HorizontalPodAutoscaler (HPA) support¶
Table of Contents¶
Introduction¶
autoscaling
option enables HPA scaling for relevant service components:
https://kubernetes.io/docs/tasks/run-application/horizontal-pod-autoscale/
Autoscaling is based on custom application metrics provided through Prometheus.
Pre-conditions¶
Read post-install steps before installation!
Resource requests¶
HPA controlled CPU pods SHOULD have appropriate resource requests or affinity rules (enabled in their subcharts and tested to work) so that k8s scheduler does not schedule too many of them on the same node(s). Otherwise they never reach ready state.
If you use different models than the default ones, update inferencing services (vLLM, TGI, TEI) resource requests to match model requirements.
Too large requests would not be a problem as long as pods still fit to available nodes. However, unless rules have been added to pods preventing them from being scheduled on same nodes, too small requests would be an issue:
Multiple inferencing instances interfere / slow down each other, especially if there are no NRI policies that provide further isolation
Containers can become non-functional when their actual resource usage crosses the specified limits
Prometheus metrics¶
Autoscaling requires k8s Prometheus installation and monitoring to be enabled in the top level chart. See monitoring instructions for details.
Prometheus adapter¶
Prometheus adapter is also needed, to provide k8s custom metrics based on collected service metrics: https://github.com/prometheus-community/helm-charts/tree/main/charts/prometheus-adapter
Install adapter after installing Prometheus:
$ prom_ns=monitoring # namespace for Prometheus+adapter
$ kubectl get svc -n $prom_ns
$ helm install prometheus-adapter prometheus-community/prometheus-adapter --version 4.10.0 -n $prom_ns \
--set prometheus.url=http://prometheus-stack-kube-prom-prometheus.$prom_ns.svc \
--set prometheus.prometheusSpec.serviceMonitorSelectorNilUsesHelmValues=false
NOTE: the service name given above in prometheus.url
must match the listed Prometheus service name,
otherwise adapter cannot access it!
(Alternative for setting the above prometheusSpec
variable to false
is making sure that
prometheusRelease
value in top-level chart matches the release name given to the Prometheus
install i.e. when it differs from prometheus-stack
used above. That is used to annotate
created serviceMonitors with a label Prometheus requires when above option is true
.)
Gotchas¶
Why HPA is opt-in:
Installing custom metrics for HPA requires manual post-install steps, as both Prometheus operator and adapter are missing support needed to automate that
Top level chart name needs to conform to Prometheus metric naming conventions, as it is also used as a metric name prefix (with dashes converted to underscores)
Unless pod resource requests, affinity rules, scheduling topology constraints and/or cluster NRI policies are used to better isolate CPU inferencing pods from each other, service instances scaled up on same node may never get to ready state
Current HPA rules are just examples, for efficient scaling they need to be fine-tuned for given setup performance (underlying HW, used models and data types, OPEA version etc)
Debugging missing custom metric issues is hard as logs rarely include anything helpful
Enable HPA¶
Install¶
ChatQnA includes pre-configured values files for scaling the services.
To enable HPA, add -f chatqna/hpa-values.yaml
option to your helm install
command line.
If CPU versions of TGI (and TEI) services are being scaled, resource requests and probe timings
suitable for CPU usage need to be used. chatqna/cpu-values.yaml
provides example of such constraints
which can be added (with -f
option) to your Helm install. As those values depend on the underlying HW,
used model, data type and image versions, the specified resource values may need to be updated.
Post-install¶
The above step created a custom metrics configuration for the Prometheus adapter, suitable for HPA use.
However, Helm does not allow OPEA chart to overwrite adapter configMap, as it belongs to another chart. Therefore, a manual step is needed to overwrite its current custom metric rules.
The following will overwrite the current adapter custom metric rules with the ones generated by OPEA Helm and will restart the adapter to apply the new rules.
scripts/install-custom-metrics.sh monitoring chatqna
(It assumes adapter to be in the monitoring
namespace, and the new rules are expected to be
generated by a ChatQnA chart release named chatqna
.)
YAML backups of the new and previous rules are saved to the current directory.
Verify¶
After verifying that service metrics work, one can verify that HPA rules can access custom metrics based on them.
Verify that custom metric values are available for scaling the services:
watch -n 5 scale-monitor-helm.sh default chatqna
(Assumes that HPA scaled chart is installed to default
namespace with chatqna
release name.)
NOTE: inferencing services provide metrics only after they’ve processed their first request. The reranking service is used only after the query context data has been uploaded. Until then, no metrics will be available for them.