[ 
https://issues.apache.org/jira/browse/SINGA-399?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16666289#comment-16666289
 ] 

Ngin Yun Chuan edited comment on SINGA-399 at 10/28/18 5:09 AM:
----------------------------------------------------------------

Regarding the error encountered in `rafiki-3.PNG`, the error signifies that the 
`predictor` Flask HTTP server is not up yet when the method 
`make_predictions(...)` within `examples/scripts/client-usage.py` is run, which 
sends queries to `127.0.0.1:30000`. From the images, I can think of 2 scenarios:

1) You edited `rafiki.predictor`'s code or its dependencies in a way that 
causes a runtime error during the deployment of the inference job. Check 
`/var/log/rafiki` that contains all components' logs for any error messages.

2) The `predictor` Flask HTTP server took more than 20s (the delay I added in 
`client-usage.py` before the script sends queries) to setup and start accepting 
queries. This is possible as `wait_until_inference_job_is_running(...)` doesn't 
actually yet ensure that the `predictor`'s HTTP 
server is up before returning. To test this, you can try sending predictions 
with `curl` after a longer delay i.e. follow the instructions on 
https://nginyc.github.io/rafiki/docs/docs/user/quickstart-app-users.html

Either way, I will be adding sufficient documentation and modify the code to 
better account for both scenarios. Do let me know if you have any suggestions 
or figured out any sources of errors.


was (Author: nginyc):
Regarding the error encountered in `rafiki-3.PNG`, the error signifies that the 
`predictor` Flask HTTP server is not up yet when the method 
`make_predictions(...)` within `examples/scripts/client-usage.py` is run, which 
sends queries to `127.0.0.1:30000`. From the images, I can think of 2 scenarios:

1) You edited `rafiki.predictor`'s code or its dependencies in a way that 
causes a runtime error during the deployment of the inference job. Check 
`/var/log/rafiki` that contains all components' logs for any error messages.

2) The `predictor` Flask HTTP server took more than 20s (the delay I added in 
`client-usage.py` before the script sends queries) to setup and start accepting 
queries. This is possible as `wait_until_inference_job_is_running(...)` doesn't 
actually yet ensure that the `predictor`'s HTTP server is up before returning. 

Either way, I will be adding sufficient documentation and modify the code to 
better account for both scenarios. Do let me know if you have any suggestions 
or figured out any sources of errors.

> Rafiki cannot test rebuilt image
> --------------------------------
>
>                 Key: SINGA-399
>                 URL: https://issues.apache.org/jira/browse/SINGA-399
>             Project: Singa
>          Issue Type: Bug
>            Reporter: Zhu Lei
>            Priority: Major
>         Attachments: rafiki-1.PNG, rafiki-2.PNG, rafiki-3.PNG, rafiki-4.PNG
>
>
> After downloading the newest rafiki code, at commit 
> 7b3b04e15c62233e515c4d82051cd5dfb799215f, with comments "Add more error 
> handling to notify user of invalid train job; compact exceptions", I ran 
> "bash ./scripts/build_images.sh" to build the new admin, advisor, predictor 
> and worker images. I got the images shown in attached image 'rafiki-1.PNG'. 
> Then I run "bash ./script/start.sh" to build the containers as shown in the 
> attached image 'rafiki-2.PNG'. Finally when I ran the client-usage.py 
> example. I got the error in attached image 'rafiki-3.PNG'.
> And I find very surprising that the images of admin, advisor, predictor and 
> worker I built just now, become some images built weeks ago, shown in 
> attached image 'rafiki-4.PNG'. Could you kindly provide me some explanations 
> on why this happens? I really do not understand why this happened.
> And finally, when I run "bash ./script/stop.sh" and leave the swarm and 
> repeat my previous procedure again, now there is no errors. The only thing 
> difference between the two runs I think is only the images are different. So 
> the current code of rafiki does not support newly build images, that is my 
> speculation.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to