Live OpenSearch Online Training starting on October 12! See all classes

Using Solr Operator to Autoscale Solr on Kubernetes

In this tutorial, you’ll see how to deploy Solr on Kubernetes. You’ll also see how to use the Solr Operator to autoscale a SolrCloud cluster based on CPU with the help of the Horizontal Pod Autoscaler. Let’s get going! 🙂


For SolrCloud we’ll need Zookeeper. Also, if we want to autoscale Solr with the Horizontal Pod Autoscaler (HPA) based on CPU – admittedly, the simplest scenario – we’ll need a Metrics Server.


Here, you have three options. First, is to set up Zookeeper manually. Basically you can create a StatefulSet, with your Zookeeper nodes (typically 3), then add a headless service on top. Here’s an example gist.

A cleaner way is to use Helm. There’s a well-maintained chart by Bitnami that you can install like this:

helm install bitnami-zookeeper oci:// \
--set image.tag=3.8 \
--set fourlwCommandsWhitelist="mntr\,conf\,ruok" \
--set autopurge.purgeInterval=1 \
--set heapSize=512 \
--set replicaCount=3


  • bitnami-zookeeper is the name we give to this release (package installation)
  • image.tag is your Zookeeper version. In this tutorial, we’ll use Solr 9.3, which uses Zookeeper 3.8
  • fourlwCommandsWhitelist will inject these four letter commands into Zookeeper’s configuration file, so that Solr can check the health of the Zookeeper ensemble
  • autopurge.purgeInterval makes Zookeeper check every hour for old snapshots and remove them
  • heapSize is the JVM heap in MB that we allow Zookeeper to use
  • replicaCount is the number of Zookeeper nodes

Have a look at the chart readme for more details. In this tutorial, the most important thing that we’re missing is persistence: you’ll want your Zookeeper nodes to use persistent volumes to store data. We’re not doing persistence here to keep things simple and environment-agnostic (i.e. you can run this tutorial on your local Docker Desktop).

Once the Zookeeper pods are up, you can try to connect to them via a port forward:

kubectl port-forward svc/bitnami-zookeeper 2181:2181

And then (in a different terminal), check if it responds:

echo ruok | nc localhost 2181

It should say imok. These are Zookeeper’s four-letter words for “Are you OK?” and “I am OK” respectively.

There is a third way to deploy Zookeeper: directly with the Solr Operator. This is actually the default method, you’d do nothing at this stage. In fact, for the other two options, we have to tell the Solr Operator not to deploy Zookeeper and use the “external” Zookeeper instead. We’ll do that later.

Why an external Zookeeper? I don’t have a strong preference, it’s just that I find the “embedded” Zookeeper Operator to be too much on the complex side and too little on the reliable side. Don’t quote me on it 😅

Metrics Server

Metrics Server is a lightweight source of basic metrics (CPU, memory) specifically designed to work with Horizontal Pod Autoscaler or Vertical Pod Autoscaler. Later in this tutorial, we’ll dynamically add&remove nodes based on CPU and we’ll need a provider for these metrics. The simplest way to get there is with a Metrics Server.

As always, Helm would be the most straightforward way. Add the official repo:

helm repo add metrics-server

Then install it:

helm install metrics-server metrics-server/metrics-server --set args\[0\]="--kubelet-insecure-tls"

Notice the --kubelet-insecure-tls parameter. This is needed for development setups (e.g. Docker Desktop) where the kubelet’s certificate isn’t signed by a trusted CA. In a production Kubernetes you shouldn’t need this. 

Installing Solr

Solr will be managed by the Solr Operator and we’ll need three things:

  • Set up CRDs for the object we’ll use. In this tutorial we’ll only refer to SolrClouds, but there are others under in the “crds” directory of your Solr Operator version
  • Install the Solr Operator
  • Set up a SolrCloud cluster


If you want to be lean and mean, you’d just install the SolrCloud CRD for this tutorial:

kubectl create -f

If you need to work with backups and other objects, you’ll likely want to try the all.yaml or all-with-dependencies.yaml instead. The latter is useful if you want to use the Zookeeper that Solr Operator can install on its own.

Installing the Solr Operator

We’ll use Helm once again. Add the repository:

helm repo add apache-solr

If you had the repo already there, you might want to do helm repo update. This applies to all repositories, like you’d do apt-get update for a local package manager.

When installing the Solr Operator, you’ll want to specify the version and whether you want it to use the Zookeeper Operator to manage Zookeeper. We don’t, we installed Zookeeper separately earlier:

helm install solr-operator apache-solr/solr-operator --version 0.7.1 \
--set zookeeper-operator.install=false

You’ll find other configuration options in the Solr Operator Helm chart values documentation. Note that this is the main branch, some options might differ for older versions of the Solr Operator.

