Neo4j Causal Cluster setup for AWS by GetSocial

A setup for a Neo4j Enterprise Causal Cluster on top of AWS ECS.

You can obtain Neo4j from the official website. Please contact [email protected] for Enterprise licensing.

---

Upgrade guide for 3.5.x -> 4.2.x

---

Content

1. Why
2. Features
3. Includes
4. Limitations
5. Prerequisites
6. About
   1. Core Servers
   2. Read Replicas
   3. Discovery endpoints
   4. Spot Setup
7. Usage
8. Upgrade version
   1. Patch version upgrades
   2. Major and minor version upgrades
9. Neo4j cluster operations manual
10. Troubleshooting

Why

Here at GetSocial we started using Neo4j a few years ago to enhance our product with the power of social connections and we have been trying to find the architecture that can keep up with our growing service.

We run our infrastructure on AWS and the first approach to host Neo4j was using the suggested (at that time) HA Cluster architecture neo4j-aws-ha-cluster, which lately become deprecated and that the new Causal Cluster architecture would be the preferred approach given its scaling possibilities and resilience to failing nodes.

Features

• Automatic daily backups to S3 using a spot based AWS Batch Jobs

• Bootstrap a cluster from a backup snapshot.

• Cluster self-healing with AutoScaling group and data volumes reuse.

• CloudWatch alerts setup.

• Logging to CloudWatch Logs (general log by default and slow query log can be optionally enabled).

• Automatically create users+credentials for read-only and read/write access.

• DNS based discovery without ELB (because Causal Cluster has built-in load balancing and routing via Bolt Driver).

• Automated and safe rolling operational procedures like upgrade, upscale, downscale, restart (powered by ecs-drain-lambda as external dependency, see additional information).

  WARNING Rolling upgrades have limitations

• Spot setup for reduced costs (but with additional risks).

Includes

• Customizable CloudFormation template.
• Extended docker image on top official Neo4j image. Current version - Neo4j 4.2.4:
  • Custom extension ecs-extension.sh for official docker-entrypoint.sh
  • Custom healthcheck script healthcheck.sh
  • Custom script for initial db users creation init-db.sh

Limitations

• 1 core server per AWS Availability Zone

  This setup using RexRay Docker Plugin to reuse/remount EBS data volumes upon container restarts, instances replacement, etc. However, this creates a major limitation, you can have only 1 instance per AZ and you should have an equal amount of instances to the amount of configured Availability Zones.

  For Example:

  • A setup where № of instances bigger than № of Availability Zones will not work
  • A setup where № of instances smaller than № of Availability Zones will generate operational issues

  Most of AWS regions has 3 AZs in place, so you can have only 3 CORE servers. Except us-east-1 which has 6 or ca-central-1 and us-west-1 which has 2 (as of September 2019).

• Backups related limitations

  • You can't restore a cluster from a backup without a downtime. Rolling backup restore is not possible. You need to stop the cluster and then do the imports. See further instructions
  • Restore from the specific backup file on the same cluster possible only once. Successful restore from backup will create file-marker on the data volume to avoid the backup re-import on every docker container restart.
  • During backup restore from the backup follower nodes may get restarted once, after they start next time cluster should successfully form itself.
  • Backup only for the default database is supported at the moment

• ECS Instances draining support

  • Most of the ECS/ASG related operations are based on the feature of draining ECS instances, however automatic draining is not a native feature and requires external lambda function to be deployed in your AWS Region. Recommended tool to use is ecs-drain-lambda.

• Autoscaling Groups doesn't scale cluster automatically but used to keep cluster at desired capacity and replace unhealthy nodes.

• Neo4j instances must have internet access (so your Subnets must have NAT Gateway or Internet Gateway attached (and Auto-assign public IPv4 address should be enabled)

Prerequisites

• Install Docker to build the image.
• AWS CLI for the ECR Auth.

About

Neo4j graph database is deployed as a Causal Cluster (HA clustering is deprecated in the latest neo4j versions). It uses Bolt – a highly efficient, lightweight binary client-server protocol designed for database applications.

Essentially it's a Neo4j cluster with a minimum of 3 nodes (for successful startup and leader election, the cluster will still function with 2 healthy nodes in runtime, this is used for rolling operations, nodes will be removed/restarted one by one while other two keep functioning).

