Fetching contributors…
Cannot retrieve contributors at this time
906 lines (700 sloc) 57.8 KB

Cross-Datacenter Replication Mode

Cross-Datacenter Replication mode is for when you want to run {project_name} in a cluster across multiple data centers, most typically using data center sites that are in different geographic regions. When using this mode, each data center will have its own cluster of {project_name} servers.

This documentation will refer the following example architecture diagram to illustrate and describe a simple Cross-Datacenter Replication use case.

Example Architecture Diagram

cross dc architecture


As this is an advanced topic, we recommend you first read the following, which provide valuable background knowledge:

Technical details

This section provides an introduction to the concepts and details of how {project_name} Cross-Datacenter Replication is accomplished.


{project_name} is stateful application. It uses the following as data sources:

  • A database is used to persist permanent data, such as user information.

  • An Infinispan cache is used to cache persistent data from the database and also to save some short-lived and frequently-changing metadata, such as for user sessions. Infinispan is usually much faster than a database, however the data saved using Infinispan are not permanent and is not expected to persist across cluster restarts.

In our example architecture, there are two data centers called site1 and site2. For Cross-Datacenter Replication, we must make sure that both sources of data work reliably and that {project_name} servers from site1 are eventually able to read the data saved by {project_name} servers on site2 .

Based on the environment, you have the option to decide if you prefer:

  • Reliability - which is typically used in Active/Active mode. Data written on site1 must be visible immediately on site2.

  • Performance - which is typically used in Active/Passive mode. Data written on site1 does not need to be visible immediately on site2. In some cases, the data may not be visible on site2 at all.

For more details, see Modes.

Request processing

An end user’s browser sends an HTTP request to the front end load balancer. This load balancer is usually HTTPD or WildFly with mod_cluster, NGINX, HA Proxy, or perhaps some other kind of software or hardware load balancer.

The load balancer then forwards the HTTP requests it receives to the underlying {project_name} instances, which can be spread among multiple data centers. Load balancers typically offer support for sticky sessions, which means that the load balancer is able to always forward all HTTP requests from the same user to the same {project_name} instance in same data center.

HTTP requests that are sent from client applications to the load balancer are called backchannel requests. These are not seen by an end user’s browser and therefore can not be part of a sticky session between the user and the load balancer. For backchannel requests, the loadbalancer can forward the HTTP request to any {project_name} instance in any data center. This is challenging as some OpenID Connect and some SAML flows require multiple HTTP requests from both the user and the application. Because we can not reliably depend on sticky sessions to force all the related requests to be sent to the same {project_name} instance in the same data center, we must instead replicate some data across data centers, so the data are seen by subsequent HTTP requests during a particular flow.


According your requirements, there are two basic operating modes for Cross-Datacenter Replication:

  • Active/Passive - Here the users and client applications send the requests just to the {project_name} nodes in just a single data center. The second data center is used just as a backup for saving the data. In case of the failure in the main data center, the data can be usually restored from the second data center.

  • Active/Active - Here the users and client applications send the requests to the {project_name} nodes in both data centers. It means that data need to be visible immediately on both sites and available to be consumed immediately from {project_name} servers on both sites. This is especially true if {project_name} server writes some data on site1, and it is required that the data are available immediately for reading by {project_name} servers on site2 immediately after the write on site1 is finished.

The active/passive mode is better for performance. For more information about how to configure caches for either mode, see: SYNC or ASYNC backups.


{project_name} uses a relational database management system (RDBMS) to persist some metadata about realms, clients, users, and so on. See this chapter of the server installation guide for more details. In a Cross-Datacenter Replication setup, we assume that either both data centers talk to the same database or that every data center has its own database node and both database nodes are synchronously replicated across the data centers. In both cases, it is required that when a {project_name} server on site1 persists some data and commits the transaction, those data are immediately visible by subsequent DB transactions on site2.

Details of DB setup are out-of-scope for {project_name}, however many RDBMS vendors like MariaDB and Oracle offer replicated databases and synchronous replication. We test {project_name} with these vendors:

  • Oracle Database 12c Release 1 (12.1) RAC

  • Galera 3.12 cluster for MariaDB server version 10.1.19-MariaDB

Infinispan caches

This section begins with a high level description of the Infinispan caches. More details of the cache setup follow.

Authentication sessions

In {project_name} we have the concept of authentication sessions. There is a separate Infinispan cache called authenticationSessions used to save data during authentication of particular user. Requests from this cache usually involve only a browser and the {project_name} server, not the application. Here we can rely on sticky sessions and the authenticationSessions cache content does not need to be replicated across data centers, even if you are in Active/Active mode.

