High Availability: Backend Cluster

[edit on GitHub]

_images/chef_automate_full.png

This topic introduces the underlying concepts behind the architecture of the highly available Chef server cluster. The topic then describes the setup and installation process for a highly available Chef server cluster comprised of five total nodes (two frontend and three backend).

Note

This feature is included as part of the Chef Automate license agreement and is available via subscription.

Overview

The Chef server can operate in a high availability configuration that provides automated load balancing and failover for stateful components in the system architecture. This type of configuration typically splits the servers into two segments: The backend cluster, and the frontend group.

  • The frontend group, comprised of one (or more) nodes running the Chef server. Nodes in the frontend group handle requests to the Chef server API and access to the Chef management console. Frontend group nodes should be load balanced, and may be scaled horizontally by increasing the number of nodes available to handle requests.

  • The backend cluster, comprised of three nodes working together, provides highly available data persistence for the frontend group.

    Note

    At this time, backend clusters can only have three nodes.

_images/chef_server_ha_cluster.svg

Important

When doing cloud deployments, Chef HA clusters are not meant to be geographically dispersed across multiple regions or datacenters; however, in cloud providers such as AWS, you can deploy HA clusters across multiple Availability Zones within the same region.

Key Differences From Standalone Chef server

There are several key differences between the highly available Chef server cluster and a standalone Chef server instance.

  • While Apache Solr is used in standalone Chef server instances, in the highly available Chef server cluster it is replaced with Elasticsearch. Elasticsearch provides more flexible clustering options while maintaining search API compatibility with Apache Solr.
  • Writes to the search engine and the database are handled asynchronously via RabbitMQ and chef-expander in standalone Chef server instances. However, the highly available Chef server cluster writes to the search engine and the database simultaneously. As such the RabbitMQ and chef-expander services are no longer present in the highly available Chef server cluster.
  • Standalone Chef server instances write Bookshelf data to the filesystem. In a highly available Chef server cluster, Bookshelf data is written to the database.

Installation

These instructions assume you are using the following versions or newer:

  • chef-server : 12.5.0
  • chef-backend : 0.8.0

Download Chef server and Chef High Availability (chef-backend) if you do not have them already.

Before creating the backend HA cluster and building at least one Chef server to be part of the frontend group, verify:

  • The user who will install and build the backend HA cluster and frontend group has root access to all nodes.
  • The number of backend and frontend nodes that are desired. It is required to have three backend nodes, but the number of frontend nodes may vary from a single node to a load-balanced tiered configuration.
  • SSH access to all boxes that will belong to the backend HA cluster from the node that will be the initial bootstrap.
  • A time synchronization policy is in place, such as Network Time Protocol (NTP). Drift of less than 1.5 seconds must exist across all nodes in the backend HA cluster.

Step 1: Create Cluster

The first node must be bootstrapped to initialize the cluster. The node used to bootstrap the cluster will be the cluster leader when the cluster comes online. After bootstrap completes this node is no different from any other back-end node.

  1. Install the chef-backend package on the first backend node as root.

  2. Update /etc/chef-backend/chef-backend.rb with the following content:

    publish_address 'external_IP_address_of_this_box' # External ip address of this backend box

  3. If any of the backends or frontends are in different networks from each other then add a postgresql.md5_auth_cidr_addresses line to /etc/chef-backend/chef-backend.rb with the following content where , "<NET-1_IN_CIDR>", ..., "<NET-N_IN_CIDR>" is the list of all of the networks that your backends and frontends are in. See the Configuring Frontend and Backend Members on Different Networks section for more information:

    publish_address 'external_IP_address_of_this_box' # External ip address of this backend box
postgresql.md5_auth_cidr_addresses = ["samehost", "samenet", "<NET-1_IN_CIDR>", ..., "<NET-N_IN_CIDR>"]

  4. Run chef-backend-ctl create-cluster.

Step 2: Shared Credentials

The credentials file /etc/chef-backend/chef-backend-secrets.json generated by bootstrapping must be shared with the other nodes. You may copy them directly, or expose them via a common mounted location.

For example, to copy using ssh:

$ scp /etc/chef-backend/chef-backend-secrets.json <USER>@<IP_BE2>:/home/<USER>
$ scp /etc/chef-backend/chef-backend-secrets.json <USER>@<IP_BE3>:/home/<USER>

Delete this file from the destination after Step 4 has been completed for each backend being joined to the cluster.