Setup is split logically into 2 ECS clusters (yet still it's 1 Neo4j cluster):

• Main ECS Cluster of Core servers

  • Should have 3 or more nodes
  • Electing cluster leader which accepts writes
  • Persists data on external EBS volumes (see related limitations)

  IMPORTANT: please read official documentation about Core Servers

• Optional ECS Cluster of Read Replicas

  • Does not persist data on EBS volumes, upon each restart/replacement pull data from the Core servers
  • Can be horizontally scaled endlessly - as many nodes as needed
  • Does not participate in the leader election
  • Can not become the leader
  • Can not accept write queries
  • Read queries will be automatically routed to the Read replicas if bolt+routing:// is used
  • Backups are performed on these nodes to avoid performance hits on Core Servers
  • Can be deployed as a different EC2 instance type (for temporary heavy analytics, etc)
  • Can handle complex queries without affecting the performance of the R/W cluster.

  IMPORTANT: please read official documentation about Read Replicas

• Discovery endpoints

  Discovery endpoints powered by AWS Cloud Map integration with ECS. AWS Cloud Map automatically generates SRV DNS records in the Route53 Private hosted zone for healthy ECS Tasks (taking into account docker healthchecks). However, discovery via SRV will require some additional code in your application to be able to connect to the Neo4j servers, as well you can not open it in your browser. For convenience this CloudFormation template includes small Lambda function which generates the Multivalue A records based on automatically generated SRV ones.

  You can use the A records in production, however, since they are replicated from the SRV records (which are the source of truth) it may take a bit longer to propagate.

  Example	Type	Description
  discovery.core.neo4j.service	SRV	Autogenerated discovery SRV record for Core nodes in format discovery.<core_prefix>.<cloudmap_service_name>
  core.neo4j.service	Multivalue A	Replicated via lambda A record for Core nodes in format <core_prefix>.<cloudmap_service_name>
  discovery.replica.neo4j.service	SRV	Autogenerated discovery SRV record for Read replicas in format discovery.<replica_prefix>.<cloudmap_service_name>
  replica.neo4j.service	Multivalue A	Replicated via lambda A record for Read replicas in format <replica_prefix>.<cloudmap_service_name>

  NOTE: Read Replica endpoints created only if Read Replica is enabled for the deployment.

  WARNING: CloudFormation stack removal may fail due to the still existing A records created by lambda, in this case, you need to remove them manually from AWS CloudMap console and try CloudFormation stack removal once again.

  WARNING: Since AWS CloudMap creates Private Hosted zone you will not be able to reach it from your laptop or other VPC. To be able to reach the private records you may need to have a VPN connection to your VPC or some other ways.

• Spot setup

  This CloudFormation template allows you, optionally, to deploy your cluster on the EC2 Spot Instances. Deploying a cluster on the Spot instances allows you to save a lot of money on the EC2 instances, however, produces additional risks. These risks are mitigated by using mixed instance types in different Availability Zones, this is minimizing the chance that more than one Neo4j instance will be interrupted at the same time and Neo4j cluster of 3 nodes is tolerant to the interruption of one node.

  If you are unwilling to take the risk, it is possible to use Spot Instances only for Read replicas, interruption of which doesn't result in any risks.

  Usage of spot instances is configured via CloudFormation template parameters ClusterInstanceType and ReplicasInstanceType. To enable spot setup pick from the drop-down one of the "custom" instance types with spot prefix (underlying instance types under the "custom" ones, can be found in the CloudFormation template mappings)

Usage

1. Create an ECR repository for Neo4j custom images. You will use its URL. (URL looks like 1234567890.dkr.ecr.us-east-1.amazonaws.com/neo).

2. Save environment variable for use in the makefile (customize them first)

   export NEO_ECR_REPO=<paste here URL of your ECR repo>
   export NEO_AWS_REGION=<your AWS region>

3. Build a Docker image and push it to your ECR:

   make push_image

4. If you know what are you doing feel free to modify cloudformation.yml in any way you like before spinning up infrastructure, however, most of the things are customizable via parameters.

