ArangoDB Starter Architecture

What does the Starter do

The ArangoDB Starter is a program used to create ArangoDB database deployments on bare-metal (or virtual machines) with ease. It enables you to create everything from a simple Single server instance to a full blown Cluster with datacenter to datacenter replication in under 5 minutes.

The Starter is intended to be used in environments where there is no higher level orchestration system (e.g. Kubernetes) available.

Starter versions

The Starter is a separate process in a binary called arangodb (or arangodb.exe on Windows). This binary has its own version number that is independent of a ArangoDB (database) version.

This means that Starter version a.b.c can be used to run deployments of ArangoDB databases with different version. For example, the Starter with version 0.11.2 can be used to create ArangoDB deployments with ArangoDB version 3.2.<something> as well as deployments with ArangoDB version 3.3.<something>.

It also means that you can update the Starter independently from the ArangoDB database.

Note that the Starter is also included in all binary ArangoDB packages.

To find the versions of you Starters & ArangoDB database, run the following commands:

# To get the Starter version
arangodb --version
# To get the ArangoDB database version
arangod --version

Starter deployment modes

The Starter supports 3 different modes of ArangoDB deployments:

  1. Single server
  2. Active failover
  3. Cluster

Note: Datacenter replication is an option for the cluster deployment mode.

You select one of these modes using the --starter.mode command line option.

Depending on the mode you’ve selected, the Starter launches one or more (arangod / arangosync) server processes.

No matter which mode you select, the Starter always provides you a common directory structure for storing the servers data, configuration & log files.

Starter operating modes

The Starter can run as normal processes directly on the host operating system, or as containers in a docker runtime.

When running as normal process directly on the host operating system, the Starter launches the servers as child processes and monitors those. If one of the server processes terminates, a new one is started automatically.

When running in a docker container, the Starter launches the servers as separate docker containers, that share the volume namespace with the container that runs the Starter. It monitors those containers and if one terminates, a new container is launched automatically.

Starter data-directory

The Starter uses a single directory with a well known structure to store all data for its own configuration & logs, as well as the configuration, data & logs of all servers it starts.

This data directory is set using the command line option. It contains the following files & sub-directories.

  • setup.json The configuration of the “cluster of Starters”. For details see below. DO NOT edit this file.
  • arangodb.log The log file of the Starter
  • single<port>, agent<port>, coordinator<port>, dbserver<port>: directories for launched servers. These directories contain among others the following files:
    • apps: A directory with Foxx applications
    • data: A directory with database data
    • arangod.conf: The configuration file for the server. Editing this file is possible, but not recommended.
    • arangod.log: The log file of the server
    • arangod_command.txt: File containing the exact command line of the started server (for debugging purposes only)

Running on multiple machines

For the activefailover & cluster mode, it is required to run multiple Starters, as every Starter will only launch a subset of all servers needed to form the entire deployment. For example in cluster mode, a Starter will launch a single Agent, a single DB-Server and a single Coordinator.

It is the responsibility of the user to run the Starter on multiple machines such that enough servers are started to form the entire deployment. The minimum number of Starters needed is 3.

The Starters running on those machines need to know about each other’s existence. In order to do so, the Starters form a “cluster” of their own (not to be confused with the ArangoDB database cluster). This cluster of Starters is formed from the values given to the --starter.join command line option. You should pass the addresses (<host>:<port>) of all Starters.

For example a typical commandline for a cluster deployment looks like this:

arangodb --starter.mode=cluster --starter.join=hostA:8528,hostB:8528,hostC:8528
# this command is run on hostA, hostB and hostC.

The state of the cluster (of Starters) is stored in a configuration file called setup.json in the data directory of every Starter and the ArangoDB Agency is used to elect a master among all Starters.

The master Starter is responsible for maintaining the list of all Starters involved in the cluster and their addresses. The slave Starters (all Starters except the elected master) fetch this list from the master Starter on regular basis and store it to its own setup.json config file.

Note: The setup.json config file MUST NOT be edited manually.

Running on multiple machines (under the hood)

As mentioned above, when the Starter is used to create an activefailover or cluster deployment, it first creates a “cluster” of Starters.

These are the steps taken by the Starters to bootstrap such a deployment from scratch.

  1. All Starters are started (either manually or by some supervisor)
  2. All Starters try to read their config from setup.json. If that file exists and is valid, this bootstrap-from-scratch process is aborted and all Starters go directly to the running phase described below.
  3. All Starters create a unique ID
  4. The list of --starter.join arguments is sorted
  5. All Starters request the unique ID from the first server in the sorted --starter.join list, and compares the result with its own unique ID.
  6. The Starter that finds its own unique ID, is continuing as bootstrap master the other Starters are continuing as bootstrap slaves.
  7. The bootstrap master waits for at least 2 bootstrap slaves to join it.
  8. The bootstrap slaves contact the bootstrap master to join its cluster of Starters.
  9. Once the bootstrap master has received enough (at least 2) requests to join its cluster of Starters, it continues with the running phase.
  10. The bootstrap slaves keep asking the bootstrap master about its state. As soon as they receive confirmation to do so, they also continue with the running phase.

In the running phase all Starters launch the desired servers and keeps monitoring those servers. Once a functional Agency is detected, all Starters will try to be running master by trying to write their ID in a well known location in the Agency. The first Starter to succeed in doing so wins this master election.

The running master will keep writing its ID in the Agency in order to remaining the running master. Since this ID is written with a short time-to-live, other Starters are able to detect when the current running master has been stopped or is no longer responsible. In that case the remaining Starters will perform another master election to decide who will be the next running master.

API requests that involve the state of the cluster of Starters are always answered by the current running master. All other Starters will refer the request to the current running master.