Step 3: Install and Configure Remaining Backend Nodes

For each additional node do the following in sequence (if you attempt to join nodes in parallel the cluster may fail to become available):

  1. Install backend package on the node.

  2. If you added a postgresql.md5_auth_cidr_addresses line to the leader’s /etc/chef-backend/chef-backend.rb in Step 1: Create Cluster then update this node’s /etc/chef-backend/chef-backend.rb with the following content where postgresql.md5_auth_cidr_addresses is set to the same value used in the leader’s chef-backend.rb. If all of the backends and frontends are in the same network then you don’t need to modify this node’s /etc/chef-backend/chef-backend.rb at all.

    publish_address 'external_IP_address_of_this_box' # External ip address of this backend box
postgresql.md5_auth_cidr_addresses = ["samehost", "samenet", "<NET-1_IN_CIDR>", ..., "<NET-N_IN_CIDR>"]

  3. As root or with sudo:

    $ chef-backend-ctl join-cluster <IP_BE1> -s /home/<USER>/chef-backend-secrets.json

  4. Answer the prompts regarding which public IP to use. As an alternative, you may specify them on the chef-backend join-cluster command line. See chef-backend-ctl join-cluster --help for more information. If you manually added the publish_address line to /etc/chef-backend/chef-backend.rb then you will not be prompted for the public IP and you should not use the --publish-address option to specify the the public IP on the chef-backend join-cluster command line.

  5. If you copied the shared chef-backend-secrets.json file to a user HOME directory on this host, remove it now.

  6. Repeat these steps for each follower node, after which the cluster is online and available. From any node in the backend HA cluster, run the following command:

    $ chef-backend-ctl status

    should return something like:

    Service        Local Status        Time in State  Distributed Node Status
elasticsearch  running (pid 6661)  1d 5h 59m 41s  state: green; nodes online: 3/3
etcd           running (pid 6742)  1d 5h 59m 39s  health: green; healthy nodes: 3/3
leaderl        running (pid 6788)  1d 5h 59m 35s  leader: 1; waiting: 0; follower: 2; total: 3
postgresql     running (pid 6640)  1d 5h 59m 43s  leader: 1; offline: 0; syncing: 0; synced: 2

Step 4: Generate Chef server Configuration

Log into the node from Step 1, and we will generate our chef-server frontend node configuration:

$ chef-backend-ctl gen-server-config <FE1-FQDN> -f chef-server.rb.FE1
$ scp chef-server.rb.FE1 USER@<IP_FE1>:/home/<USER>

Note

/etc/chef-backend/chef-backend-secrets.json is not made available to Chef server frontend nodes.

Step 5: Install and Configure First Frontend

On the first frontend node, assuming that the generated configuration was copied as detailed in Step 4:

  1. Install the current chef-server-core package
  2. Run cp /home/<USER>/chef-server.rb.<FE1> /etc/opscode/chef-server.rb
  3. As the root user, run chef-server-ctl reconfigure

Adding More Frontends

For each additional frontend node you wish to add to your cluster:

  1. Install the current chef-server-core package.

  2. Generate a new /etc/opscode/chef-server.rb from any of the backend nodes via

    $ chef-backend-ctl gen-server-config <FE_NAME-FQDN> > chef-server.rb.<FE_NAME>

  3. Copy it to /etc/opscode on the new frontend node.

  4. From the first frontend node configured in Step 5, copy the following files from the first frontend to /etc/opscode on the new frontend node:

    • /etc/opscode/private-chef-secrets.json

    Note

    For Chef server versions prior to 12.14, you will also need to copy the key files:

    • /etc/opscode/webui_priv.pem
    • /etc/opscode/webui_pub.pem
    • /etc/opscode/pivotal.pem

  5. On the new frontend node run mkdir -p /var/opt/opscode/upgrades/.

  6. From the first frontend node, copy /var/opt/opscode/upgrades/migration-level to the same location on the new node.

  7. On the new frontend run touch /var/opt/opscode/bootstrapped.

  8. On the new frontend run chef-server-ctl reconfigure as root.

Upgrading Chef Server on the Frontend Machines

  1. On one frontend server, follow the standalone upgrade process.
  2. Copy /var/opt/opscode/upgrades/migration-level from the first upgraded frontend to /var/opt/opscode/upgrades/migration-level on each of the remaining frontends.
  3. Once the updated file has been copied to each of the remaining frontends, perform the standaloneupgradeprocess/upgrade_server.html#standalone on each of the frontend servers.