Caching and invalidation of persistent data

{project_name} uses Infinispan to cache persistent data to avoid many unnecessary requests to the database. Caching improves performance, however it adds an additional challenge. When some {project_name} server updates any data, all other {project_name} servers in all data centers need to be aware of it, so they invalidate particular data from their caches. {project_name} uses local Infinispan caches called realms, users, and authorization to cache persistent data.

We use a separate cache, work, which is replicated across all data centers. The work cache itself does not cache any real data. It is used only for sending invalidation messages between cluster nodes and data centers. In other words, when data is updated, such as the user john, the {project_name} node sends the invalidation message to all other cluster nodes in the same data center and also to all other data centers. After receiving the invalidation notice, every node then invalidates the appropriate data from their local cache.

User sessions

There are Infinispan caches called sessions, clientSessions, offlineSessions, and offlineClientSessions, all of which usually need to be replicated across data centers. These caches are used to save data about user sessions, which are valid for the length of a user’s browser session. The caches must handle the HTTP requests from the end user and from the application. As described above, sticky sessions can not be reliably used in this instance, but we still want to ensure that subsequent HTTP requests can see the latest data. For this reason, the data are usually replicated across data centers.

Brute force protection

Finally the loginFailures cache is used to track data about failed logins, such as how many times the user john entered a bad password. The details are described here. It is up to the admin whether this cache should be replicated across data centers. To have an accurate count of login failures, the replication is needed. On the other hand, not replicating this data can save some performance. So if performance is more important than accurate counts of login failures, the replication can be avoided.

For more detail about how caches can be configured see Tuning the JDG cache configuration.

Communication details

{project_name} uses multiple, separate clusters of Infinispan caches. Every {project_name} node is in the cluster with the other {project_name} nodes in same data center, but not with the {project_name} nodes in different data centers. A {project_name} node does not communicate directly with the {project_name} nodes from different data centers. {project_name} nodes use external JDG (actually {jdgserver_name} servers) for communication across data centers. This is done using the Infinispan HotRod protocol.

The Infinispan caches on the {project_name} side must be configured with the remoteStore to ensure that data are saved to the remote cache. There is separate Infinispan cluster between JDG servers, so the data saved on JDG1 on site1 are replicated to JDG2 on site2 .

Finally, the receiving JDG server notifies the {project_name} servers in its cluster through the Client Listeners, which are a feature of the HotRod protocol. {project_name} nodes on site2 then update their Infinispan caches and the particular user session is also visible on {project_name} nodes on site2.

See the Example Architecture Diagram for more details.

Basic setup

For this example, we describe using two data centers, site1 and site2. Each data center consists of 1 {jdgserver_name} server and 2 {project_name} servers. We will end up with 2 {jdgserver_name} servers and 4 {project_name} servers in total.

  • Site1 consists of {jdgserver_name} server, jdg1, and 2 {project_name} servers, node11 and node12 .

  • Site2 consists of {jdgserver_name} server, jdg2, and 2 {project_name} servers, node21 and node22 .

  • {jdgserver_name} servers jdg1 and jdg2 are connected to each other through the RELAY2 protocol and backup based {jdgserver_name} caches in a similar way as described in the JDG documentation.

  • {project_name} servers node11 and node12 form a cluster with each other, but they do not communicate directly with any server in site2. They communicate with the Infinispan server jdg1 using the HotRod protocol (Remote cache). See Communication details for the details.

  • The same details apply for node21 and node22. They cluster with each other and communicate only with jdg2 server using the HotRod protocol.

Our example setup assumes all that all 4 {project_name} servers talk to the same database. In production, it is recommended to use separate synchronously replicated databases across data centers as described in Database.

{jdgserver_name} server setup

Follow these steps to set up the {jdgserver_name} server:

  1. Download {jdgserver_name} {jdgserver_version} server and unzip to a directory you choose. This location will be referred in later steps as JDG1_HOME .

  2. Change those things in the JDG1_HOME/standalone/configuration/clustered.xml in the configuration of JGroups subsystem:

    1. Add the xsite channel, which will use tcp stack, under channels element:

      <channels default="cluster">
          <channel name="cluster"/>
          <channel name="xsite" stack="tcp"/>
    2. Add a relay element to the end of the udp stack. We will configure it in a way that our site is site1 and the other site, where we will backup, is site2:

      <stack name="udp">
          <relay site="site1">
              <remote-site name="site2" channel="xsite"/>
              <property name="relay_multicasts">false</property>
    3. Configure the tcp stack to use TCPPING protocol instead of MPING. Remove the MPING element and replace it with the TCPPING. The initial_hosts element points to the hosts jdg1 and jdg2:

