feat(k6): add support for k8s loading/unloading of Seldon CRs #5563
Conversation
This adds initial support for k8s via the `USE_KUBE_CONTROL_PLANE` environment variable: - models/pipelines/experiments in components/model.js are modified to also return CR yamls - updates functions in components/utils.js (including `setupBase(...)` and `teardownBase(...)` to use xk6-kubernetes, when configured - all scenarios already using `setupBase(...)` should work unchanged **Which issue(s) this PR fixes:** - INFRA-949 (internal issue) Extend existing k6 scenarios to use xk6-kubernetes - [ ] Test functionality in kind - [ ] Extend existing k6 scenarios to use xk6-kubernetes
lc525
changed the title
lc525
changed the title
- works: model creation + load - fails: getting model status via k8s (permissions) - fails: deleting/tear down models (permissions)
| - name: SCHEDULER_ENDPOINT | ||
| value: "${SCHEDULER_ENDPOINT}:9004" | ||
| - name: INFER_HTTP_ITERATIONS | ||
| value: "1" | ||
| - name: INFER_GRPC_ITERATIONS | ||
| value: "1" | ||
| - name: MODELNAME_PREFIX | ||
| value: "tfsimplea,pytorch_cifar10a,tfmnista,mlflow_winea,irisa" | ||
| value: "tfsimplea,pytorch-cifar10a,tfmnista,mlflow-winea,irisa" |
There was a problem hiding this comment.
names changed because the ones with _ were not getting loaded in k8s
| @@ -97,32 +101,32 @@ export function getModelInferencePayload(modelName, inferBatchSize) { | |||
| const shape = [inferBatchSize, 16] | |||
| var httpBytes = [] | |||
| var grpcBytes = [] | |||
|
|
|||
There was a problem hiding this comment.
changes from here to line 208 are linting changes
| @@ -0,0 +1,103 @@ | |||
| import { Kubernetes } from "k6/x/kubernetes"; | |||
| import { getConfig } from '../components/settings.js' | |||
| import { | |||
There was a problem hiding this comment.
in the future we might want also to check the status via kube. not a blocker for this PR though.
There was a problem hiding this comment.
Indeed, was already planning for that. The xk6-kubernetes extension is a bit wierd in that it doesn't have built-in functionality for waiting on a given condition/status. Also, it throws exceptions for any unexpected conditions, so I'll have to code that a bit defensively (via repeated resource gets).
tests/k6/components/k8s.js
Outdated
| } | ||
|
|
||
| export function loadModel(modelName, data, awaitReady=true) { | ||
| //console.log(data) |
There was a problem hiding this comment.
yep, will clean those up. They were there in the scheduler code so I've left them for dev debugging
|
|
||
| export function loadModel(modelName, data, awaitReady=true) { | ||
| //console.log(data) | ||
| if(!seldonObjExists(seldonObjectType.MODEL, modelName, namespace)) { |
There was a problem hiding this comment.
do we have to check if the model already exists?
There was a problem hiding this comment.
If a model already exists here we get an exception. The exception likely happens only when the loaded model has the same CR as an existing one (it is an apply after all).
tests/k6/components/k8s.js
Outdated
| } | ||
|
|
||
| export function unloadModel(modelName, awaitReady=true) { | ||
| // console.log("Unloading model "+modelName) |
tests/k6/components/k8s.js
Outdated
| } | ||
|
|
||
| export function loadPipeline(pipelineName, data, awaitReady=true) { | ||
| //console.log(data) |
tests/k6/components/model.js
Outdated
| @@ -1,3 +1,7 @@ | |||
| // import { dump as yamlDump } from "../import/js-yaml.mjs" | |||
| for (var i = 0; i < 16 * inferBatchSize; i++) { | ||
| grpcBytes.push("MQ=="); // base64 of 1 | ||
| httpBytes.push("97") | ||
| } | ||
| const payload = { | ||
| "http": {"inputs":[{"name":"INPUT0","data":httpBytes,"datatype":"BYTES","shape":shape},{"name":"INPUT1","data":httpBytes,"datatype":"BYTES","shape":shape}]}, | ||
| "grpc": {"inputs":[{"name":"INPUT0","contents":{"bytes_contents":grpcBytes},"datatype":"BYTES","shape":shape},{"name":"INPUT1","contents":{"bytes_contents":grpcBytes},"datatype":"BYTES","shape":shape}]} | ||
| "http": { "inputs": [{ "name": "INPUT0", "data": httpBytes, "datatype": "BYTES", "shape": shape }, { "name": "INPUT1", "data": httpBytes, "datatype": "BYTES", "shape": shape }] }, |
There was a problem hiding this comment.
question: how did you reformat this file?
There was a problem hiding this comment.
It was the automated formatter from intelliJ (we should try to find a cli formatter and add linting/formatting to the makefile)
tests/k6/components/model.js
Outdated
| "storageUri": uri, | ||
| "requirements": modelTemplate.requirements, | ||
| "memory": (memoryBytes == null) ? modelTemplate.memoryBytes : memoryBytes, | ||
| "minReplicas": 1, |
There was a problem hiding this comment.
not sure if we need minReplicas, it will trigger I think autoscaling of models which we might want to do.
maybe at least align with the grpc api path.
| @@ -0,0 +1,5 @@ | |||
| export const seldonObjectType = { | |||
- cleanup commented debug calls - remove minReplicas from CRs
This adds initial support for k8s via the
USE_KUBE_CONTROL_PLANEenvironment variable:setupBase(...)andteardownBase(...)to use xk6-kubernetes, when configuredsetupBase(...)should work unchangedWhich issue(s) this PR fixes:
TODO