Configuring Frontend and Backend Members on Different Networks

By default, PostgreSQL only allows systems on its local network to connect to the database server that runs it and the pg_hba.conf used by PostgreSQL controls network access to the server. The default pg_hba.conf has the following four entries:

host    all         all         samehost               md5
hostssl replication replicator  samehost               md5
host    all         all         samenet                md5
hostssl replication replicator  samenet                md5

To allow other systems to connect, such as members of a frontend group that might exist on a different network, you will need to authorize that usage by adding the following line to the /etc/chef-backend/chef-backend.rb file on all of the backend members.

postgresql.md5_auth_cidr_addresses = ["samehost", "samenet", "<YOURNET IN CIDR>"]

Afer setting the md5_auth_cidr_addresses value and reconfiguring the server, two entries will be created in pg_hba.conf for each value in the md5_auth_cidr_addresses array. Existing values in pg_hba.conf will be overwritten by the values in the array, so we must also specify “samehost” and “samenet”, which will continue to allow systems on a local network to connect to PostgreSQL.

For example, if a frontend host at 192.168.1.3 can reach a backend member over the network, but the backend’s local network is 192.168.2.x, you would add the following line to /etc/chef-backend/chef-backend.rb

postgresql.md5_auth_cidr_addresses = ["samehost", "samenet", "192.168.1.3/24"]

which would result in the following two entries being added to the pg_hba.conf file.

host    all         all         samehost               md5
hostssl replication replicator  samehost               md5
host    all         all         samenet                md5
hostssl replication replicator  samenet                md5
host    all         all         192.168.1.3/24         md5
hostssl replication replicator  192.168.1.3/24         md5

Running chef-backend-ctl reconfigure on all the backends will allow that frontend to complete its connection.

Important

The postgresql.md5_auth_cidr_addresses subnet settings must be identical for all members of the backend cluster. In the case where the subnet settings of the frontend cluster are different from the subnet settings of the backend cluster, the values set on the members of the backend cluster should contain the subnet of the frontend cluster. This guarantees that all members of a cluster can still communicate with each other after a cluster change of state occurs. For example, if the frontend subnet setting is “192.168.1.0/24” and the backend subnet setting is “192.168.2.0/24”, then the postgresql.md5_auth_cidr_addresses subnet settings must be postgresql.md5_auth_cidr_addresses = ["samehost", "samenet", "192.168.1.0/24", 192.168.2.0/24]

Cluster Security Considerations

A backend cluster is expected to run in a trusted environment. This means that untrusted users that communicate with and/or eavesdrop on services provided by the backend cluster can potentially view sensitive data.

Communication Between Nodes

PostgreSQL communication between nodes in the backend cluster is encrypted, and uses password authentication. All other communication in the backend cluster is unauthenticated and happens in the clear (without encryption).

Communication Between Frontend Group & Backend Cluster

PostgreSQL communication from nodes in the frontend group to the leader of the backend cluster uses password authentication, but communication happens in the clear (without encryption).

Elasticsearch communication is unauthenticated and happens in the clear (without encryption).

Securing Communication

Because most of the peer communication between nodes in the backend cluster happens in the clear, the backend cluster is vulnerable to passive monitoring of network traffic between nodes. To help prevent an active attacker from intercepting or changing cluster data, Chef recommends using iptables or an equivalent network ACL tool to restrict access to PostgreSQL, Elasticsearch and etcd to only hosts that need access.

By service role, access requirements are as follows:

Service Access Requirements
PostgreSQL All backend cluster members and all Chef server frontend group nodes.
Elasticsearch All backend cluster members and all Chef server frontend group nodes.
etcd All backend cluster members and all Chef server frontend group nodes.

