External Exposure of Neo4j Clusters when using client routing
This chapter describes how to route traffic from the outside world or Internet to a Neo4j cluster running in Kubernetes when using client routing.
Generally these instructions are only required for versions of Neo4j before 4.3.0. If you are using Neo4j 4.3.0 or later look at external exposure instructions
As described in the user guide, by default when you install Neo4j, each node in your cluster gets a private internal DNS address, which it advertises to its clients.
This works "out of the box" without any knowledge of your local addressing or DNS situation. The downside is that external clients cannot use the bolt+routing or neo4j protocols to connect to the cluster, because they cannot route traffic to strictly cluster internal DNS names. With the default helm install, connections from the outside fail even with proper exposure of the pods, because:
The client connects to Neo4j
Fetches a routing table, which contains entries like
External clients attempt and fail to connect to routing table entries
Overall connection fails or times out.
This article discusses these background issues in depth. These instructions are intended as a quick method of exposing Neo4j Clusters, but you may have to do additional work depending on your configuration.
To fix external clients, we need two things:
dbms.connector.*_addresssettings inside of each Neo4j node set to the externally routable address
An externally valid DNS name or IP address that clients can connect to, that routes traffic to the kubernetes pod
Some visual diagrams about what’s going on can be found in the architectural documentation here.
We’re going to address point 1 with some special configuration of the Neo4j pods themselves. I’ll explain the Neo4j config bits first, and then we’ll tie it together with the external. The most complex bit of this is ensuring each pod has the right config.
We’re going to address point 2 with Kubernetes Load Balancers. We will create one per pod in our Neo4j stateful set. We will associate static IP addresses to those load balancers. This enables packets to flow from outside of Kubernetes to the right pod / Neo4j cluster member.
In the helm chart within this repo, Neo4j core members are part of a stateful set, and get indexes. Given a deployment in a particular namespace, you end up with the following hostnames:
The helm chart in this repo can take a configurable ConfigMap for setting env vars on these pods. So
we can define our own configuration and pass it to the StatefulSet on startup. The
file in this directory is an example of that.
I’m using GCP, so it is done like this. Important notes here, on GCP the region must match your GKE region, and the network tier must be premium. On other clouds, the conceptual step here is the same, but the details will differ: you need to allocate 3 static IP addresses, which we’ll use in a later step.
# Customize these next 2 for the region of your GKE cluster,
# and your GCP project ID
for idx in 0 1 2 ; do
gcloud compute addresses create \
neo4j-static-ip-$idx --project=$PROJECT \
gcloud compute addresses describe neo4j-static-ip-$idx \
--region=$REGION --project=$PROJECT --format=json | jq -r '.address'
If you are doing this with Azure please note that the static IP addresses must be in the same
resource group as your kubernetes cluster, and can be created with
az network public-ip create like this (just one single sample):
az network public-ip create -g resource_group_name -n core01 --sku standard --dns-name neo4jcore01 --allocation-method Static. The Azure SKU used must be standard, and the resource group you need can be found in the kubernetes Load Balancer that [following the Azure Tutorial](https://docs.microsoft.com/en-us/azure/aks/kubernetes-walkthrough) sets up for you.
For the remainder of this tutorial, let’s assume that the core IP addresses I’ve allocated here are as follows; I’ll refer to them as these environment variables:
We will also need 3 exposure addresses that we want to advertise to the clients. I’m going to set these to be the same as the IP addresses, but if you have mapped DNS, you could use DNS names instead here.
It’s important for later steps that we have both IPs and addresses, because they’re used differently.
Recall that the Helm chart will let us configure core nodes with a custom config map. That’s good. But the problem with 1 configmap for all 3 cores is that each host needs different config for proper exposure. So in the helm chart, we’ve divided the neo4j settings into basic settings, and over-rideable settings. In the custom configmap example, you’ll see lines like this:
In a minute, after expanding $DEPLOYMENT to be "graph",
these variables have "host prefixes" -
graph_neo4j_core_0_* settings will only apply to the host
graph-neo4j-core-0. (The dashes are changed to _ because dashes aren’t supported in env var naming).
Very important to notice that these override settings have the pod name/hostname already "baked into them",
so it’s important to know how you’re planning to deploy Neo4j prior to setting this up.
These "address settings" need to be changed to match the 3 static IPs that we allocated in the previous
step. There are four critical env vars, all of which need to be configured, for each host:
With overrides, that’s 12 special overrides (4 vars each for 3 containers)
So using this "override approach" we can have 1 ConfigMap that specifies all the config for 3 members
of a cluster, while still allowing per-host configuration settings to differ. The override approach in
question is implemented in a small amount of bash that is in the
core-statefulset.yaml file. It simply
reads the environment and applies default values, permitting overrides if the override matches the host
where the changes are being applied.
