7 High Availability

High availability can be defined as follows:

The system is designed for highest requirements in terms of performance and reliability. Several platform capabilities allow easy scaling without downtimes.

Since version 2.2, Intershop Order Management is ready for High Availability.

The following sections describe a tested and working approach to enabling IOM to be HA.

7.1 Symmetric cluster Nodes

High availability can be provided by using symmetric cluster nodes. Therefore every server node consists of:

• A WildFly application server acting as clustered server. A clustered server requires access to a database to register its own IP address required for JGroups cluster communication.
• A mount of a shared file-system. Of course, the shared file-system has to be highly available too, in order to create a highly available IOM cluster. GlusterFS is one solution to fulfill this requirement, a single NFS server would not.

The configuration of all cluster nodes is identical, so every server node is able to replace any other. As a result, the entire system can answer requests as long as at least one node exists.

The following diagram shows two clustered nodes with most cluster specific infrastructure and communication. GlusterFS can be used to replicate data between IOM servers. IOM Watchdog for managing the availability of application servers is not shown here.

7.2 WildFly Application Server as Part of Cluster Node

7.2.1 Scalable Applications

Most IOM applications (e.g., base, customization, process) running in WildFly application server are scaleable. They can be executed on many application servers in parallel and hence they do not need our attention.

7.2.2 Singleton Applications

There are two IOM applications running in WildFly application server that must never be executed in parallel:

• Impex
• Control

In order to guarantee that these applications are running on one node only, the singleton subsystem of WildFly application server is used. In short, this subsystem does the following:

• Impex and Control are marked as singleton-deployments, in order to get them under control of the singleton subsystem, see Singleton Deployments in the WildFly documentation
• Singleton subsystem use a replicated Infinispan cache for internal communication (information about available nodes, election of node, etc.), see Singleton Subsystem in the WildFly documentation
• Infinispan cache replication uses JGroups subsystem to transfer data between machines, see Infinispan Subsystem in the WildFly documentation
• JGroups subsystem uses TCP/IP for communication between cluster nodes. Every node registers the IP-Address stored in property JBOSS_BIND_ADDRESS_PRIVATE (in $OMS_ETC/installation.properties) via JDBC_PING at the database, see JDBCPING in the JBossDeveloper Wiki and JGroups Subsystem in the WildFly documentation.

7.2.3 IOM Watchdog

Additionally to the health checks made on the WildFly application servers controlling the behavior of the load balancer, it is required to run the WildFly application server under control of server monitoring tool, i.e., IOM Watchdog. Additionally, The Watchdog provides the ability to restart the server whenever it becomes unhealthy or the process dies unexpectedly. Please see Guide - IOM Watchdog 2.2 - 2.11 and IOM Technical Overview - High Availability (2.12 - 2.17)#Health Check Requests and Cluster Status for more information.

Since traffic to IOM servers is already controlled by health check requests made by the load balancer, which can react quite fast (it can enable/disable a WildFly application server within some seconds), the reaction of the Watchdog on unhealthy IOM application servers can be more lazy, e.g., react after 5 minutes in unhealthy state. The according configuration is made in $OMS_ETC/watchdog.properties.

# number of seconds before failed health-checks lead to restart of the
# watched process
watchdog.healthcheck.timeout = 300

7.2.4 File Synchronization

Some directories of IOM servers are containing stateful runtime data, that have to be shared by all IOM servers. These directories are placed as sub-directories within $OMS_SHARE.

It is recommended to use a kind of clustered files system for synchronization (e.g., GlusterFS). Clustered file systems are highly available. The failure of a single node has impact on this node only. Using a simple shared file system server is not providing a highly available solution. If the file server fails, all IOM servers are affected too.

7.3 Load Balancer

If two or more IOM servers are used, a load balancer has to be placed in front of them. This load balancer has to support:

• Session stickiness
• Session failover for undertow-subsystem
• Health check requests

7.3.1 Session Stickiness

In general all session data are persisted within the database, but there are some transient data stored at the session, which are used in AJAX requests only. The called page prepares some information stored at the session and AJAX calls embedded into the response of the page are accessing them. This mechanism works only if page- and AJAX-requests are both handled by the same frontend application server.

