Sender | Message | Time |
---|---|---|
13 Oct 2021 | ||
Nick | In reply to@_slack_kubeflow_UFVUV2UFP:matrix.orgDan Sun actually I think it’s 2.30pm tomorrow https://kccncna2021.sched.com/event/lV2s/serving-machine-learning-models-at-scale-using-kserve-animesh-singh-ibm?iframe=no | 18:11:35 |
Peilun Li joined the room. | 18:44:29 | |
Peilun Li | So I guess we have to: 1. First upgrade our manifest from 0.5 to 0.6 (without changing existing services -- still on 0.5) 2. Then install 0.7 manifest to the cluster 3. Meanwhile for new service deployment we can deploy with 0.7 directly, while the old 0.5 deployments are still controlled by 0.6. 4. Finally use the migration job to migrate existing 0.5 services (controlled by 0.6) to 0.7's api group and controller. Would that be a working path? | 18:44:29 |
Peilun Li | In reply toundefined(edited) ... controller. Would ... => ... controller. 5. Delete the 0.6 manifest afterwards. Would ... | 18:49:12 |
Dan Sun | In reply to@_slack_kubeflow_U0127AUTPMH:matrix.orgoops sorry my bad | 19:12:04 |
Dan Sun | In reply toundefined(edited) ... KServe today at ... => ... KServe tomorrow at ... | 19:12:12 |
Dan Sun | In reply toundefined(edited) ... tomorrow at ... => ... tomorrow 10/14 at ... | 19:12:34 |
14 Oct 2021 | ||
Riley Green | Download image.png | 02:35:49 |
Riley Green | Hi, I have installed Kubeflow 1.4 from manifests and I tried to deploy the sklearn-iris predictor sample and I am getting the following error: Any ideas on a way forward? Many thanks 🙏 | 02:35:49 |
Z Omari joined the room. | 03:21:00 | |
Dan Sun | Suresh Nakkiran I do not see any reason why we can't go from 0.5.0 to 0.7.0 as long as the version supports v1beta1 API and you are still using the same version of istiod/knative | 05:02:24 |
Dan Sun | actually to be precise if your istio/knative version meets the min requirement of kserve 0.7 then the migration path from 0.5 to 0.7 should work in my mind, but we will test out to make sure | 05:14:18 |
Dan Sun | Peilun Li what's your istio/knative version used along side with KFS 0.5 | 05:15:35 |
Dan Sun | In reply toundefined(edited) ... KFS 0.5 => ... KFS 0.5? | 05:15:39 |
Suresh Nakkiran | Dan Sun In quick install we have istio 1.6.2.. | 06:21:41 |
Dan Sun | we should have istio 1.9 there | 06:37:30 |
iamlovingit | In reply to@_slack_kubeflow_U02HTJ9GLKT:matrix.orgRiley Green Hi, can you show kfserving logs here? | 08:02:04 |
Peilun Li | We are currently on K8S 1.19, Istio 1.8, Knative 0.17 -- Definitely some work on our end to upgrade to Istio 1.9 and Knative 0.23+ before we upgrade kserve to 0.6/0.7 | 16:42:16 |
15 Oct 2021 | ||
Yulong joined the room. | 02:42:51 | |
Yulong | Hi, guys. I am using kfserving v0.6.0 on kubeflow v1.3.0. I was trying to use replace or patch function to apply a new version of TensorFlow model, but it it seems that after running the code, the InferenceService and pod didn’t changed.
predictor_2 = V1beta1PredictorSpec(tensorflow=V1beta1TFServingSpec(storage_uri=storage_uri, image=image), min_replicas=1, max_replicas=3, canary_traffic_percent=50, ) isvc_2 = V1beta1InferenceService(api_version="serving.kubeflow.org/v1beta1", kind=constants.KFSERVING_KIND, metadata=client.V1ObjectMeta(name=predictor_name, namespace=namespace), spec=V1beta1InferenceServiceSpec(predictor=predictor_2)) kfs.replace(name=predictor_name, inferenceservice=isvc_2, namespace=namespace) (base) jovyan@tensorboard-normal-0:~$ k get InferenceService NAME URL READY PREV LATEST PREVROLLEDOUTREVISION LATESTREADYREVISION AGE demo http://demo.test-ns.example.com True 100 demo-predictor-default-5sn57 8m59sWhat should I do to make it right? | 02:52:37 |
Yulong | (edited) Hi, guys. I am ... => Hi, guys. I am ... | 02:52:50 |
Yulong | By the way, if I use different storage_uri , it will replace or patch successfully, but we are using TFX. Pusher, a component of TFX, saves trained-model under a folder, e.g., ./models/1, /models/2, … we need to specify the storage_uri to the ./models level for loading saved model of TF. In this case, kfserving won’t do replace or patch . Does any possible way to deal with it? | 06:40:20 |
Dan Sun | In reply to@_slack_kubeflow_U0253V66WCT:matrix.orgYulong that's by design, if the storage uri is not changed then nothing is deployed as we want to make storage immutable. But I think in this case you can add an annotation to indicate the version change then InferenceService will redeploy and load the correct version | 13:02:29 |
Dan Sun | In reply to@_slack_kubeflow_UFVUV2UFP:matrix.orgAny change should be surfaced on inference service yaml which is the whole point of declarative deployment | 13:03:40 |
Vedant Padwal joined the room. | 13:49:25 | |
Vedant Padwal changed their display name from jordan sumitomo to Vedant Padwal. | 18:55:44 | |
Vedant Padwal changed their profile picture. | 18:55:45 | |
Vedant Padwal | Hi Everyone, The KServe v0.7 release blog is out!, please check it out at https://kserve.github.io/website/blog/articles/2021-10-11-KServe-0.7-release/ | 18:55:51 |
Kevin Hu joined the room. | 21:04:14 | |
Kevin Hu changed their display name from _slack_kubeflow_U02JP9VPHH6 to Kevin Hu. | 21:15:48 |