Multi-Master Replication Clusters in Kubernetes and Docker Swarm

For more examples visit – https://github.com/franzinc/agraph-examples

Introduction

In this document we primarily discuss running a Multi-Master Replication cluster (MMR) inside Kubernetes. We will also show a Docker Swarm implementation.

This directory and subdirectories contain code you can use to run an MMR cluster. The second half of this document is entitled Setting up and running MMR under Kubernetes and that is where you’ll see the steps needed to run the MMR cluster in Kubernetes.

MMR replication clusters are different from distributed AllegroGraph clusters in these important ways:

  1. Each member of the cluster needs to be able to make a TCP connection to each other member of the cluster. The connection is to a port computed at run time. The range of port numbers to which a connection is made can be constrained by the agraph.cfg file but typically this will be a large range to ensure that at least one port in that range is not in used.
  2. All members of the cluster hold the complete database (although for brief periods of time they can be out of sync and catching up with one another).

MMR replication clusters don’t quite fit the Kubernetes model in these ways

  1. When the cluster is running normally each instance knows the DNS name or IP address of each other instance. In Kubernetes you don’t want to depend on the IP address of another cluster’s pod as those pods can go away and a replacement started at a different IP address. We’ll describe below our solution to this.
  2. Services are a way to hide the actual location of a pod however they are designed to handle a set of known ports.. In our case we need to connect from one pod to a known-at-runtime port of another pod and this isn’t what services are designed for.
  3. A key feature of Kubernetes is the ability to scale up and down the number of processes in order to handle the load appropriately. Processes are usually single purpose and stateless. An MMR process is a full database server with a complete copy of the repository. Scaling up is not a quick and simple operation – the database must be copied from another node. Thus scaling up is a more deliberate process rather than something automatically done when the load on the system changes during the day.

The Design

  1. We have a headless service for our controlling instance StatefulSet and that causes there to be a DNS entry for the name controlling that points to the current IP address of the node in which the controlling instance runs. Thus we don’t need to hardwire the IP address of the controlling instance (as we do in our AWS load balancer implementation).
  2. The controlling instance uses two PersistentVolumes to store: 1. The repo we’re replicating and 2. The token that other nodes can use to connect to this node. Should the controlling instance AllegroGraph server die (or the pod in which it runs dies) then when the pod is started again it will have access to the data on those two persistent volumes.
  3. We call the other instances in the cluster Copy instances. These are full read-write instances of the repository but we don’t back up their data in a persistent volume. This is because we want to scale up and down the number of Copy instances. When we scale down we don’t want to save the old data since when we scale down we remove that instance from the cluster thus the repo in the cluster can never join the cluster again. We denote the Copy instances by their IP addresses. The Copy instances can find the address of the controlling instance via DNS. The controlling instance will pass the cluster configuration to the Copy instance and that configuration information will have the IP addresses of the other Copy instances. This is how the Copy instances find each other.
  4. We have a load balancer that allows one to access a random Copy instance from an external IP address. This load balancer doesn’t support sessions so it’s only useful for doing queries and quick inserts that don’t need a session.
  5. We have a load balancer that allows access to the Controlling instance via HTTP. While this load balancer also doesn’t have session support, because there is only one controlling instance it’s not a problem if you start an AllegroGraph session because all sessions will live on the single controlling instance.

We’ve had the most experience with Kubernetes on the Google Cloud Platform. There is no requirement that the load balancer support sessions and the GCP version does not at this time, but that doesn’t mean that session support isn’t present in the load balancer in other cloud platforms. Also there is a large community of Kubernetes developers and one may find a load balancer with session support available from a third party.

Implementation

We build and deploy in three subdirectories. We’ll describe the contents of the directories first and then give step by step instructions on how to use the contents of the directories.