Services and Secrets

Communication with PostgreSQL requires password authentication. The backend cluster generates PostgreSQL users and passwords during the initial cluster-create. These passwords are present in the following files on disk:

Secret Owner Group Mode
/etc/chef-backend/secrets.json root chef_pgsql 0640
/var/opt/chef-backend/leaderl/data/sys.config chef_pgsql chef_pgsql 0600
/var/opt/chef-backend/PostgreSQL/9.5/recovery.conf chef_pgsql chef_pgsql 0600

The following services run on each node in the backend cluster. The user account under which the service runs as listed the second column:

Service Process Owner
postgresql chef_pgsql
elasticsearch chef-backend
etcd chef-backend
leaderl chef_pgsql
epmd chef_pgsql (or first user launching an erlang process)

Chef server frontend

The chef-backend-ctl gen-server-config command, which can be run as root from any node in the backend cluster, will automatically generate a configuration file containing the superuser database access credentials for the backend cluster PostgreSQL instance.

Software Versions

The backend HA cluster uses the Chef installer to package all of the software necessary to run the services included in the backend cluster. For a full list of the software packages included (and their versions), see the file located at /opt/chef-backend/version-manifest.json.

Do not attempt to upgrade individual components of the Chef package. Due to the way Chef packages are built, modifying any of the individual components in the package will lead to cluster instability. If the latest version of the backend cluster is providing an out-of-date package, please bring it to the attention of Chef by filling out a ticket with support@chef.io.

chef-backend.rb Options

The chef-backend.rb file is generated using chef-backend-ctl gen-sample-backend-config and controls most of the various feature and configuration flags going into a Chef HA backend node. A number of these options control the reliability, stability and uptime of the backend PostgreSQL databases, the elastic search index, and the leader election system. Please refrain from changing them unless you have been advised to do so.

  • fqdn Host name of this node.
  • hide_sensitive Set to false if you wish to print deltas of sensitive files and templates during chef-backend-ctl reconfigure runs. true by default.
  • ip_version Set to either 'ipv4' or 'ipv6'. 'ipv4' by default.
  • publish_address Externally resolvable IP address of this back-end node.

Common ‘Runit’ flags for any backend service

See https://github.com/chef-cookbooks/runit for details. Many of the flags are repeated across the various backend services - they are only documented once at the top here. The same defaults are used unless specified below.

  • postgresql.enable Sets up and runs this service. true by default.
  • postgresql.environment A hash of environment variables with their values as content used in the service’s env directory.
  • postgresql.log_directory The directory where the svlogd log service will run. '/var/log/chef-backend/postgresql/<version>' by default.
  • postgresql.log_rotation.file_maxbytes The maximum size a log file can grow to before it is automatically rotated. 104857600 by default (100MB).
  • postgresql.log_rotation.num_to_keep The maximum number of log files that will be retained after rotation. 10 by default.
  • etcd.enable
  • etcd.log_directory '/var/log/chef-backend/etcd' by default
  • etcd.log_rotation.file_maxbytes
  • etcd.log_rotation.num_to_keep
  • elasticsearch.enable
  • elasticsearch.log_directory '/var/log/chef-backend/elasticsearch' by default. Also affects path.logs in the elastic search configuration yml.
  • elasticsearch.log_rotation.file_maxbytes
  • elasticsearch.log_rotation.num_to_keep
  • leaderl.enable
  • leaderl.log_directory '/var/log/chef-backend/leaderl' by default.
  • leaderl.start_down Set the default state of the runit service to ‘down’ by creating <sv_dir>/down file. true by default.
  • leaderl.log_rotation.file_maxbytes
  • leaderl.log_rotation.num_to_keep

PostgreSQL settings

  • postgresql.db_superuser Super user account to create. Password is in chef-backend-secrets.json. 'chef_pgsql' by default.
  • postgresql.md5_auth_cidr_addresses A list of authorized addresses from which other backend nodes can connect to perform streaming replication. samehost and samenet are special symbols to allow connections from the this node’s IP address and its subnet. You may also use all to match any IP address. You may specify a hostname or IP address in CIDR format (172.20.143.89/32 for a single host, or 172.20.143.0/24 for a small network. See https://www.postgresql.org/docs/9.5/static/auth-pg-hba-conf.html for alternative formats. ["samehost", "samenet"] by default.
  • postgresql.replication_user Username used by postgres streaming replicator when accessing this node. 'replicator' by default.
  • postgresql.username 'chef_pgsql' by default.