5. Create a Cloud Formation stack using cloudformation.yml with your parameters.

   If you don't need the Read Replicas you can set the ReplicasCount=0 and ignore the rest of Slave related parameters (except SlaveSubnetID you need to choose any subnet there)

   Parameters reference

   ---

   Neo4j License

   Parameter	Description
   AcceptLicense	Before using Neo4j, you must accept license

   ---

   Global configurations

   Parameter	Description
   VpcId	Existing AWS VPC to deploy the Neo4j cluster in
   KeyName	SSH key to use for cluster EC2 instances access
   ECSAMI	ECS Optimized AMI Version as SSM AMI metadata parameter path By default use the recommended one, however keeping this setting as is during further stack updates may result in unexpected AMI update (when new AMI version will become the recommended one) If you don't want to update AMI, pin AMI version to specific value with value like /aws/service/ecs/optimized-ami/amazon-linux-2/amzn2-ami-ecs-hvm-2.0.20181112-x86_64-ebs
   NodeSecurityGroups	List of additional Security Groups to assign to the EC2 instances (for example, your custom SG group for SSH access or VPN access, etc)
   SNSTopicArn	SNS topic to send CloudWatch Alerts to, you could provide an ARN of the existing topic or new topic will be created if you don't specify any

   ---

   Core Nodes Configuration

   Parameter	Description
   ClusterInstanceType	AWS Instance type to use for Neo4j Cluster Core nodes (possible to use spot instances, see details)
   SubnetID	List of subnets to deploy your cluster into. Must include at least 3 subnets in different AZ see details
   DesiredCapacity	Number of desired Neo4j Core nodes. Must be at least 3 nodes and must much the number of subnets in different AZs see details
   EBSSize	Size of EBS volume for Neo4j data in GBs
   EBSType	Type of EBS volume

   ---

   Read replicas configuration

   Parameter	Description
   ReplicasInstanceType	AWS Instance type to use for Neo4j Cluster Read Replicas (possible to use spot instances, see details)
   ReplicasCount	Number of the desired Neo4j Read replicas. Set to 0 if you don't want to deploy read replicas (all the other resources associated with Read replicas will be not created as well. In this case all other the replica related parameters can be ignored)
   ReplicasSubnetID	List of subnets to deploy your read replicas into.

   ---

   Docker image configurations

   Parameter	Description
   DockerImage	URL of your custom build Neo4j image (in the following format 111111111111.dkr.ecr.us-east-1.amazonaws.com/neo:c531de3a6655b8c885330ca91b867431760392bf)
   DockerECRARN	ARN of your Private ECR repo (in the following format arn:aws:ecr:us-east-1:111111111111:repository/neo)

   ---

   Neo4j users configuration

   Parameter	Description
   AdminUser	Must be neo4j
   AdminPassword	Password for the neo4j user
   ReadOnlyUser	Name for the Neo4j Read-Only user
   ReadOnlyUserPassword	Password for the Neo4j Read-Only user

   ---

   Cloud Map discovery settings

   Parameter	Description
   CloudMapNamespaceID	ID of an existing CloudMap Namespace to use for discovery. If not set, a new CloudMap Namespace will be created for you automatically
   CloudMapNamespaceName	Name of the CloudMap Namespace. If CloudMapNamespaceID is set to use an existing Namespace then CloudMapNamespaceName should be match the existing Namespace name. If CloudMapNamespaceID is not set, then a new CloudMap Namespace will be created automatically with the provided name. See more details
   Neo4jCoreSubdomain	The subdomain that will be used for the Neo4j Core Cluster. Will look like <subdomain>.<namespace>. For example: core.neo4j.service. By default is set to core
   Neo4jReplicasSubdomain	The subdomain that will be used for the Neo4j Read replicas. Will look like <subdomain>.<namespace>. For example: replica.neo4j.service, by default is set to replica

   ---

   Neo4j operations

   Parameter	Description
   BackupPath	Full S3 path (in format <bucket_name>/path/to/backup.zip). This parameter is used ONLY when you want to start/restore Neo4j cluster from backup. See related limitations
   BackupHourlyStoreForDays	Number of days to keep hourly backups
   BackupDailyStoreForDays	Number of days to keep daily backups (hourly backup made on midnight considered as daily backup)
   IsDrainSupported	If set to true, instances replacement will be done automatically in rolling way with the usage of ASG Termination Hooks. Set this parameter to true only after deploying ecs-drain-lambda, see additional information. If this parameter set to false Rolling Update of AutoScaling Group will not be triggered automatically, because that will result in downtime and you will need to do the operations manually (not recommended)
   SlowQueryLog	Enable logging of Cypher queries that takes longer than 500 ms to the CloudWatch log. If enabled additional tiny sidecar container will be deployed next to Neo4j containers to tail the slow_query.log and push it to the CloudWatch log (additional CloudWatch log group will be created as well)

Upgrade version

Patch version upgrades

Upgrade between patch versions can be done using rolling upgrades. However, it is possible only when a store format upgrade is not needed (see release notes for a particular change).