Directory ag/

In this directory we build a Docker image holding an installed AllegroGraph. The Dockerfile is

FROM centos:7

#
# AllegroGraph root is /app/agraph
#

RUN yum -y install net-tools iputils bind-utils wget hostname

ARG agversion=agraph-6.6.0
ARG agdistfile=${agversion}-linuxamd64.64.tar.gz

# This ADD command will automatically extract the contents
# of the tar.gz file
ADD ${agdistfile} .

# needed for agraph 6.7.0 and can't hurt for others
# change to 11 if you only have OpenSSL 1.1 installed
ENV ACL_OPENSSL_VERSION=10

# so prompts are readable in an emacs window
ENV PROMPT_COMMAND=

RUN groupadd agraph && useradd -d /home/agraph -g agraph agraph 
RUN mkdir /app 

# declare ARGs as late as possible to allow previous lines to be cached
# regardless of ARG values

ARG user
ARG password

RUN (cd ${agversion} ;  ./install-agraph /app/agraph -- --non-interactive \
		--runas-user agraph \
		--super-user $user \
		--super-password $password ) 

# remove files we don't need
RUN rm -fr /app/agraph/lib/doc /app/agraph/lib/demos

# we will attach persistent storage to this directory
VOLUME ["/app/agraph/data/rootcatalog"]

# patch to reduce cache time so we’ll see when the controlling instance moves.
# ag 6.7.0 has config parameter StaleDNSRetainTime which allows this to be
# done in the configuration.
COPY dnspatch.cl /app/agraph/lib/patches/dnspatch.cl

RUN chown -R agraph.agraph /app/agraph

The Dockerfile installs AllegroGraph in /app/agraph and creates an AllegroGraph super user with the name and password passed in as arguments. It creates a user agraph so that the AllegroGraph server will run as the user agraph rather than as root.

We have to worry about the controlling instance process dying and being restarted in another pod with a different IP address. Thus if we’ve cached the DNS mapping of controlling we need to notice as soon as possible that the mapping as changed. The dnspatch.cl file changes a parameter in the AllegroGraph DNS code to reduce the time we trust our DNS cache to be accurate so that we’ll quickly notice if the IP address of controlling changes.

We also install a number of networking tools. AllegroGraph doesn’t need these but if we want to do debugging inside the container they are useful to have installed.

The image created by this Dockerfile is pushed to the Docker Hub using an account you’ve specified (see the Makefile in this directory for details).

Directory agrepl/

Next we take the image created above and add the specific code to support replication clusters.

The Dockerfile is

ARG DockerAccount=specifyaccount

FROM ${DockerAccount}/ag:latest

#
# AllegroGraph root is /app/agraph

RUN mkdir /app/agraph/scripts
COPY . /app/agraph/scripts

# since we only map one port from the outside into our cluster
# we need any sessions created to continue to use that one port.
RUN echo "UseMainPortForSessions true" >> /app/agraph/lib/agraph.cfg

# settings/user will be overwritten with a persistent mount so copy
# the data to another location so it can be restored.
RUN cp -rp /app/agraph/data/settings/user /app/agraph/data/user

ENTRYPOINT ["/app/agraph/scripts/repl.sh"]

When building an image using this Dockerfile you must specify

--build-arg DockerAccount=MyDockerAccount

where MyDockerAccount is a Docker account you’re authorized to push images to.

The Dockerfile installs the scripts repl.shvars.sh and accounts.sh. These are run when this container starts.

We modify the agraph.cfg with a line that ensures that even if we create a session that we’ll continue to access it via port 10035 since the load balancer we’ll use to access AllegroGraph only forwards 10035 to AllegroGraph.

Also we know that we’ll be installing a persistent volume at /app/agraph/data/user so we make a copy of that directory in another location since the current contents will be invisible when a volume is mounted on top of it. We need the contents as that is where the credentials for the user we created when AllegroGraph was installed.

Initially the file settings/user/username will contain the credentials we specified when we installed AllegroGraph in first Dockerfile. When we create a cluster instance a new token is created and this is used in place of the password for the test account. This token is stored in settings/user/username which is why we need this to be an instance-specific and persistent filesystem for the controlling instance.