WildFly application server supports sticky sessions by appending a route-information to the session-cookie. The route is simply a unique identifier of the server that should handle the session. As long as this server is available, the load balancer has to send all requests of this session to the same server. To guarantee this behavior, the load balancer has to support this kind of session stickiness.

Both, frontend application server and load balancer, have to use identical routes. At the application server side, the route is configured by the following property, located at $OMS_ETC/system.std.frontend.properties:

# Make sure the "route" property at the load balancer uses the same ID as set here for "instance-id".
# The value assigned to instance-id will be appended to the session-ID. With the help of this information
# the load balancer (e.g., mod_prox_balancer) realizes session stickiness.
/subsystem=undertow: instance-id="${installation.SERVER_ID}"

This example shows the default configuration of IOM. instance-id is set to the value ${installation.SERVER_ID}, which itself is defined as SERVER_ID in $OMS_ETC/installation.properties.

7.3.2 Session Failover

If one of the WildFly servers disappears, e.g., for maintenance, due to a hardware failure, etc., the load balancer has to be able to send all incoming requests dedicated for this WildFly server to the remaining WildFly servers.

7.3.3 Health Check Requests

The load balancer should support health check requests. The health check requests support the load balancer in deciding which WildFly application server to use and which not. E.g., if the WildFly server is currently redeployed, is unable to connect to the database or it is unable to access the shared file system, the load balancer should not use the according WildFly server any longer. 

7.3.4 Load Balancer - Using Apache HTTP Server Configuration

You can use any load balancer you want, as long it is able to fulfill the requirements described in the sections above.

If you choose Apache HTTP server to be used as load balancer, the combination of mod_proxy_balancer and mod_proxy_hcheck is able to fulfill the requirements. The example below shows a configuration snippet of mod_proxy_balancer and mod_proxy_hcheck made for two WildFly application servers. The according machines have IPs 10.10.10.1 and 10.10.10.2 with OMT application listening on port 8080. The value of route uses the default setting of SERVER_ID, which is set in $OMS_ETC/installation.properties to $(hostname)_$(OMS_SERVER_TYPE). Health-check URL is set for both servers to /monitoring/services/health/status. The interval for checking health status is set to 10 seconds.

# load modules required for reverse proxy
LoadModule proxy_module modules/mod_proxy.so
LoadModule proxy_http_module modules/mod_proxy_http.so
LoadModule proxy_balancer_module modules/mod_proxy_balancer.so
LoadModule proxy_hcheck_module modules/mod_proxy_hcheck.so
LoadModule watchdog_module modules/mod_watchdog.so
LoadModule lbmethod_byrequests_module modules/mod_lbmethod_byrequests.so
LoadModule lbmethod_bytraffic_module modules/mod_lbmethod_bytraffic.so
LoadModule lbmethod_bybusyness_module modules/mod_lbmethod_bybusyness.so
LoadModule slotmem_shm_module modules/mod_slotmem_shm.so

# make sure, not to act as forward proxy
ProxyRequests Off

# do not rewrite Host-Header when forwarding request
ProxyPreserveHost on

# define proxy with balancer-members and settings
# make sure, route is identical to the settings made at the WildFly application servers
<Proxy balancer://omscluster>
    BalancerMember http://10.10.10.1:8180 max=5 timeout=40 retry=40 acquire=30000 route=fe1.iom.com_cluster hcmethod=GET hcinterval=10 hcuri=/monitoring/services/health/status
    BalancerMember http://10.10.10.2:8180 max=5 timeout=40 retry=40 acquire=30000 route=fe2.iom.com_cluster hcmethod=GET hcinterval=10 hcuri=/monitoring/services/health/status
</Proxy>

# enable balancer-manager, avoid proxying
<Location /balancer-manager>
  SetHandler balancer-manager
  ProxyPass !
</Location>
# enable server-status, avoid proxying
ExtendedStatus on
<Location /server-status>
  SetHandler server-status
  ProxyPass !
</Location>

# proxying everything else
ProxyPass / balancer://omscluster/ stickysession=JSESSIONID
ProxyPassReverse / balancer://omscluster/ stickysession=JSESSIONID

7.4 Health Check Requests and Cluster Status

7.4.1 Health Check Requests

All application servers provide a health check that can be requested using the URL /monitoring/services/health/status. It responds with HTTP status code 200 if the application server is healthy, otherwise it responds with 5XX.