      <stack name="tcp">
          <transport type="TCP" socket-binding="jgroups-tcp"/>
          <protocol type="TCPPING">
              <property name="initial_hosts">jdg1[7600],jdg2[7600]</property>
              <property name="ergonomics">false</property>
          <protocol type="MERGE3"/>
      This is just an example setup to have things quickly running. In production, you are not required to use tcp stack for the JGroups RELAY2, but you can configure any other stack. For example, you could use the default udp stack, if the network between your data centers is able to support multicast. Just make sure that the {jdgserver_name} and {project_name} clusters are mutually indiscoverable. Similarly, you are not required to use TCPPING as discovery protocol. And in production, you probably won’t use TCPPING due it’s static nature. Finally, site names are also configurable. Details of this more-detailed setup are out-of-scope of the {project_name} documentation. See the {jdgserver_name} documentation and JGroups documentation for more details.
  3. Add this into JDG1_HOME/standalone/configuration/clustered.xml under cache-container named clustered:

    <cache-container name="clustered" default-cache="default" statistics="true">
            <replicated-cache-configuration name="sessions-cfg" mode="SYNC" start="EAGER" batching="false">
                <locking acquire-timeout="0" />
                    <backup site="site2" failure-policy="FAIL" strategy="SYNC" enabled="true">
                        <take-offline min-wait="60000" after-failures="3" />
            <replicated-cache name="work" configuration="sessions-cfg"/>
            <replicated-cache name="sessions" configuration="sessions-cfg"/>
            <replicated-cache name="clientSessions" configuration="sessions-cfg"/>
            <replicated-cache name="offlineSessions" configuration="sessions-cfg"/>
            <replicated-cache name="offlineClientSessions" configuration="sessions-cfg"/>
            <replicated-cache name="actionTokens" configuration="sessions-cfg"/>
            <replicated-cache name="loginFailures" configuration="sessions-cfg"/>
    Details about the configuration options inside replicated-cache-configuration are explained in Tuning the JDG cache configuration, which includes information about tweaking some of those options.
  4. Some {jdgserver_name} server releases require authorization before accessing protected caches over network.

    You should not see any issue if you use recommended {jdgserver_name} {jdgserver_version} server and this step can (and should) be ignored. Issues related to authorization may exist just for some other versions of {jdgserver_name} server.

    {project_name} requires updates to ___script_cache cache containing scripts. If you get errors accessing this cache, you will need to set up authorization in clustered.xml configuration as described below:

    1. In the <management> section, add a security realm:

              <security-realm name="AllowScriptManager">
                          <user username="___script_manager">
    2. In the server core subsystem, add <security> as below:

      <subsystem xmlns="urn:infinispan:server:core:8.4">
          <cache-container name="clustered" default-cache="default" statistics="true">
                      <role name="___script_manager" permissions="ALL"/>
    3. In the endpoint subsystem, add authentication configuration to Hot Rod connector:

      <subsystem xmlns="urn:infinispan:server:endpoint:8.1">
          <hotrod-connector cache-container="clustered" socket-binding="hotrod">
              <authentication security-realm="AllowScriptManager">
                  <sasl mechanisms="DIGEST-MD5" qop="auth" server-name="keycloak-jdg-server">
                          <no-anonymous value="false" />
  5. Copy the server to the second location, which will be referred to later as JDG2_HOME.

  6. In the JDG2_HOME/standalone/configuration/clustered.xml exchange site1 with site2 and vice versa, both in the configuration of relay in the JGroups subsystem and in configuration of backups in the cache-subsystem. For example:

    1. The relay element should look like this:

      <relay site="site2">
          <remote-site name="site1" channel="xsite"/>
          <property name="relay_multicasts">false</property>
    2. The backups element like this:

                      <backup site="site1" ....

      It is currently required to have different configuration files for the JDG servers on both sites as the Infinispan subsystem does not support replacing site names with expressions. See this issue for more details.