When this container starts it runs repl.sh which first runs accounts.sh and vars.sh.

accounts.sh is a file created by the top level Makefile to store the account information for the user account we created when we installed AllegroGraph.

vars.sh is

# constants need by scripts
port=10035
reponame=myrepl

# compute our ip address, the first one printed by hostname
myip=$(hostname -I | sed -e 's/ .*$//')

In vars.sh we specify the information about the repository we’ll create and our IP address.

The script repl.sh is this:

#!/bin/bash
#
## to start ag and then create or join a cluster
##

cd /app/agraph/scripts

set -x
. ./accounts.sh
. ./vars.sh

agtool=/app/agraph/bin/agtool

echo ip is $myip


# move the copy of user with our login to the newly mounted volume
# if this is the first time we've run agraph on this volume
if [ ! -e /app/agraph/data/rootcatalog/$reponame ] ; then
    cp -rp /app/agraph/data/user/* /app/agraph/data/settings/user
fi  

# due to volume mounts /app/agraph/data could be owned by root
# so we have to take back ownership
chown -R agraph.agraph /app/agraph/data


## start agraph
/app/agraph/bin/agraph-control --config /app/agraph/lib/agraph.cfg start

term_handler() {
    # this signal is delivered when the pod is
    # about to be killed.  We remove ourselves
    # from the cluster.
   echo got term signal
   /bin/bash ./remove-instance.sh
   exit
}

sleepforever() {
    # This unusual way of sleeping allows
    # a TERM signal sent when the pod is to
    # die to then cause the shell to invoke
    # the term_handler function above.
    date
    while true
    do
        sleep 99999 & wait ${!}
    done
}    

if [ -e /app/agraph/data/rootcatalog/$reponame ] ; then
    echo repository $reponame already exists in this persistent volume
    sleepforever
fi    

controllinghost=controlling

controllingspec=$authuser:$authpassword@$controllinghost:$port/$reponame

if [ x$Controlling == "xyes" ] ;
then
   # It may take a little time for the dns record for 'controlling' to be present
   # and we need that record because the agtool program below will use it
   until host controlling ; do  echo controlling not in DNS yet; sleep 5 ; done
    
   ## create first and controlling cluster instance
   $agtool repl create-cluster $controllingspec controlling
    

else
    # wait for the controlling ag server to be running
    until curl -s http://$authuser:$authpassword@$controllinghost:$port/version ; do echo wait for controlling ; sleep 5; done

    # wait for server in this container to be running
    until curl -s http://$authuser:$authpassword@$myip:$port/version ; do echo wait for local server ; sleep 5; done

    
   # wait for cluster repo on the controlling instance to be present
   until $agtool repl status $controllingspec > /dev/null ; do echo wait for repo ; sleep 5; done
    
   
   myiname=i-$myip
   echo $myiname > instance-name.txt

   # construct the remove-instance.sh shell script to remove this instance
   # from the cluster when the instance is terminated.
   echo $agtool repl remove $controllingspec $myiname > remove-instance.sh
   chmod 755 remove-instance.sh
   #

   # note that
   #  % docker kill container
   # will send a SIGKILL signal by default  we can't trap on  SIGKILL.
   # so
   #  % docker kill -s TERM container
   # in order to test this handler
   
   trap term_handler SIGTERM SIGHUP SIGUSR1
   trap -p
   echo this pid is $$

   # join the cluster
   echo joining the cluster
   $agtool repl grow-cluster $controllingspec $authuser:$authpassword@$myip:$port/$reponame $myiname
   
fi
sleepforever

This script can be run under three different conditions

  1. Run when the Controlling instance is starting for the first time
  2. Run when the Controlling instance is restarting having run before and died (perhaps the machine on which it was running crashed or the AllegroGraph process had some error)
  3. Run when a Copy instance is starting for the first time. Copy instances are not restarted when they die. Instead a new instance is created to take the place of the dead instance. Therefore we don’t need to handle the case of a Copy instance restarting.

