different for different component. The all in one bundle has 300+ jars, but API
server and orchestrator when independent will have around 50 or so jars. When I want to deploy
api server and orchestrator there is significant different in small light weight components
vs one monolithic core. 
>>> 
>>> Another problem is component evolution. Lets say there is a production deployment
running 1.1.4 version. Lets say the single job execution is stable enough and there is a 6
month focused effort on workflow. Say the master moves to 1.2.8 with all the changes to workflow
and only few to single application execution. We can more comfortably upgrade is they are
cleanly separated modules. But if it is one core and so many changes to it (even though technically
they are to different classes, the perception will remain), the upgrades will get behind.
>>> 
>>> Bottom line I am + 1 for cleaning up the modules. Past few years we have been
moving towards micro service architectures and your suggestions will reverse this back to
monolithic architecture. I am -1 for this change in direction. 
>>> 
>>> 
>>> By looking at your miniature component suggestion above, it has 30+ modules.
Do you think we really need this number of categorization? With my industrial experience,
I have seen ​number of modules in a project always increase with time. Hence​ If we start
30+ we will come to 40+ and then 50+​ and so on so forth​.  Why we make this complicated?

>>> 
>>> ​Thanks,
>>> Shameera.​
>>> 
>>>  
>>> 
>>> Suresh
>>> 
>>>> 
>>>> ​​Thanks,
>>>> ​Shameera.
>>>> ​
>>>> 
>>>> Suresh
>>>> 
>>>>> On May 29, 2015, at 11:29 AM, Suresh Marru <smarru@apache.org <mailto:smarru@apache.org>>
wrote:
>>>>> 
>>>>> + 1. 
>>>>> 
>>>>> I was planning to bring up this issue also. Probably it will not address
what you are raising, but here is a tree output from airavata labs code I was toying with
locally. I did not yet compare it with what you proposed, I will do so later today.
>>>>> 
>>>>> ├── airavata-api
>>>>> │   ├── airavata-api-interface-descriptions
>>>>> │   ├── airavata-api-java-stubs
>>>>> │   ├── airavata-api-server
>>>>> │   ├── airavata-data-models
>>>>> │   ├── api-security-manager
>>>>> ├── clients
>>>>> │   ├── airavata-client-cpp-sdk
>>>>> │   ├── airavata-client-java-sdk
>>>>> │   ├── airavata-client-php-sdk
>>>>> │   ├── airavata-client-python-sdk
>>>>> │   ├── airavata-sample-examples
>>>>> │   └── airavata-xbaya-gui
>>>>> ├── components
>>>>> │   ├── commons
>>>>> │   ├── component-interface-descriptions
>>>>> │   ├── component-services
>>>>> │   │   ├── credential-store-service
>>>>> │   │   ├── orchestrator-service
>>>>> │   │   ├── task-executor-service
>>>>> │   │   └── workflow-interpreter-service
>>>>> │   ├── component-clients
>>>>> │   │   ├── credential-store-client
>>>>> │   │   ├── orchestrator-client
>>>>> │   │   ├── task-executor-client
>>>>> │   │   ├── workflow-interpreter-client
>>>>> │   │   └── messaging
>>>>> │   ├── task-adaptors
>>>>> │   │   ├── compute
>>>>> │   │   └── data-movement
>>>>> │   ├── registry
>>>>> │   │   ├── app-catalog
>>>>> │   │   ├── experiment-catalog
>>>>> │   │   └── resource-catalog
>>>>> │   └── workflow-interpreter
>>>>> ├── distribution
>>>>> ├── integration-tests
>>>>> 
>>>>> 
>>>>> 
>>>>>> On May 29, 2015, at 10:15 AM, Shameera Rathnayaka <shameera@apache.org
<mailto:shameera@apache.org>> wrote:
>>>>>> 
>>>>>> Hi Devs, 
>>>>>> 
>>>>>> As we are using different modules to package different type of functionalities,
which will help us to maintain loosely couple codes. Now the project has 49 leaf module (
one to hit half century :) ). If we allow project to grow this way, having too fine grain
modules will be huge headache in future. IMO we should clean this ASAP before it become really
mess. Actually we half way there, I experienced cyclic dependency issues when I was writing
workflow implementation and email monitoring. Please see the modules in current repo below.