PostgreSQL settings given to postgresql.conf

See https://www.postgresql.org/docs/9.5/static/runtime-config.html for details. Some defaults are provided:

  • postgresql.archive_command ''
  • postgresql.archive_mode 'off'
  • postgresql.archive_timeout 0
  • postgresql.checkpoint_completion_target 0.5
  • postgresql.checkpoint_timeout '5min'
  • postgresql.checkpoint_warning '30s'
  • postgresql.effective_cache_size Automatically calculated based on available memory.
  • postgresql.hot_standby 'on'
  • postgresql.keepalives_count 2 Sets tcp_keepalives_count
  • postgresql.keepalives_idle 60 Sets tcp_keepalives_idle
  • postgresql.keepalives_interval 15 Sets tcp_keepalives_interval
  • postgresql.log_checkpoints true
  • postgresql.log_min_duration_statement -1
  • postgresql.max_connections 350
  • postgresql.max_replication_slots 12
  • postgresql.max_wal_senders 12
  • postgresql.max_wal_size 64
  • postgresql.min_wal_size 5
  • postgresql.port 5432
  • postgresql.shared_buffers Automatically calculated based on available memory.
  • postgresql.wal_keep_segments 32
  • postgresql.wal_level 'hot_standby'
  • postgresql.wal_log_hints on
  • postgresql.work_mem '8MB'

etcd settings

  • etcd.client_port 2379 Port to use for ETCD_LISTEN_CLIENT_URLS and ETCD_ADVERTISE_CLIENT_URLS.
  • etcd.peer_port 2380 Port to use for ETCD_LISTEN_PEER_URLS and ETCD_ADVERTISE_PEER_URLS.

The following settings relate to etcd’s consensus protocol. Chef Backend builds its own leader election on top of etcd’s consensus protocol. Updating these settings may be advisable if you are seeing frequent failover events as a result of spurious etcd connection timeouts. The current defaults assume a high-latency environment, such those you might find if deploying Chef Backend to various cloud providers.

  • etcd.heartbeat_interval 500 ETCD_HEARTBEAT_INTERVAL in milliseconds. This is the frequency at which the leader will send heartbeats to followers. Etcd’s documentation recommends that this is set roughly to the round-trip times between members. (The default before 1.2 was 100)
  • etcd.election_timeout 5000 ETCD_ELECTION_TIMEOUT in milliseconds. This controls how long an etcd node will wait for heartbeat before triggering an election. Per Etcd’s documentation, this should be 5 to 10 times larger than the etcd.heartbeat_interval. Increasing etcd.election_timeout increases the time it will take for etcd to detect a failure. (The default value before 1.2 was 1000)
  • etcd.snapshot_count 5000 ETCD_SNAPSHOT_COUNT which is the number of committed transactions to trigger a snapshot to disk.

Note

Even though the defaults assume a high-latency environment, cloud deployments should be restricted to the same datacenter, or in AWS, in the same region. This means that geographically-dispersed cluster deployments are not supported. Multiple Availability Zones are supported as long as they are in the same region.

For additional information on the etcd tunables, see https://coreos.com/etcd/docs/latest/tuning.html.

Elastic Search JVM settings

  • elasticsearch.heap_size Automatically computed by elastic search based on available memory. Specify in MB if you wish to override.
  • elasticsearch.java_opts Flags to directly pass to the JVM when launching elastic search. If you override a heap flag here, the setting here takes precedence.
  • elasticsearch.new_size Java heap’s new generation size.

Elastic Search configuration

See https://www.elastic.co/guide/en/elasticsearch/reference/current/settings.html for details.

  • elasticsearch.plugins_directory '/var/opt/chef-backend/elasticsearch/plugins' Sets path.plugins.
  • elasticsearch.port 9200 Sets http.port.
  • elasticsearch.scripts_directory '/var/opt/chef-backend/elasticsearch/scripts' Sets path.scripts.

