Skip to content

Commit

Permalink
Consolidate HPA documentation
Browse files Browse the repository at this point in the history 
Signed-off-by: Alexey Fomenko <alexey.fomenko@intel.com>
  • Loading branch information
@byako
byako committed Aug 23, 2024
1 parent 3f97188 commit 2b89cd3
Show file tree
Hide file tree
Showing 5 changed files with 76 additions and 221 deletions.
67 changes: 67 additions & 0 deletions helm-charts/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,10 @@ This directory contains helm charts for [GenAIComps](https://github.com/opea-pro
- [Components](#components)
- [How to deploy with helm charts](#deploy-with-helm-charts)
- [Helm Charts Options](#helm-charts-options)
- [HorizontalPodAutoscaler (HPA) support](#horizontalpodautoscaler-hpa-support)
- [Pre-conditions](#pre-conditions)
- [Gotchas](#gotchas)
- [Verify HPA metrics](#verify-hpa-metrics)
- [Using Persistent Volume](#using-persistent-volume)
- [Using Private Docker Hub](#using-private-docker-hub)
- [Helm Charts repository](#helm-chart-repository)
Expand Down Expand Up @@ -62,8 +66,71 @@ There are global options(which should be shared across all components of a workl
| global | http_proxy https_proxy no_proxy | Proxy settings. If you are running the workloads behind the proxy, you'll have to add your proxy settings here. |
| global | modelUsePVC | The PersistentVolumeClaim you want to use as huggingface hub cache. Default "" means not using PVC. Only one of modelUsePVC/modelUseHostPath can be set. |
| global | modelUseHostPath | If you don't have Persistent Volume in your k8s cluster and want to use local directory as huggingface hub cache, set modelUseHostPath to your local directory name. Note that this can't share across nodes. Default "". Only one of modelUsePVC/modelUseHostPath can be set. |
| global | horizontalPodAutoscaler.enabled | Enable HPA autoscaling for TGI and TEI service deployments based on metrics they provide. See #pre-conditions and #gotchas before enabling! |
| tgi | LLM_MODEL_ID | The model id you want to use for tgi server. Default "Intel/neural-chat-7b-v3-3". |

## HorizontalPodAutoscaler (HPA) support

`horizontalPodAutoscaler` option enables HPA scaling for the TGI and TEI inferencing deployments:
https://kubernetes.io/docs/tasks/run-application/horizontal-pod-autoscale/

Autoscaling is based on custom application metrics provided through [Prometheus](https://prometheus.io/).

### Pre-conditions

If cluster does not run [Prometheus operator](https://github.com/prometheus-operator/kube-prometheus)
yet, it SHOULD be be installed before enabling HPA, e.g. by using:
https://github.com/prometheus-community/helm-charts/tree/main/charts/kube-prometheus-stack

Enabling HPA in top-level Helm chart (e.g. `chatqna`), overwrites cluster's current _PrometheusAdapter_
configuration with relevant custom metric queries. If that has queries you wish to retain, _or_ HPA is
otherwise enabled only in TGI or TEI subchart(s), you need add relevat queries to _PrometheusAdapter_
configuration _manually_ (e.g. from `chatqna` custom metrics Helm template).

### Gotchas

Why HPA is opt-in:

- Enabling (top level) chart `horizontalPodAutoscaler` option will _overwrite_ cluster's current
`PrometheusAdapter` configuration with its own custom metrics configuration.
Take copy of the existing one before install, if that matters:
`kubectl -n monitoring get cm/adapter-config -o yaml > adapter-config.yaml`
- `PrometheusAdapter` needs to be restarted after install, for it to read the new configuration:
`ns=monitoring; kubectl -n $ns delete $(kubectl -n $ns get pod --selector app.kubernetes.io/name=prometheus-adapter -o name)`
- By default Prometheus adds [k8s RBAC rules](https://github.com/prometheus-operator/kube-prometheus/blob/main/manifests/prometheus-roleBindingSpecificNamespaces.yaml)
for accessing metrics from `default`, `kube-system` and `monitoring` namespaces. If Helm is
asked to install OPEA services to some other namespace, those rules need to be updated accordingly
- Current HPA rules are examples for Xeon, for efficient scaling they need to be fine-tuned for given setup
performance (underlying HW, used models and data types, OPEA version etc)

### Verify HPA metrics

To verify that metrics required by horizontalPodAutoscaler option work, check following...