To ease error analysis, the content delivered by the health check URL contains further information about processed checks. This information is provided in a machine readable format (JSON), which can also easily be understood by humans.

Health checks require the deployment of oms.monitoring-app-2.x.x.x.war.

Health checks of shared file system can be enabled or disabled. If the IOM installation consists of a single node only (e.g., for tests or demos), a shared file system is not required. The according setting is provided by the property is.oms.sharedfs.healthcheck:

/system-property=is.oms.sharedfs.healthcheck: "enabled"|"disabled"

See Concept - IOM Server Health Check for more information.

7.4.2 Cluster Status

For an overview of all application servers within a cluster, each health check request updates the current status of the application server within the database. The cluster status can be seen using the URL /monitoring/services/health/clusterstatus. Basic authentication is required. The user also requires the permission Basic application management assigned at root organization.

See Concept - IOM Server Health Check for more information.

7.5 High Availability Database

Frontend and backend servers are all connecting to the PostgreSQL database.

For high availability also the database and connection to it has to support HA:

• Database reconnect
• PostgreSQL HA cluster

7.5.1 DB Reconnect

To provide HA, the application servers are able to reconnect to the database without a restart.

To work properly, invalid connections must be recognized and evicted from the pool. The xa-datasource configuration defines how this happens.

Intershop recommends to use the background validation checker rather than the validate-on-match method to reduce the check overhead. Moreover, the timeouts configuration parameters may influence the reconnect behavior (old connections might not get evicted as long as the timeouts are not reached).
For more information about datasource configuration, see Datasource Parameters in the redhat customer portal. 

xa- datasource configuration example
# The pool size depends on the number of application servers and the database server ressources
min-pool-size="30"
max-pool-size="80"
pool-prefill="true"

# timeouts
set-tx-query-timeout="true"
query-timeout="3600"
blocking-timeout-wait-millis="3000"
idle-timeout-minutes="60"

#connection validation
validate-on-match="false"
background-validation="true"
background-validation-millis="20000"
exception-sorter-class-name="org.jboss.jca.adapters.jdbc.extensions.postgres.PostgreSQLExceptionSorter"
valid-connection-checker-class-name="org.jboss.jca.adapters.jdbc.extensions.postgres.PostgreSQLValidConnectionChecker"

# resources that are unused or not supported by IOM
interleaving="false"
pad-xid="false"
wrap-xa-resource="false"
same-rm-override="false"
share-prepared-statements="false"

7.5.2 PostgreSQL HA Cluster

IOM supports access to PostgreSQL HA clusters but has always connect to the master database.

A PostgreSQL HA cluster usually consists of one master server and one or more hot-standby servers. The master server is the only one, which is allowed to change data. An additional witness server is needed by the failover process when the total number of servers (master + standbys) is odd.

During the failover, the IOM application must be redirected to the new master. One solution is to add a proxy layer between the IOM application servers and the Postgres HA-cluster. This proxy-layer can be realized by PgBouncer. PgBouncer has to be reconfigured on the fly (without restart) whenever the current master changes. PgBouncer being a connection pool, can also be used to limit the number of connections to PostgreSQL. More than one instance of PgBoucer should be defined to avoid single points of failures.

The IOM database connection address is defined in $OMS_ETC/cluster.properties by the property is.oms.db.hostlist which supports a number of one or more database host addresses. For more information see Guide - Setup Intershop Order Management 2.10.

For more information about PostgreSQL HA clusters see http://repmgr.org and https://pgbouncer.github.io

7.6 Clock Synchronization

All hosts of the cluster should use a clock synchronization service (daemon).