Chef HA backend leader management service settings

  • leaderl.db_timeout Socket timeout when connecting to PostgreSQL in milliseconds. 2000 by default.
  • leaderl.http_acceptors Http threads that responds to monitoring and leadership status requests from HAProxy. 10 by default.
  • leaderl.http_address The address that leaderl listens on. This address should not be 127.0.0.1. It should be reachable from any front-end node. '0.0.0.0' by default.
  • leaderl.http_port 7331 by default.
  • leaderl.leader_ttl_seconds The number of seconds it takes the leader key to expire. Increasing this value will increase the amount of time the cluster will take to recognize a failed leader. Lowering this value may lead to frequent leadership changes and thrashing. 30 by default (10 by default before 1.2).
  • leaderl.required_active_followers The number of followers that must be syncing via a PostgreSQL replication slot before a new leader will return 200 to /leader HTTP requests. If an existing leader fails to maintain this quorum of followers, the /leader endpoint will return 503 but active connections will still be able to complete their writes to the database. 0 by default.
  • leaderl.runsv_group The group that sensitive password files will belong to. This is used internally for test purposes and should never be modified otherwise. 'chef_pgsql' by default.
  • leaderl.status_internal_update_interval_seconds How often we check for a change in the leader service’s status. 5 seconds by default.
  • leaderl.status_post_update_interval_seconds How often etcd is updated with the leader service’s current status. 10 seconds by default.
  • leaderl.username 'chef_pgsql'
  • leaderl.log_rotation.max_messages_per_second Rate limit for the number of messages that the Erlang error_logger will output. 1000 by default.
  • leaderl.etcd_pool.ibrowse_options Internal options to affect how requests to etcd are made (see https://github.com/cmullaparthi/ibrowse/blob/master/doc/ibrowse.html).
  • leaderl.epmd_monitor.check_interval How often (in milliseconds) to check that leaderl is registered with the Erlang Port Mapping Daemon (epmd). 60000 by default.

Chef HA backend leader health status settings

  • leaderl.health_check.interval_seconds How frequently, in seconds, to poll the service for health status. We recommend setting this to at least 5 times the value of leaderl.leader_ttl_seconds. 5 by default (2 by default before version 1.2)
  • leaderl.health_check.max_bytes_behind_leader Limit on maximum different between elected leader and current node in bytes. 52428800 (50MB) by default.
  • leaderl.health_check.max_elasticsearch_failures Number of Elastic Search API failures allowed before health check fails. 5 by default.
  • leaderl.health_check.max_etcd_failures Number of etcd failures allowed before health check fails. 5 by default.
  • leaderl.health_check.max_pgsql_failures Number of PostgreSQL connection failures allowed before health check fails. 5 by default.
  • leaderl.health_check.fatal_system_checks Whether or not system check failures (such as disk space failures) will result in the node being marked ineligible for leadership. false by default. Added in Backend 1.4.
  • leaderl.health_check.disk_paths An array containing the paths to check for sufficient disk space. [/var/log/chef-backend, /var/opt/chef-backend] by default. Added in Backend 1.4.
  • leaderl.health_check.disk_min_space_mb The minimum amount of disk space (in megabytes) required for a disk health check to pass. 250 by default. Added in Backend 1.4.

Chef HA backend leader connection pool settings

See https://github.com/seth/pooler/blob/master/README.org for details. These are internal settings that affect the responsiveness, uptime and reliability of the backend cluster. They should not be modified unless you are advised to do so by Support.

  • leaderl.etcd_pool.cull_interval_seconds 60
  • leaderl.etcd_pool.http_timeout_ms 5000
  • leaderl.etcd_pool.init_count 10
  • leaderl.etcd_pool.max_age_seconds 60
  • leaderl.etcd_pool.max_connection_duration_seconds 300
  • leaderl.etcd_pool.max_count 10

SSL settings

If certificate and certificate_key are nil, the SSL Certificate will be auto-generated using the other parameters provided. Otherwise, they are on-disk locations to user-provided certificate.

  • ssl.certificate Provide this path if you have a pre-generated SSL cert.
  • ssl.certificate_key Provide this path if you have a pre-generated SSL cert.
  • ssl.ciphers Ordered list of allowed SSL ciphers. This will be updated based on security considerations and the version of OpenSSL being shipped.
  • ssl.company_name
  • ssl.country_name
  • ssl.data_dir Where certificates will be stored. '/var/opt/chef-backend/ssl/' by default
  • ssl.duration 3650 days by default (10 years).
  • ssl.key_length 2048 by default.
  • ssl.organizational_unit_name

chef-backend-ctl

The Chef server backend HA cluster includes a command-line utility named chef-backend-ctl. This command-line tool is used to manage the Chef server backend HA cluster, start and stop individual services, and tail Chef server log files. For more information, see the chef-backend-ctl documentation.