Major and minor version upgrades

Major and minor version upgrades not tested yet and probably may require offline upgrade.

Neo4j cluster operations manual

Most of cluster operations is done via ECS Console:

• Stop cluster

  1. Go to ECS Console
  2. Find your Neo4j cluster, try searching by Neo4jCluster (for Core nodes cluster) or Neo4jReplicas (for Read replicas cluster)
  3. Open the Neo4j service (the only one)
  4. Click Update
  5. Set Number of tasks to 0
  6. Click Skip to review and Update Service
  7. Wait until all Neo4j containers are stopped on the Service Deployments page (you will see Running count: 0)

• Start cluster

  1. Go to ECS Console
  2. Find your Neo4j cluster, try searching by Neo4jCluster (for Core nodes cluster) or Neo4jReplicas (for Read replicas cluster)
  3. Open the Neo4j service (the only one)
  4. Click Update
  5. Set Number of tasks to the number of nodes in your cluster (if it is Core cluster, the number is usually 3)
  6. Click Skip to review and Update Service
  7. Wait until all Neo4j containers are started on the Service Deployments page (you will see Running count: 3)

• Restart cluster

  1. Go to ECS Console
  2. Find your Neo4j cluster, try searching by Neo4jCluster (for Core nodes cluster) or Neo4jReplicas (for Read replicas cluster)
  3. Open the Neo4j service (the only one)
  4. Click Update
  5. Check Force new deployment
  6. Click Skip to review and Update Service
  7. Wait until all Neo4j containers are restarted in a rolling way on the Service Deployments page (you will see Running count of the old and new deployments)

• Reset cluster data

  Sometimes you may need to wipe all the data from the cluster (for testing purpose or, for example, you want to import data from the backup). The fastest way to do this is following:

  1. Stop cluster
  2. Go to AWS EC2 Volumes Console
  3. Search for the volumes with State: Available and tag:Name: <cloudformation-stack-name-data>
  4. This should probably result in 3 volumes
  5. Remove this volumes (DO THIS ON YOUR OWN RISK)
  6. Start cluster

Troubleshooting

In most of the cases, algorithm will be following:

1. Try neo4j UI on the port 7474. Run :sysinfo to see nodes present in the cluster.
2. Check the CloudWatch Logs output of neo4j container for any problems.
3. Check the Neo4j ECS Cluster/Service/Tasks for stopped ones and ecs/docker errors, container exist codes.
4. Check the Neo4j debug logs for any problems. Debug logs can be found on the disk or in the CloudWatch logs if CaptureDebugLogs enabled. And ECS tasks can be found in the AWS ECS Console.

Possible problems:

1. Cluster leader keep changing. Possible reasons are:

   • Neo4j containers are being restarted. Check ECS for the stopped tasks, if there are some, this means that containers are being restarted. You need to figure out why. They can be killed due the HealthCheck failure, due to instance replacement or some internal error. You should be able to find out that from the ECS Console. If it is some internal error, check the Neo4j container output logs for errors. So you need to find the reason and fix it.

   • Neo4j re-elects the cluster leader because of long GC pauses. You should be able see corresponding logs in the debug log. There are usually application related reasons, like very heavy query.

2. Cluster not forming. Check the output of Neo4j containers (in the CloudWatch logs) or the debug logs. Discovery happens in the cloudmap_discover function in the ecs-extension.sh script via calls to the AWS CloudMap API. So on the cluster forming stage, if any of the containers cannot start, discovery will fail.

   Also, check for configuration problems. Remember, that amount of configured Availability zones (via Subnets) should match the number of core nodes in the cluster, so RexRay plugin can create one EBS volume per AZ.

TODO

• Allow to customize backup frequency
• Automatically create CloudWatch dashboard with main cluster metrics
• Automatically scale Read Replicas?!
• Allow defining static On-Demand capacity for Spot setup
• Make stack simpler to deploy:
  • Publish custom neo4j image on DockerHub
  • Add LauchStack button
• More advanced monitoring (Halin or sending runtime metrics to CloudWatch)
• Backup of multiple databases