19 May 2022 |
Ian Miller | Hi all, question on KServe and KFServing. We deployed KServe/KFServing (depending on the cluster + KF version) using the installation yaml for deploying with kubeflow. When we deploy inference services, the inference service gets a URL that doesn't work through the istio gateway. I see reference in the configmap (and therefore virtual services) referencing a kfserving (or kserve respectively) gateway which doesn't seem to exist. Additionally, when I switch that config to use cluster-local-gateway instead which does seem to exist, it still doesn't route the calls through to my deployed inference service. I am able to invoke my inference service via the route listed in the corresponding ksvc however. Wondering if anyone knows why I might not be able to invoke through the istio gateway. My kserve deployment is not altered other than as described above. The main issue here is the Model UI displays the route listed on the inference service so having trouble directing users how to interact with their deployed model. | 01:51:32 |
Dan Sun | Yes you should be able to change via knative networking configuration | 08:14:38 |
Dan Sun | Ian Miller You are probably on an old version of Kubeflow? | 08:49:37 |
Ian Miller | Done the bulk of this testing on KF 1.4 with KFServing 0.6.1. Saw the same configuration on KF 1.5 + KServe 0.7 though, but admittedly have tried less to get it to work there so far as we're still upgrading our primary clusters. | 13:58:34 |
_slack_kubeflow_U9UFLSBM4 | I'm working on a slightly strange deployment. KServe-raw on openshift without istio. I am able to serve models and get predictions no problem. I can hit /metrics on the kserve-controller-manager-metrics-service and I see metrics, but they all seem to be only for the controller itself rather than anything about the models being served (ie: number of predictions, time per prediction, etc). Is there something else I need to enable or somewhere else I need to be looking to get the metrics about the InferenceService itself? | 15:29:53 |
| @californiatok:matrix.org joined the room. | 22:02:59 |
20 May 2022 |
| 레몬버터구이 joined the room. | 07:14:28 |
| @californiatok:matrix.org left the room. | 09:34:04 |
Dan Sun | croberts that depends on the model server you are running, most of the model servers are producing metrics. | 19:25:09 |
Shri Javadekar | I'm sure curious about this too.
One way I was thinking about this was to write my own logger that exported metrics such as # of predictions, time, prediction confidence score, etc.
Dan Sun: Is there any specific model server that already does this? Triton? | 21:50:13 |
Dan Sun | if you are running knative then queue proxy already produces a set of standard metrics for request count, latency etc, confidence score is a bit tricky as KServe has no idea about the model accuracy | 22:11:08 |
Dan Sun | Triton itself does produce a set of metrics too
https://github.com/triton-inference-server/server/blob/main/docs/metrics.md | 22:11:39 |
| _slack_kubeflow_U03FZJA1GBZ joined the room. | 23:44:39 |
21 May 2022 |
| _slack_kubeflow_U03GF0NC9QS joined the room. | 14:12:13 |
Dan Sun | If anyone from Asia interested in doing a monthly KServe community call in Asia friendly time, please help give a thumbs up here and DM me your email address, thanks! | 16:27:49 |
| Amalesh joined the room. | 16:45:01 |
22 May 2022 |
| mobin nikkhesal joined the room. | 11:46:25 |
| mobin nikkhesal changed their display name from _slack_kubeflow_U03GUM55T0R to mobin nikkhesal. | 11:58:34 |
| mobin nikkhesal set a profile picture. | 11:58:36 |
Shri Javadekar | Is the current schedule of KServe community calls published somewhere? | 15:25:15 |
| mobin nikkhesal changed their profile picture. | 17:00:14 |
23 May 2022 |
zorba(손주형) | what will be done in community call? | 02:50:04 |
| Lize Cai joined the room. | 08:25:53 |
| Abhishek Sharma joined the room. | 11:34:49 |
Jan Buijnsters | https://kserve.github.io/website/0.8/blog/articles/2022-02-18-KServe-0.8-release/#join-the-community
Biweekly Wednesday 9am PST | 14:22:45 |
| Thomas Ounnas joined the room. | 16:11:39 |
24 May 2022 |
| Benjamin Tan changed their profile picture. | 15:25:17 |
_slack_kubeflow_U9UFLSBM4 | Trying to debug a kserve-raw deployment. Can anyone think of reasons why the storage-initializer init container is not getting attached to the tensorflow pod? | 15:57:02 |
Dan Sun | which namespace it is deployed to? it does not work in the kserve control plane namespace | 16:07:24 |
_slack_kubeflow_U9UFLSBM4 | The person trying it originally put it in the kserve namespace, but later tried in another namespace and had the same result | 16:22:01 |