#acumos-meeting: Architecture Committtee
Meeting started by farheen_att at 14:02:59 UTC
(full logs).
Meeting summary
- Licensing (farheen_att, 14:04:45)
- Working on things that are difficult to demo
because we still need the front end pieces. We created the api to
interact with the portal. We have handed over what portal can
deliver in their parts. LUM is what we demonstrated on the
Community Call yesterday. It isn't a great user demo but we will
review the enhancements that have been (farheen_att,
14:07:10)
- cut short demo for on-boarding team offshore
who have to leave early. (farheen_att,
14:07:55)
- On-boarding Priya, Guy Spark Java Spark on-boarding (farheen_att, 14:08:37)
- https://wiki.acumos.org/display/MOB/Acumos+spark
(farheen_att,
14:11:38)
- Priya today we are assuming that you are using
the Spark ML library. (farheen_att,
14:13:01)
- with respect to spark lib. Does the lib have
to be packaged as a k8 container? (farheen_att,
14:13:29)
- yes (farheen_att,
14:13:32)
- current ms gen method we put everything inside
the container. (farheen_att,
14:14:11)
- so every container will contain the spark
lib? (farheen_att,
14:14:22)
- it will be packaged as part of the jar
(farheen_att,
14:14:50)
- the model docker image will be independent of
the docker image. (farheen_att,
14:15:03)
- so I can have a spark cluster outside for my
model to run? (farheen_att,
14:15:16)
- correct (farheen_att,
14:15:20)
- regarding mode of deployment there are two ways
client spark and cluster mode. We are considering the standalone
spark for this version of implementation. The apis used are
different based on which manager is being used as part of the
deployment. (farheen_att,
14:16:57)
- we don't really know which one should be used
and we wanted to verfiy based on configuration. YML or K8. Today
we are focusing on stand alone deployment of spark. (farheen_att,
14:18:01)
- YARN not YML* (farheen_att,
14:19:02)
- being able to onboard spark models. But for
Clio we are limiting it to java spark based on boarding.
(farheen_att,
14:19:35)
- Java spark based model will remain as a docker
image where a modeler will deploy? (farheen_att,
14:20:06)
- it is the same process and ms generation is the
same as any other type of model. (farheen_att,
14:20:25)
- how does it differ? (farheen_att,
14:20:52)
- the apis used are different. Whatever apis
used to invoke the instances are different. (farheen_att,
14:21:11)
- there are two types of libraries available
today. and Mlib and a spark MLIB. Available from spark 2.3
onwards. (farheen_att,
14:22:41)
- 2 ways to create docker image. 1. python
based models today we are deploying all the keras dependencies
contained. But in the case of spark it is going to be outside the
docker image. The docker won't have the docker engine only the
spark libs to talk to spark instance. (farheen_att,
14:24:13)
- important to have a clear input output. Spark
ML models may be reading from a hadoop cluster or a different data
source as of today. Spark ML model will define a clear input and
output per protobuf. (farheen_att,
14:25:22)
- whatever required has to be specified.
(farheen_att,
14:25:38)
- can we not have a single container for model is
consuming resoures and performance will hit. (farheen_att,
14:26:16)
- agree the spark engine will be
independent. (farheen_att,
14:26:46)
- don't be tightly coupled with protobuf
(farheen_att,
14:27:11)
- long term we should involve. JSON and Parkay
file support should be there. If you look at customers they are
using other file formats. My recommendation is to keep it
generic. (farheen_att,
14:28:28)
- We want to support both protobuf and
JSON. (farheen_att,
14:28:57)
- for now we will support only protobuf but
eventually we will support other types. v.1 protobuf and v.2 JSON
and then others. (farheen_att,
14:30:07)
- these are the apis that model runner will
support. Invoking the model will have a sequence when you invoke
the model. The spark config that you have can be different than
what is in the deployed format. Accepting and submitting the config
at run time. Other models are self contained but in term of spark
since the engine is outside the (farheen_att,
14:32:16)
- Data source and output are both coming from the
same response. It is independent of what fs to use. These are
consideration to invoke the model. Docker image will be submitting
the job but the input output will be outside of
request/response. (farheen_att,
14:33:21)
- this is the view we thought that kafka and
spark to go together. We will need a pipeline for kafka to support
spark. (farheen_att,
14:34:20)
- java client is impacted by this change.