Prometheus has found the metric endpoints, i.e. last number on `curl` output is non-zero:

```console
chart=chatqna; # OPEA services prefix
ns=monitoring; # Prometheus namespace
prom_url=http://$(kubectl -n $ns get -o jsonpath="{.spec.clusterIP}:{.spec.ports[0].port}" svc/prometheus-k8s);
curl --no-progress-meter $prom_url/metrics | grep scrape_pool_targets.*$chart
```

**NOTE**: TGI and TEI inferencing services provide metrics endpoint only after they've processed their first request!

PrometheusAdapter lists TGI and/or TGI custom metrics (`te_*` / `tgi_*`):

```console
kubectl get --raw /apis/custom.metrics.k8s.io/v1beta1 | jq .resources[].name
```

HPA rules list valid (not `<unknown>`) TARGET values for service deployments:

```console
ns=default; # OPEA namespace
kubectl -n $ns get hpa
```

## Using Persistent Volume

It's common to use Persistent Volume(PV) for model caches(huggingface hub cache) in a production k8s cluster. We support to pass the PersistentVolumeClaim(PVC) to containers, but it's the user's responsibility to create the PVC depending on your k8s cluster's capability.
Expand Down
41 changes: 6 additions & 35 deletions helm-charts/chatqna/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,35 +34,6 @@ helm install chatqna chatqna --set global.HUGGINGFACEHUB_API_TOKEN=${HFTOKEN} --

1. Make sure your `MODELDIR` exists on the node where your workload is schedueled so you can cache the downloaded model for next time use. Otherwise, set `global.modelUseHostPath` to 'null' if you don't want to cache the model.

## HorizontalPodAutoscaler (HPA) support

`horizontalPodAutoscaler` option enables HPA scaling for the TGI and TEI inferencing deployments:
https://kubernetes.io/docs/tasks/run-application/horizontal-pod-autoscale/

Autoscaling is based on custom application metrics provided through [Prometheus](https://prometheus.io/).

### Pre-conditions

If cluster does not run [Prometheus operator](https://github.com/prometheus-operator/kube-prometheus)
yet, it SHOULD be be installed before enabling HPA, e.g. by using:
https://github.com/prometheus-community/helm-charts/tree/main/charts/kube-prometheus-stack

### Gotchas

Why HPA is opt-in:

- Enabling chart `horizontalPodAutoscaler` option will _overwrite_ cluster's current
`PrometheusAdapter` configuration with its own custom metrics configuration.
Take copy of the existing one before install, if that matters:
`kubectl -n monitoring get cm/adapter-config -o yaml > adapter-config.yaml`
- `PrometheusAdapter` needs to be restarted after install, for it to read the new configuration:
`ns=monitoring; kubectl -n $ns delete $(kubectl -n $ns get pod --selector app.kubernetes.io/name=prometheus-adapter -o name)`
- By default Prometheus adds [k8s RBAC rules](https://github.com/prometheus-operator/kube-prometheus/blob/main/manifests/prometheus-roleBindingSpecificNamespaces.yaml)
for accessing metrics from `default`, `kube-system` and `monitoring` namespaces. If Helm is
asked to install OPEA services to some other namespace, those rules need to be updated accordingly
- Provided HPA rules are examples for Xeon, for efficient scaling they need to be fine-tuned for given setup
(underlying HW, used models, OPEA version etc)

## Verify

To verify the installation, run the command `kubectl get pod` to make sure all pods are running.
Expand Down Expand Up @@ -112,9 +83,9 @@ Access `http://localhost:5174` to play with the ChatQnA workload through UI.

## Values

| Key | Type | Default | Description |
| -------------------------------------- | ------ | ----------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| image.repository | string | `"opea/chatqna"` | |
| service.port | string | `"8888"` | |
| tgi.LLM_MODEL_ID | string | `"Intel/neural-chat-7b-v3-3"` | Models id from https://huggingface.co/, or predownloaded model directory |
| global.horizontalPodAutoscaler.enabled | bop; | false | HPA autoscaling for the TGI and TEI service deployments based on metrics they provide. See #pre-conditions and #gotchas before enabling! (If one doesn't want one of them to be scaled, given service `maxReplicas` can be set to `1`) |
| Key | Type | Default | Description |
| -------------------------------------- | ------ | ----------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- |
| image.repository | string | `"opea/chatqna"` | |
| service.port | string | `"8888"` | |
| tgi.LLM_MODEL_ID | string | `"Intel/neural-chat-7b-v3-3"` | Models id from https://huggingface.co/, or predownloaded model directory |
| global.horizontalPodAutoscaler.enabled | bop; | false | HPA autoscaling for the TGI and TEI service deployments based on metrics they provide. See HPA section in ../README.md before enabling! |
63 changes: 1 addition & 62 deletions helm-charts/common/tei/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -21,38 +21,6 @@ MODELDIR=/mnt/opea-models

MODELNAME="/data/BAAI/bge-base-en-v1.5"

## HorizontalPodAutoscaler (HPA) support

`horizontalPodAutoscaler` option enables HPA scaling for the deployment:
https://kubernetes.io/docs/tasks/run-application/horizontal-pod-autoscale/

Autoscaling is based on custom application metrics provided through [Prometheus](https://prometheus.io/).