In cases 1 and 2 the environment variable Controlling will have the value “yes”.

In case 2 there will be a directory at /app/agraph/data/rootcatalog/$reponame.

In all cases we start an AllegroGraph server.

In case 1 we create a new cluster. In case 2 we just sleep and let the AllegroGraph server recover the replication repository and reconnect to the other members of the cluster.

In case 3 we wait for the controlling instance’s AllegroGraph to be running. Then we wait for our AllegroGraph server to be running. Then we wait for the replication repository we want to copy to be up and running. At that point we can grow the cluster by copying the cluster repository.

We also create a script which will remove this instance from the cluster should this pod be terminated. When the pod is killed (likely due to us scaling down the number of Copy instances) a termination signal will be sent first to the process allowing it to run this remove script before the pod completely disappears.

Directory kube/

This directory contains the yaml files that create kubernetes resources which then create pods and start the containers that create the AllegroGraph replication cluster.

controlling-service.yaml

We begin by defining the services. It may seem logical to define the applications before defining the service to expose the application but it’s the service we create that puts the application’s address in DNS and we want the DNS information to be present as soon as possible after the application starts. In the repl.sh script above we include a test to check when the DNS information is present before allowing the application to proceed.

apiVersion: v1
kind: Service
metadata:
 name: controlling
spec:
 clusterIP:  None
 selector:
   app: controlling
 ports:
 - name: http
   port: 10035
   targetPort: 10035

This selector defines a service for any container with a label with a key app and a value controlling. There aren’t any such containers yet but there will be. You create this service with

% kubectl create -f controlling-service.yaml

In fact for all the yaml files shown below you create the object they define by running

% kubectl create -f  filename.yaml

copy-service.yaml

We do a similar service for all the copy applications.

apiVersion: v1
kind: Service
metadata:
 name: copy
spec:
 clusterIP: None
 selector:
   app: copy
 ports:
 - name: main
   port: 10035
   targetPort: 10035

controlling.yaml

This is the most complex resource description for the cluster. We use a StatefulSet so we have a predictable name for the single pod we create. We define two persistent volumes. A StatefulSet is designed to control more than one pod so rather than a VolumeClaim we have a VolumeClaimTemplate so that each Pod can have its own persistent volume… but as it turns out we have only one pod in this set and we never scale up. There must be exactly one controlling instance.

We setup a liveness check so that if the AllegroGraph server dies Kubernetes will restart the pod and thus the AllegroGraph server. Because we’ve used a persistent volume for the AllegroGraph repositories when the AllegroGraph server restarts it will find that there is an existing MMR replication repository that was in use when the AllegroGraph server was last running. AllegroGraph will restart that replication repository which will cause that replication instance to reconnect to all the copy instances and become part of the cluster again.

We set the environment variable Controlling to yes and this causes this container to start up as a controlling instance (you’ll find the check for the Controlling environment variable in the repl.sh script above).

We have a volume mount for /dev/shm, the shared memory filesystem, because the default amount of shared memory allocated to a container by Kubernetes is too small to support AllegroGraph.

#
# stateful set of controlling instance
#

apiVersion: apps/v1beta1
kind: StatefulSet
metadata:
  name: controlling
spec:
  serviceName: controlling
  replicas: 1
  template:
    metadata:
      labels:
        app: controlling
    spec:
        containers:
        - name: controlling
          image: dockeraccount/agrepl:latest
          imagePullPolicy: Always
          livenessProbe:
            httpGet:
              path: /hostname
              port: 10035
            initialDelaySeconds: 30
          volumeMounts:
          - name: shm
            mountPath: /dev/shm
          - name: data
            mountPath: /app/agraph/data/rootcatalog
          - name: user
            mountPath: /app/agraph/data/settings/user
          env:
          - name: Controlling
            value: "yes"
        volumes:
         - name: shm
           emptyDir:
             medium: Memory
  volumeClaimTemplates:
         - metadata:
            name: data
           spec:
            resources:
              requests:
                storage: 20Gi
            accessModes:
            - ReadWriteOnce
         - metadata:
            name: user
           spec:
            resources:
              requests:
                storage: 10Mi
            accessModes:
            - ReadWriteOnce

copy.yaml

This StatefulSet is responsible for starting all the other instances. It’s much simpler as it doesn’t use Persistent Volumes

