#acumos-meeting: architecture committee
Meeting started by farheen_att at 15:03:51 UTC
(full logs).
Meeting summary
- agenda (farheen_att, 15:03:56)
- Download user experience requested by Philippe
D. (farheen_att,
15:08:43)
- No portal presence. (farheen_att,
15:09:04)
- If anyone wants to discuss topics contact
Manoop (farheen_att,
15:09:21)
- Nat - Vasu on the Clio point release who are
all still pending? (farheen_att,
15:09:35)
- Deployment, DS, and ML Workbench are
outstanding (farheen_att,
15:10:05)
- Justin - Deployment model usage tracking.
Having issues but there is no one is assigned to deployment client
anymore. (farheen_att,
15:10:34)
- Santosh has committer access. Vineet also has
committer access. (farheen_att,
15:11:45)
- Vineet can help with merge requests.
(farheen_att,
15:12:05)
- in the deployment client (farheen_att,
15:12:19)
- Manoop - Clio maintenance release is our top
priority. (farheen_att,
15:13:00)
- Sayee - Sprint 1 new feature that the DS/ML Workbench team is working on. (farheen_att, 15:13:38)
- Sayee - YML workbench enhancement. We are
adding deployment of a model from ML Workbench. (farheen_att,
15:14:47)
- you can create and associate notebook and nifi
pipeline to a project. Model has to be onboarded on Acumos to be
associated. This blue Deploy icon added. (farheen_att,
15:15:34)
- this is similar to the portal version. You can
give a predictor name. Once complete a deployment in process. A
jenkins job is created where the model is being deployed.
(farheen_att,
15:16:13)
- do you differentiate between models and
predictors? (farheen_att,
15:16:28)
- this is for trained model. (farheen_att,
15:16:38)
- Sayee - predictor is a running model. If you
want to scale it to one or two pods. Currently that is not being
tested. We want to see where it is being scaled to cutomers.
(farheen_att,
15:17:31)
- You can check the deployment in the UI . it is
a jenkins job. (farheen_att,
15:18:10)
- you can redeploy the model. No restrictions on
the number of times you can deploy a model. (farheen_att,
15:18:51)
- if i have my own jenkins job and deploy in
azure I can link it here. We want to provide but we will see if it
is needed . This is a way a data admin can manage their model
across many clusters and troubleshoot as needed. (farheen_att,
15:19:55)
- manoop - is this feature represented as model
deployment from ML workbench? (farheen_att,
15:20:18)
- yes list of predictors that can be deployed to
multiple clusters. (farheen_att,
15:20:38)
- Nat - this has a resource constraint. This is
a huge gap and going to be an issue. (farheen_att,
15:21:21)
- Manoop - deploying the model. Do you have a
goal for which environment? (farheen_att,
15:22:08)
- k8 (farheen_att,
15:22:12)
- Justin got it working. (farheen_att,
15:22:35)
- Ken - having trouble getting deployment to
work. (farheen_att,
15:22:58)
- Justin offered to support Ken. (farheen_att,
15:23:41)
- Ken - Recommending a code review. (farheen_att,
15:24:03)
- Sayee - It needs to be scrutinized. It needs a
lot of work. Sometimes demos don't work. (farheen_att,
15:24:34)
- Reuben - We need to start planning to support
re-training. License and dist. models is I can plug my data. We
need to extend this to training and evaluation. (farheen_att,
15:25:23)
- Sayee - I couldn't get Wenting's time.
(farheen_att,
15:25:37)
- model builder can re-train. Reuben will talk
to Wenting and Sayee about the re-training aspect. Sayee will have
a separate store for data sets and training. (farheen_att,
15:26:58)
- Reuben wants to bring in IBM face detection.
We want to be able to run it here (demo of local ML
workbenchd) (farheen_att,
15:27:45)
- I can deploy at run time. Another is to run in
one pod. (farheen_att,
15:28:15)
- When you deploy a model as a predictor and then
evaluate the model. (farheen_att,
15:28:52)
- ACTION: Sayee bring
one single draft. (farheen_att,
15:29:09)
- Reuben cautions putting a single draft alone.
ML flow deals with a model registry trained with different data sets
and allows selection of what should be deployed where. (farheen_att,
15:30:35)
- Reuben - we want to look at the logic of other
projects before designing our own. (farheen_att,
15:31:29)
- one way of increasing the value of acumos is to
make sure that we have inter-operability. IF we an make this
inter-operability a requirement of all the projects, not only do we
get the synergy but we also get the benefit of bringing in the
expertise such as ML workflow team who understand model version
management. We will get more (farheen_att,
15:32:56)
- https://wiki.lfai.foundation/display/DL/ML+Workflow+Committee?preview=/10518537/18481275/ML%20stack_v10.pptx
(farheen_att,
15:33:21)
- Nat sharing the LF ecosystem. (farheen_att,
15:33:38)
- can we learn from these sub-projects?
(farheen_att,
15:35:46)
- Nat - yes, i know the person who already has a
deployment. They focus on sprk and python. (farheen_att,
15:36:19)
- Sayee - we can work with these teams and
leverage already developed resources. (farheen_att,
15:36:56)
- Reuben - Angel is being used predominantly with
Chinese companies. (farheen_att,
15:37:34)
- Manoop is someone at the LF level reaching out
to these projects? (farheen_att,
15:37:57)
- Nat - I can introduce the people I know with
Guy and Philippe. This is the TAG committee (LFAI). Angel has
graduated. ONNX also. (farheen_att,
15:39:07)
- Manoop can we review with Angel? (farheen_att,
15:39:21)
- you clearly see the overlap over the project.
Should we bring to discuss with PTLs. Rather than general
meetings (farheen_att,
15:40:02)
- Reuben - we should know which projects are
doing what. We should have an easy way to access these
resources. (farheen_att,
15:40:36)
- Guy and Philippe have discussed deploying a
model. As far as ML workbench and serving pipeline. (farheen_att,
15:41:19)
- ACTION: Nat introduce
ml workbench with the Angel project. (farheen_att,
15:41:45)
- Sayee bring them to this call. (farheen_att,
15:42:05)
- Manoop we don't want a marketing call we want
to be specific. (farheen_att,
15:42:30)
- how can we leverage the feature from open
source. (farheen_att,
15:42:42)
- ACTION: Reuben will
bring information about his contacts on the community meeting next
Wednesday. (farheen_att,
15:50:39)
- Philippe - UX model deployment (farheen_att, 15:51:08)
- Philippe created a user story about downloading
multiple artifacts. Today you have to download artifacts one at a
time. ACUMOS-3679 (farheen_att,
15:52:16)
- There is enough time to move this user story
into an epic. We would like to add docker image. Today when you
download the image it takes time. (farheen_att,
15:53:57)
- https://wiki.acumos.org/display/MOB/Enhance+UX+when+downloading+artifacts
(farheen_att,
15:54:00)
- should we work on it in this release or save it
for the next release? (farheen_att,
15:54:51)
- Manoop - remove artifacts... was addressed in
the past releases. Manoop will check. The requirement was
addressed. Some are hidden that are internally used. Manoop will
check. About selecting artifacts. What is the motivation?
Downloading one zip file is easier than multiple clicks and time.
Easier to select the artifacts and download (farheen_att,
15:56:52)
Meeting ended at 16:07:33 UTC
(full logs).
Action items
- Sayee bring one single draft.
- Nat introduce ml workbench with the Angel project.
- Reuben will bring information about his contacts on the community meeting next Wednesday.
People present (lines said)
- farheen_att (67)
- collabot` (4)
- farheen (1)
Generated by MeetBot 0.1.4.