  7. Start server jdg1:

    cd JDG1_HOME/bin
    ./ -c clustered.xml \
      -Djboss.default.multicast.address= \ -b PUBLIC_IP_ADDRESS
  8. Start server jdg2. There is a different multicast address, so the jdg1 and jdg2 servers are not directly clustered with each other; rather, they are just connected through the RELAY2 protocol, and the TCP JGroups stack is used for communication between them. The start up command looks like this:

    cd JDG2_HOME/bin
    ./ -c clustered.xml \
      -Djboss.default.multicast.address= \ -b PUBLIC_IP_ADDRESS
  9. To verify that channel works at this point, you may need to use JConsole and connect either to the running JDG1 or the JDG2 server. When you use the MBean jgroups:type=protocol,cluster="cluster",protocol=RELAY2 and operation printRoutes, you should see output like this:

    site1 --> _jdg1:site1
    site2 --> _jdg2:site2

    When you use the MBean jgroups:type=protocol,cluster="cluster",protocol=GMS, you should see that the attribute member contains just single member:

    1. On JDG1 it should be like this:

      (1) jdg1
    2. And on JDG2 like this:

      (1) jdg2
      In production, you can have more {jdgserver_name} servers in every data center. You just need to ensure that {jdgserver_name} servers in same data center are using the same multicast address (In other words, the same jboss.default.multicast.address during startup). Then in jconsole in GMS protocol view, you will see all the members of current cluster.
{project_name} servers setup
  1. Unzip {project_name} server distribution to a location you choose. It will be referred to later as NODE11.

  2. Configure a shared database for KeycloakDS datasource. It is recommended to use MySQL or MariaDB for testing purposes. See Database for more details.

    In production you will likely need to have a separate database server in every data center and both database servers should be synchronously replicated to each other. In the example setup, we just use a single database and connect all 4 {project_name} servers to it.

  3. Edit NODE11/standalone/configuration/standalone-ha.xml :

    1. Add the attribute site to the JGroups UDP protocol:

                        <stack name="udp">
                            <transport type="UDP" socket-binding="jgroups-udp" site="${}"/>
    2. Add this module attribute under cache-container element of name keycloak :

       <cache-container name="keycloak" module="org.keycloak.keycloak-model-infinispan">
    3. Add the remote-store under work cache:

      <replicated-cache name="work">
          <remote-store cache="work" remote-servers="remote-cache" passivation="false" fetch-state="false" purge="false" preload="false" shared="true">
              <property name="rawValues">true</property>
              <property name="marshaller">org.keycloak.cluster.infinispan.KeycloakHotRodMarshallerFactory</property>
    4. Add the remote-store like this under sessions cache:

      <distributed-cache name="sessions" owners="1">
          <remote-store cache="sessions" remote-servers="remote-cache" passivation="false" fetch-state="false" purge="false" preload="false" shared="true">
              <property name="rawValues">true</property>
              <property name="marshaller">org.keycloak.cluster.infinispan.KeycloakHotRodMarshallerFactory</property>
    5. Do the same for offlineSessions, clientSessions, offlineClientSessions, loginFailures, and actionTokens caches (the only difference from sessions cache is that cache property value are different):

      <distributed-cache name="offlineSessions" owners="1">
          <remote-store cache="offlineSessions" remote-servers="remote-cache" passivation="false" fetch-state="false" purge="false" preload="false" shared="true">
              <property name="rawValues">true</property>
              <property name="marshaller">org.keycloak.cluster.infinispan.KeycloakHotRodMarshallerFactory</property>
      <distributed-cache name="clientSessions" owners="1">
          <remote-store cache="clientSessions" remote-servers="remote-cache" passivation="false" fetch-state="false" purge="false" preload="false" shared="true">
              <property name="rawValues">true</property>
              <property name="marshaller">org.keycloak.cluster.infinispan.KeycloakHotRodMarshallerFactory</property>
      <distributed-cache name="offlineClientSessions" owners="1">
          <remote-store cache="offlineClientSessions" remote-servers="remote-cache" passivation="false" fetch-state="false" purge="false" preload="false" shared="true">
              <property name="rawValues">true</property>
              <property name="marshaller">org.keycloak.cluster.infinispan.KeycloakHotRodMarshallerFactory</property>
      <distributed-cache name="loginFailures" owners="1">
          <remote-store cache="loginFailures" remote-servers="remote-cache" passivation="false" fetch-state="false" purge="false" preload="false" shared="true">
              <property name="rawValues">true</property>
              <property name="marshaller">org.keycloak.cluster.infinispan.KeycloakHotRodMarshallerFactory</property>
      <distributed-cache name="actionTokens" owners="2">
          <object-memory size="-1"/>
          <expiration max-idle="-1" interval="300000"/>
          <remote-store cache="actionTokens" remote-servers="remote-cache" passivation="false" fetch-state="false" purge="false" preload="true" shared="true">
              <property name="rawValues">true</property>
              <property name="marshaller">org.keycloak.cluster.infinispan.KeycloakHotRodMarshallerFactory</property>
    6. Add outbound socket binding for the remote store into socket-binding-group element configuration:

      <outbound-socket-binding name="remote-cache">
          <remote-destination host="${}" port="${remote.cache.port:11222}"/>
    7. The configuration of distributed cache authenticationSessions and other caches is left unchanged.