#
# stateful set of copies of the controlling instance
#

apiVersion: apps/v1beta1
kind: StatefulSet
metadata:
  name: copy
spec:
  serviceName: copy
  replicas: 2
  template:
    metadata:
      labels:
        app: copy
    spec:
        volumes:
         - name: shm
           emptyDir:
             medium: Memory
        containers:
        - name: controlling
          image: dockeraccount/agrepl:latest
          imagePullPolicy: Always
          livenessProbe:
            httpGet:
              path: /hostname
              port: 10035
            initialDelaySeconds: 30
          volumeMounts:
          - name: shm
            mountPath: /dev/shm

controlling-lb.yaml

We define a load balancer so applications on the internet outside of our cluster can communicate with the controlling instance. The IP address of the load balancer isn’t specified here. The cloud service provider (i.e. Google Cloud Platform or AWS) will determine an address after a minute or so and will make that value visible if you run

% kubectl get svc controlling-loadbalancer

The file is

apiVersion: v1
kind: Service
metadata:
  name: controlling-loadbalancer
spec:
  type: LoadBalancer
  ports:
  - port: 10035
    targetPort: 10035
  selector:
    app: controlling

copy-lb.yaml

As noted earlier the load balancer for the copy instances does not support sessions. However you can use the load balancer to issue queries or simple inserts that don’t require a session.

apiVersion: v1
kind: Service
metadata:
  name: copy-loadbalancer
spec:
  type: LoadBalancer
  ports:
  - port: 10035
    targetPort: 10035
  selector:
    app: copy

copy-0-lb.yaml

If you wish to access one of the copy instances explicitly so that you can create sessions you can create a load balancer which links to just one instance, in this case the first copy instance which is named “copy-0”.

apiVersion: v1
kind: Service
metadata:
  name: copy-0-loadbalancer
spec:
  type: LoadBalancer
  ports:
  - port: 10035
    targetPort: 10035
  selector:
    app: copy
    statefulset.kubernetes.io/pod-name: copy-0

Setting up and running MMR under Kubernetes

The code will build and deploy an AllegroGraph MMR cluster in Kubernetes. We’ve tested this in Google Cloud Platform and Amazon Web Service. This code requires Persistent Volumes and load balancers and thus requires a sophisticated platform to run (such as GCP or AWS).

Prerequisites

In order to use the code supplied you’ll need two additional things

  1. A Docker Hub account (https://hub.docker.com). A free account will work. You’ll want to make sure you can push to the hub without needing a password (use the docker login command to set that up).
  2. An AllegroGraph distribution in tar.gz format. We’ve been using agraph-6.6.0-linuxamd64.64.tar.gz in our testing. You can find the current set of server files at https://franz.com/agraph/downloads/server This file should be put in the ag subdirectory. Note that the Dockerfile in that directory has the line ARG agversion=agraph-6.6.0 which specifies the version of agraph to install. This must match the version of the ...tar.gz file you put in that directory.

Steps

Do Prerequisites

Fullfill the prerequisites above

Set parameters

There are 5 parameters

  1. Docker account – Must Specify
  2. AllegroGraph user – May want to specify
  3. AllegroGraph password – May want to specify
  4. AllegroGraph repository name – Unlikely to want to change
  5. AllegroGraph port – Very unlikely to want to change

The first three parameters can be set using the Makefile in the top level directory. The last two parameters are found in agrepl/vars.sh if you wish to change them. Note that the port number of 10035 is found in the yaml files in the kube subdirectory. If you change the port number you’ll have edit the yaml files as well.