Deploying a Solr Cluster

Helm is your friend. Here’s an example:

helm install example-solr apache-solr/solr --version 0.7.1 \
--set image.tag=9.3 \
--set solrOptions.javaMemory="-Xms500m -Xmx500m" \
--set zk.address="bitnami-zookeeper-headless:2181" \
--set podOptions.resources.requests.cpu=200m


  • The Helm chart version should match your Solr Operator version, they’re tested together.
  • image.tag is the Solr version.
  • solrOptions.javaMemory is one of the Solr-specific configuration options. In this case, heap size.
  • zk.address is the Zookeeper headless service. In our case, it’s the one set up by the Bitnami Helm chart. If you’re using the Zookeeper Operator referenced by the Solr Operator, skip this option. Note that we’re using the headless service here, because we need Solr to know the IPs of the Zookeeper pods, we don’t want to just connect to them via a single IP. Which is what a regular service does.
  • podOptions.resources.requests.cpu will request 0.2 vCPUs from Kubernetes. The value isn’t very important (unless it’s too high and the default of 3 pods won’t be able to fit in your cluster). It’s just that it’s required for autoscaling: if you want to use the Horizontal Pod Autoscaler on a metric from the Metrics Server, you need to request that metric for the pod first. If you don’t want to autoscale Solr, you can skip this option – though resource requests and limits are generally a good practice. You can always scale the cluster manually with a command like:

kubectl scale --replicas=5 solrcloud/example

To see the complete list of options for the Solr Helm chart, have a look at the Solr Helm chart values documentation . As with Zookeeper, persistent storage might be a useful one.

After installing the Helm chart, you can see the overall status of your Solr cluster with:

kubectl get solrclouds

Accessing Solr from Outside

At this point, Solr can only be accessed from within the Kubernetes cluster. If you want to, say, open the Admin UI, you have two options: port forwarding and Ingress.

You’d normally use port-forwarding in a development environment. You can connect to the SolrCloud service like this (it listens to port 80 by default):

kubectl port-forward svc/example-solrcloud-common 8983:80

Then you can see the Admin UI at http://localhost:8983

For production, you’ll want to use an Ingress. There’s a tutorial on how to install it locally on the official Solr Operation documentation. Once you have Ingress set up, you’ll need to point your SolrCloud to it via addressability.external options when you deploy it – you have an example in the docs as well.

Solr + HPA = ❤️

To autoscale in Kubernetes you’re most likely going to use the Horizontal Pod Autoscaler (HPA). There’s also the Vertical Pod Autoscaler (VPA), but it’s less common. The difference is that HPA will adjust the number of pods to match your capacity needs, while VPA will adjust the resource requests and (currently) restart pods so they can use more resources. With Solr, you usually have one pod per host, so using VPA wouldn’t help that much. We’ll concentrate on HPA.

If you followed so far, you should have 3 Solr pods (that’s the default). And if you didn’t set replicas manually, HPA can change the number of replicas, if you choose to set it up. So let’s do that:

kubectl autoscale solrcloud example --cpu-percent=2 --min=3 --max=6

In plain English, we just told Kubernetes to set up an autoscale (HPA) object that looks at our SolrCloud object named “example” (our test cluster) and aim for a CPU usage of 2%. Make sure that you have at least 3 pods and no more than 6.

Why 2%? Because I want to test scaling up, your idle CPU is likely to be more than that, so if you run:

kubectl get hpa

You’ll probably see that we’re using more than 2% CPU and HPA is adding more Solr pods. It should get to 6 quickly. You can check that in Solr Admin at http://localhost:8983/solr/#/~cloud?view=nodes

Similarly, you can “provoke” the cluster to scale down by updating the rule to a high CPU value:

kubectl autoscale solrcloud example --cpu-percent=90 --min=3 --max=6 

Rebalancing Solr Shards

Adding and removing Solr pods isn’t terribly useful if data doesn’t populate new pods while scaling up. Or doesn’t move off of nodes before they get evicted on scale-down. This kind of functionality is supported starting from Solr Operator version 0.8.0, which – at the time of writing this – is just around the corner.

In short, if you have Solr Operator 0.8.0 or later, feel free to create a collection that has more shards (e.g. 12). Then you’ll see (in e.g. the nodes view of the Solr Admin UI) not only nodes being added/removed, but also shards rebalancing.

For more information about how this rebalancing works under the hood, have a look at this talk from Houston Putman:


As we’ve set up a Solr cluster in Kubernetes and we autoscaled it, I hope you learned a few things. In my opinion, the devil is in the details. For example:

  • Which metrics should determine when to scale up or down? It’s quite likely that CPU alone doesn’t cut it for your use-case. Although if it does, HPA might just be enough, as it has some useful options around how often to add nodes, etc. If it doesn’t, you might want to have a look at KEDA, which can trigger HPA to autoscale on custom metrics, as defined by Scalers.
  • What exactly happens to the cluster during autoscaling? It may be under heavy load, in which case it might be even counterproductive to start shuffling shards. It might move too much data at once (in that case maybe replication throttling will help) or it might be in the middle of an upgrade.
  • Would it be better to add/remove replicas than to shuffle shards? For example, if you have a spike in traffic every morning when people start working, you might want to add additional nodes and replicas, then remove them when the spike is gone. You might not even need HPA for this, a couple of CronJobs that change the number of replicas at specific times could do the trick.

Either way, I’d suggest watching this space, as we plan to write more in-depth content about such scenarios.

And if you’re into Solr, you might not be aware that Sematext is a one-stop-shop for it. We offer:

Last but not least, if you’d like to provide such services, we’re looking for new colleagues 🙂

Start Free Trial