In the next command, we’ll apply the custom configmap. Here you use the IP addresses from the previous step as ADDR0, ADDR1, and ADDR2. Alternatively, if those IP addresses are associated with DNS entries, you can use those DNS names instead. We’re calling them addresses because they can be any address you want to advertise, and don’t have to be an IP. But these addresses must resolve to the static IPs we created in the earlier step.
cat tools/external-exposure-legacy/custom-core-configmap.yaml | envsubst | kubectl apply -f -
Once customized, we now have a ConfigMap we can point our Neo4j deployment at, to advertise properly.
From the root of this repo, navigate to stable/neo4j and issue this command to install the helm chart with a deployment name of "graph". The deployment name must match what you did in previous steps, because remember we gave pod-specific overrides in the previous step.
helm install $DEPLOYMENT . \
--set core.numberOfServers=3 \
--set readReplica.numberOfServers=0 \
--set core.configMap=$DEPLOYMENT-neo4j-externally-addressable-config \
--set acceptLicenseAgreement=yes \
Note the custom configmap that is passed.
After a few minutes you’ll have a fully-formed cluster whose pods show ready, and which you can connect
to, but it will be advertising values that Kubernetes isn’t routing yet. So what we need to do next is to
create a load balancer per Neo4j core pod, and set the
loadBalancerIP to be the static IP address we
reserved in the earlier step, and advertised with the custom ConfigMap.
load-balancer.yaml file has been provided as a template, here’s how to make 3 of them for given static
# Reuse IP0, etc. from the earlier step here.
# These *must be IP addresses* and not hostnames, because we're
# assigning load balancer IP addresses to bind to.
export CORE_ADDRESSES=($IP0 $IP1 $IP2)
for x in 0 1 2 ; do
echo $DEPLOYMENT with IDX $IDX and IP $IP ;
cat tools/external-exposure-legacy/load-balancer.yaml | envsubst | kubectl apply -f -
You’ll notice we’re using 3 load balancers for 3 pods. In a sense it’s silly to "load balance" a single pod. But without a lot of extra software and configuration, this is the best option, because LBs will support TCP connections (ingresses won’t), and LBs can get their own independent IP addresses which can be associated with DNS later on. Had we used NodePorts, we’d be at the mercy of more dynamic IP assignment, and also have to worry about a Kubernetes cluster member itself falling over. ClusterIPs aren’t suitable at all, as they don’t give you external addresses.
Inside of these services, we use an
externalTrafficPolicy: Local. Because we’re routing to single pods and
don’t need any load spreading, local is fine. Refer to the kubernetes docs for more information on this topic.
There are other fancier options, such as the nginx-ingress controller but in this config we’re shooting for something as simple as possible that you can do with existing kubernetes primities without installing new packages you might not already have.
Potential Trip-up point: On GKE, the only thing needed to associate the static IP to the
load balancer is this
loadBalancerIP field in the YAML. On other clouds, there may be additional steps
to allocate the static IP to the Kubernetes cluster. Consult your local cloud documentation.
We can verify our services are running nicely like this:
$ kubectl get service | grep neo4j-external
zeke-neo4j-external-0 LoadBalancer 10.0.5.183 18.104.22.168 7687:30529/TCP,74.3.140843/TCP,7473:30325/TCP 115s
zeke-neo4j-external-1 LoadBalancer 10.0.9.182 22.214.171.124 7687:31059/TCP,74.3.141288/TCP,7473:31009/TCP 115s
zeke-neo4j-external-2 LoadBalancer 10.0.12.38 126.96.36.199 7687:30523/TCP,74.3.140844/TCP,7473:31732/TCP 114s
After all of these steps, you should end up with a cluster properly exposed. We can recover our password like so, and connect to any of the 3 static IPs.
export NEO4J_PASSWORD=$(kubectl get secrets graph-neo4j-secrets -o yaml | grep password | sed 's/.*: //' | base64 -d)
cypher-shell -a neo4j://188.8.131.52:7687 -u neo4j -p "$NEO4J_PASSWORD"
Additionally, since we exposed port 7474, you can go to any of the static IPs on port 7474 and end up with Neo4j browser and be able to connect.
If you have static IPs, you can of course associate DNS with them, and obtain signed certificates.
This in turn will let you expose signed cert HTTPS using standard Neo4j techniques, and will also permit advertising DNS instead of a bare IP if you wish.
For background on general Kubernetes network exposure issues, I’d recommend this article: Kubernetes $TYPE vs. LoadBalancer vs. Ingress? When should I use what?