The first three parameters are set via

% make account=DockerHubAccount user=username password=password

The account must be specified but the last two can be omitted and default to an AllegroGraph account name of test and a password of xyzzy.

If you choose to specify a password make it a simple one consisting of letters and numbers. The password will appear in shell commands and URLs and our simple scripts don’t escape characters that have a special meaning to the shell or URLs.

Install AllegroGraph

Change to the ag directory and build an image with AllegroGraph installed. Then push it to the Docker Hub

% cd ag
% make build
% make push
% cd ..

Create cluster-aware AllegroGraph image

Add scripts to create an image that will either create an AllegroGraph MMR cluster or join a cluster when started.

% cd agrepl
% make build
% make push
% cd ..

Setup a Kubernetes cluster

Now everything is ready to run in a Kubernetes cluster. You may already have a Kubernetes cluster running or you may need to create one. Both Google Cloud Platform and AWS have ways of creating a cluster using a web UI or a shell command. When you’ve got your cluster running you can do

% kubectl get nodes

and you’ll see your nodes listed. Once this works you can move into the next step.

Run an AllegroGraph MMR cluster

Starting the MMR cluster involves setting up a number of services and deploying pods. The Makefile will do that for you.

% cd kube
% make doall

You’ll see when it displays the services that there isn’t an external IP address allocated for the load balancers It can take a few minutes for an external IP address to be allocated and the load balancers setup so keep running

% kubectl  get svc

until you see an IP address given, and even then it may not work for a minute or two after that for the connection to be made.

Verify that the MMR cluster is running

You can use AllegroGraph Webview to see if the MMR cluster is running. Once you have an external IP address for the controlling-load-balancer go to this address in a web browser

http://external-ip-address:10035

Login with the credentials you used when you created the Docker images (the default is user test and password xyzzy). You’ll see a repository myrepl listed. Click on that. Midway down you’ll see a link titled

Manage Replication Instances as controller

Click on that link and you’ll see a table of three instances which now serve the same repository. This verifies that three pods started up and all linked to each other.

Namespaces

All objects created in Kubernetes have a name that is chosen either by the user or Kubernetes based on a name given by the user. Most names have an associated namespace. The combination of namespace and name must be unique among all objects in a Kubernetes cluster. The reason for having a namespace is that it prevents name clashes between multiple projects running in the same cluster that both choose to use the same name for an object.

The default namespace is named default.

Another big advantage using namespaces is that if you delete a namespace you delete all objects whose name is in that namespace. This is useful because a project in Kubernetes uses a lot of different types of objects and if you want to delete all the objects you’ve added to a Kubernetes cluster it can take a while to find all the objects by type and then delete them. However if you put all the objects in one namespace then you need only delete the namespace and you’re done.

In the Makefile we have this line

Namespace=testns

which is used by this rule

reset:
	-kubectl delete namespace ${Namespace}
	kubectl create namespace ${Namespace}
	kubectl config set-context `kubectl config current-context` --namespace ${Namespace}

The reset rule deletes all members of the Namespace named at the top of the Makefile (here testns) and then recreates the namespace and switches to it as the active namespace. After doing the reset all objects created will be created in the testns namespace.

We include this in the Makefile because you may find it useful.

Docker Swarm

The focus of this document is Kubernetes but we also have a Docker Swarm implementation of an AllegroGraph MMR cluster. Docker Swarm is significantly simpler to setup and manage than Kubernetes but has far fewer bells and whistles. Once you’ve gotten the ag and agrepl images built and pushed to the Docker Hub you need only link a set of machines running Docker together into a Docker Swarm and then

% cd swarm ; make controlling copy

and the AllegroGraph MMR cluster is running Once it is running you can access the cluster using Webview at

http://localhost:10035/