(farheen_att,
14:35:01)
- Licensing resume. (farheen_att, 14:35:18)
- Michelle, Justin (farheen_att,
14:35:28)
- We've completed the apis for license profile.
We can now review the apis for this sprint. (farheen_att,
14:36:15)
- we're starting with 3336 which is an api
manager. In this demo we are using junit test. We are using
different json files checked in as resources. A license validator
and the warnings. Once you upload your license profile. You will
be able to call this api and get feedback vs. down stream problems.
This is the back end of reading the (farheen_att,
14:39:32)
- as far as different variables missing. This is
the validation of the structure of the document and the unit test.
When it does pass as the license json. When you call this from the
api perspective we wrap it under the profile class. 3 inputs. Give
it a string, input stream, or json node if you've done the parsing
of the json for (farheen_att,
14:41:17)
- questions? (farheen_att,
14:41:24)
- Michelle we showed the schema validation api to
make sure it is captured. (farheen_att,
14:41:47)
- Please open the response? So you will have
multiple messages are there any status? We will check the field if
it's a success we won't do anything. (farheen_att,
14:43:05)
- tausif - I thought we will be calling for
onboarding but there is a different flow for just the license file
so you want portal to call when there is a license. (farheen_att,
14:45:31)
- yes we want to catch issues up front.
(farheen_att,
14:45:41)
- will there be an option to edit online or do
they have to reload the file? (farheen_att,
14:45:55)
- yes, they have to upload againg. The editor
uses the schema. Upload without using editor. The editor is baked
into (farheen_att,
14:46:55)
- is the format json? (farheen_att,
14:47:02)
- yes (farheen_att,
14:47:07)
- stored as blob in? (farheen_att,
14:47:16)
- nexus (farheen_att,
14:47:20)
- can it be hacked? (farheen_att,
14:47:28)
- only the system component. User does not
access nexus (farheen_att,
14:47:53)
- to create a custom license requirement you have
to ready readthedocs. (farheen_att,
14:51:32)
- We made sure that all of our java apis are
updated to java 11. We were able to update. (farheen_att,
14:52:12)
- you use the docker base images? (farheen_att,
14:52:20)
- we help the portal team update to open j9, It
wasn't licensing specific. There were issues. There are two libs
that licensing manager client library and we helped them with the
LUM code. In order to interface with LUM there is a code generator
that connects with LUM. LMCL is consuming that . I wanted to show
you that. It wasn't a jira. (farheen_att,
14:58:17)
- demonstrated how to register the software with
the LUM with the solution id. Reaches out to CDS and nexus to gather
the information based on the solution id. (farheen_att,
14:59:07)
- We're trying to simplify how clients interface
with LUM. Construct swid tag (revision id) so the mapping is
consisitent in the LCML. (farheen_att,
15:00:00)
- demonstrated the library working as a
bonus (farheen_att,
15:00:13)
- ACTION: Murali
co-ordinate a full end to end demo with portal. In your scrum or
their scrum have a joined session. (farheen_att,
15:05:47)
- ACTION: Manoop add
license review to the next call. (farheen_att,
15:06:39)
Meeting ended at 15:06:45 UTC
(full logs).
Action items
- Murali co-ordinate a full end to end demo with portal. In your scrum or their scrum have a joined session.
- Manoop add license review to the next call.
People present (lines said)
- farheen_att (69)
- collabot` (3)
Generated by MeetBot 0.1.4.