### Pre-conditions

If cluster does not run [Prometheus operator](https://github.com/prometheus-operator/kube-prometheus)
yet, it SHOULD be be installed before enabling HPA, e.g. by using:
https://github.com/prometheus-community/helm-charts/tree/main/charts/kube-prometheus-stack

`horizontalPodAutoscaler` enabled in top level Helm chart depending on this component (e.g. `chatqna`),
so that relevant custom metric queries are configured for PrometheusAdapter.

### Gotchas

Why HPA is opt-in:

- Enabling chart `horizontalPodAutoscaler` option will _overwrite_ cluster's current
`PrometheusAdapter` configuration with its own custom metrics configuration.
Take copy of the existing one before install, if that matters:
`kubectl -n monitoring get cm/adapter-config -o yaml > adapter-config.yaml`
- `PrometheusAdapter` needs to be restarted after install, for it to read the new configuration:
`ns=monitoring; kubectl -n $ns delete $(kubectl -n $ns get pod --selector app.kubernetes.io/name=prometheus-adapter -o name)`
- By default Prometheus adds [k8s RBAC rules](https://github.com/prometheus-operator/kube-prometheus/blob/main/manifests/prometheus-roleBindingSpecificNamespaces.yaml)
for accessing metrics from `default`, `kube-system` and `monitoring` namespaces. If Helm is
asked to install OPEA services to some other namespace, those rules need to be updated accordingly
- Provided HPA rules are examples for Xeon, for efficient scaling they need to be fine-tuned for given setup
(underlying HW, used models, OPEA version etc)

## Verify

To verify the installation, run the command `kubectl get pod` to make sure all pods are runinng.
Expand All @@ -65,35 +33,6 @@ Open another terminal and run the following command to verify the service if wor
curl http://localhost:2081/embed -X POST -d '{"inputs":"What is Deep Learning?"}' -H 'Content-Type: application/json'
```

### Verify HPA metrics

To verify that metrics required by horizontalPodAutoscaler option work, check that:

Prometheus has found the metric endpoints, i.e. last number on the line is non-zero:

```console
prom_url=http://$(kubectl -n monitoring get -o jsonpath="{.spec.clusterIP}:{.spec.ports[0].port}" svc/prometheus-k8s)
curl --no-progress-meter $prom_url/metrics | grep scrape_pool_targets.*tei
```

Prometheus adapter provides custom metrics for their data:

```console
kubectl get --raw /apis/custom.metrics.k8s.io/v1beta1 | jq .resources[].name
```

And those custom metrics have valid values for HPA rules:

```console
ns=default; # OPEA namespace
url=/apis/custom.metrics.k8s.io/v1beta1;
for m in $(kubectl get --raw $url | jq .resources[].name | cut -d/ -f2 | sort -u | tr -d '"'); do
kubectl get --raw $url/namespaces/$ns/metrics/$m | jq;
done | grep -e metricName -e value
```

NOTE: HuggingFace TGI and TEI services provide metrics endpoint only after they've processed their first request!

## Values

| Key | Type | Default | Description |
Expand All @@ -102,4 +41,4 @@ NOTE: HuggingFace TGI and TEI services provide metrics endpoint only after they'
| global.modelUseHostPath | string | `"/mnt/opea-models"` | Cached models directory, tei will not download if the model is cached here. The host path "modelUseHostPath" will be mounted to container as /data directory. Set this to null/empty will force it to download model. |
| image.repository | string | `"ghcr.io/huggingface/text-embeddings-inference"` | |
| image.tag | string | `"cpu-1.5"` | |
| horizontalPodAutoscaler.enabled | bool | false | Enable HPA autoscaling for the service deployments based on metrics it provides. See #pre-conditions and #gotchas before enabling! |
| horizontalPodAutoscaler.enabled | bool | false | Enable HPA autoscaling for the service deployment based on metrics it provides. See HPA section in ../../README.md before enabling! |
Loading

0 comments on commit 2b89cd3

Please sign in to comment.