>>>>>> 
>>>>>> <module-name> ( <num of child modules> )
>>>>>> 
>>>>>> modules  ( 43 )
>>>>>>      app-catalog ( 2 )
>>>>>>      commons ( 1 )
>>>>>>      configurations ( 2 )
>>>>>>      credential-store ( 3 )
>>>>>>      distribution ( 8 )
>>>>>>      gfac ( 10 )
>>>>>>      integration test ( 1 )
>>>>>>      messaging ( 2 )
>>>>>>      orchestrator ( 3 )
>>>>>>      registry ( 3 )
>>>>>>      security ( 1 )
>>>>>>      server ( 1 )
>>>>>>      test-suit ( 1 )
>>>>>>      workflow ( 1 )
>>>>>>      workflow-modal ( 3 )
>>>>>>      xbaya ( 1 ) 
>>>>>> airavata-api ( 5 )
>>>>>> tools ( 1 ) 
>>>>>> 
>>>>>> Most of the current modules have interfaces and implementations together,
but this violate our main goal which reduce inter module dependencies. Following is what I
am suggesting, WDYS?
>>>>>> 
>>>>>> core { has all core interfaces and basic classes of gfac-core , orchestrator-core
, message-core , monitor core, registry core, workflow-core}
>>>>>> service - all thrift services and service handlers 
>>>>>> orchestrator - orchestrator specific classes
>>>>>> gfac 
>>>>>>      SSH  
>>>>>>      BES
>>>>>>      Local
>>>>>> message - amqp implemention 
>>>>>> distribution 
>>>>>>      XBaya
>>>>>>      server - { use different mode input to start server as orchestrator
, Gfac or/and api-server }
>>>>>> commons
>>>>>> registry
>>>>>> app-catalog
>>>>>> security
>>>>>> Workflow
>>>>>> XBaya-gui
>>>>>> Integration-test 
>>>>>> 
>>>>>> Thanks, 
>>>>>> Shameera.
>>>>> 
>>>> 
>>>> 
>>>> 
>>>> 
>>>> -- 
>>>> Best Regards,
>>>> Shameera Rathnayaka.
>>>> 
>>>> email: shameera AT apache.org <http://apache.org/> , shameerainfo AT
gmail.com <http://gmail.com/>
>>>> Blog : http://shameerarathnayaka.blogspot.com/ <http://shameerarathnayaka.blogspot.com/>
>>> 
>>> 
>>> 
>>> -- 
>>> Best Regards,
>>> Shameera Rathnayaka.
>>> 
>>> email: shameera AT apache.org <http://apache.org/> , shameerainfo AT gmail.com
<http://gmail.com/>
>>> Blog : http://shameerarathnayaka.blogspot.com/ <http://shameerarathnayaka.blogspot.com/>
>> 
>> 
>> 
>> -- 
>> Best Regards,
>> Shameera Rathnayaka.
>> 
>> email: shameera AT apache.org <http://apache.org/> , shameerainfo AT gmail.com
<http://gmail.com/>
>> Blog : http://shameerarathnayaka.blogspot.com/ <http://shameerarathnayaka.blogspot.com/>
>> 
>> 
>> 
>> 
>> -- 
>> Best Regards,
>> Shameera Rathnayaka.
>> 
>> email: shameera AT apache.org <http://apache.org/> , shameerainfo AT gmail.com
<http://gmail.com/>
>> Blog : http://shameerarathnayaka.blogspot.com/ <http://shameerarathnayaka.blogspot.com/>
>> 
>> 
>> 
>> -- 
>> Best Regards,
>> Shameera Rathnayaka.
>> 
>> email: shameera AT apache.org <http://apache.org/> , shameerainfo AT gmail.com
<http://gmail.com/>
>> Blog : http://shameerarathnayaka.blogspot.com/ <http://shameerarathnayaka.blogspot.com/>
> 
> 
> 
> 
> -- 
> Best Regards,
> Shameera Rathnayaka.
> 
> email: shameera AT apache.org <http://apache.org/> , shameerainfo AT gmail.com
<http://gmail.com/>
> Blog : http://shameerarathnayaka.blogspot.com/ <http://shameerarathnayaka.blogspot.com/>


Mime
View raw message