airavata-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Suresh Marru <sma...@apache.org>
Subject Re: Suggessions for app-catalog design
Date Fri, 25 Jul 2014 13:08:05 GMT
Lahiru,

I was looking through the API and here are some more thoughts on this:

AppCatalog is designed treating Airavata as a shared resource among a common set of users.
For instance, you go onto a supercomputer you see a list of applications. You leave to users
how to use them. Similar is our deployment vs interface. A typical scientific computing application
will have dozens of parameters, any given interface will only expose a subset of them or all
them. But it will be bad to enforce for every deployment there is an interface. Similar to
when a application is deployed, you do no anticipate how it will be used, we do not want to
tightly bound deployments and interfaces. 

I think you are critiquing in the right direction but your perspective is a single user single
application (one deployment set per interface). I agree, the design did not take into consideration
how to do the hello world easily. So step back, think about how the design should be when
you get dozen’s of applications and a hand full for interfaces for one deployment and so
forth. If you still think the design is complex to handle such scenarios, lets revisit. I
am not talking about corner cases, I am talking about real-world production usage vs ease
of use for hello world developer testing. A good way for you to understand is, login to any
HPC machine and see how the system is described (once on a semi-static web page (compute resource
description), uniquely identified modules (deployment description) and leave the interfaces
to users (interface descriptions). 

Suresh

On Jul 25, 2014, at 8:37 AM, Suresh Marru <smarru@apache.org> wrote:

> Lahiru,
> 
> Glad that you took time to sink in the design and provide suggestions. I see that all
of them are hitting the similar issue, so let me respond a summary one:
> 
> The suggestions you have are not really the design but usability and ease of debugging.
The Id’s and modules are there so we disambiguate usage. In the previous approaches (blame
me) I exactly always advocated for what you ask below and we have seen over time, serious
production usage suffers from it. For instance in a ParamChem workflow where Gaussian is wrapped
in 10 different ways, its very error prone. We should not push the burden onto users to do
the right thing, but the system should enforce it. As an example, if google docs uses names,
imagine the commonly used terms like “meeting notes” or resume. There are ID’s for a
reason. But as a user you never deal with ID’s. But if you use google docs API, you very
well needs to work with ID’s.
> 
> Same here, I like the other suggestion you made in a different context, lets develop
tooling to deal with debugging. And if there are UI’s to describe documents, you never work
with ID’s. Take RegisterSampleApplications as an example, I do not think you want simper
than that. There will be additional steps. But app catalog will be register once use 1000’s
of times. For this, I think its ok to spend extra 5 minutes to register an application to
make sure users use exactly what they intend to instead of some wild card guessing. 
> 
> Suresh
> 
> On Jul 25, 2014, at 3:43 AM, Lahiru Gunathilake <glahiru@gmail.com> wrote:
> 
>> Hi All,
>> 
>> I think app-catalog design is a well-thought comprehensive design and I would like
to propose following suggestions. Please correct me if I am wrong.
>> 
>> I think we have to minimize the complexity in app-catalog model if some of the components
are not really necessary in 99% of our scenarios. If the corner cases are complex we shouldn't
make things complex for our most frequent scenarios.
>> 
>> Example: I think we do not need a layer of module and I think its not worth api dealing
with whole module layer. We can simply achieve the module layer information by giving some
meaningful name for the application Deployment document.
>> Ex:Amber-1.4,Amber-1.3
>> 
>> I think we should remove all the Id based query because its very difficult to program
against Ids, IMHO exposing the Ids in the API is a bad design. If someone try to create same
named document in same level again we can simply give an error without exposing Ids to the
API. To achieve this we should not allow users to create Application deployment documents
before they create its Application interface and parsing along the application interface name.
>> 
>> And in practice allowing users to give same names is a bad design (with Id model
we allow users to give same name) because ultimately what users will see is set of names not
the Ids, so I think we have to enforce that in the API. I think good API should always lead
users to do minimum errors.
>> 
>> For me this is more like a tree structure and we do not have scenario of random access.
If we know all the names we should be able to get it in a single call with all the objects
in it as we do in a normal tree where data will be stored in the leaf nodes. Currently we
do lot of traversing among different documents, these methods will be useful during scheduling
but in the case of using Experiment configuration where we specify everything precisely we
should be able to get it right away.
>> 
>> Regards
>> Lahiru
>> -- 
>> System Analyst Programmer
>> PTI Lab
>> Indiana University
> 


Mime
View raw message