    8. Optionally enable DEBUG logging under the logging subsystem:

      <logger category="org.keycloak.cluster.infinispan">
          <level name="DEBUG"/>
      <logger category="org.keycloak.connections.infinispan">
          <level name="DEBUG"/>
      <logger category="org.keycloak.models.cache.infinispan">
          <level name="DEBUG"/>
      <logger category="org.keycloak.models.sessions.infinispan">
          <level name="DEBUG"/>
  4. Copy the NODE11 to 3 other directories referred later as NODE12, NODE21 and NODE22.

  5. Start NODE11 :

    cd NODE11/bin
    ./ -c standalone-ha.xml \
      -Djboss.default.multicast.address= \ -b PUBLIC_IP_ADDRESS
  6. Start NODE12 :

    cd NODE12/bin
    ./ -c standalone-ha.xml \
      -Djboss.default.multicast.address= \ -b PUBLIC_IP_ADDRESS

    The cluster nodes should be connected. Something like this should be in the log of both NODE11 and NODE12:

    Received new cluster view for channel keycloak: [node11|1] (2) [node11, node12]
    The channel name in the log might be different.
  7. Start NODE21 :

    cd NODE21/bin
    ./ -c standalone-ha.xml \
      -Djboss.default.multicast.address= \ -b PUBLIC_IP_ADDRESS

    It shouldn’t be connected to the cluster with NODE11 and NODE12, but to separate one:

    Received new cluster view for channel keycloak: [node21|0] (1) [node21]
  8. Start NODE22 :

    cd NODE22/bin
    ./ -c standalone-ha.xml \
      -Djboss.default.multicast.address= \ -b PUBLIC_IP_ADDRESS

    It should be in cluster with NODE21 :

    Received new cluster view for channel keycloak: [node21|1] (2) [node21, node22]
    The channel name in the log might be different.
  9. Test:

    1. Go to http://node11:8080/auth/ and create the initial admin user.

    2. Go to http://node11:8080/auth/admin and login as admin to admin console.

    3. Open a second browser and go to any of nodes http://node12:8080/auth/admin or http://node21:8080/auth/admin or http://node22:8080/auth/admin. After login, you should be able to see the same sessions in tab Sessions of particular user, client or realm on all 4 servers.

    4. After doing any change in Keycloak admin console (eg. update some user or some realm), the update should be immediately visible on any of 4 nodes as caches should be properly invalidated everywhere.

    5. Check server.logs if needed. After login or logout, the message like this should be on all the nodes NODEXY/standalone/log/server.log :

      2017-08-25 17:35:17,737 DEBUG [org.keycloak.models.sessions.infinispan.remotestore.RemoteCacheSessionListener] (Client-Listener-sessions-30012a77422542f5) Received event from remote store.
      Event 'CLIENT_CACHE_ENTRY_REMOVED', key '193489e7-e2bc-4069-afe8-f1dfa73084ea', skip 'false'

Administration of Cross DC deployment

This section contains some tips and options related to Cross-Datacenter Replication.

  • When you run the {project_name} server inside a data center, it is required that the database referenced in KeycloakDS datasource is already running and available in that data center. It is also necessary that the {jdgserver_name} server referenced by the outbound-socket-binding, which is referenced from the Infinispan cache remote-store element, is already running. Otherwise the {project_name} server will fail to start.

  • Every data center can have more database nodes if you want to support database failover and better reliability. Refer to the documentation of your database and JDBC driver for the details how to set this up on the database side and how the KeycloakDS datasource on Keycloak side needs to be configured.

  • Every datacenter can have more {jdgserver_name} servers running in the cluster. This is useful if you want some failover and better fault tolerance. The HotRod protocol used for communication between {jdgserver_name} servers and {project_name} servers has a feature that {jdgserver_name} servers will automatically send new topology to the {project_name} servers about the change in the {jdgserver_name} cluster, so the remote store on {project_name} side will know to which {jdgserver_name} servers it can connect. Read the {jdgserver_name} and WildFly documentation for more details.

  • It is highly recommended that a master {jdgserver_name} server is running in every site before the {project_name} servers in any site are started. As in our example, we started both jdg1 and jdg2 first, before all {project_name} servers. If you still need to run the {project_name} server and the backup site is offline, it is recommended to manually switch the backup site offline on the {jdgserver_name} servers on your site, as described in Bringing sites offline and online. If you do not manually switch the unavailable site offline, the first startup may fail or they may be some exceptions during startup until the backup site is taken offline automatically due the configured count of failed operations.

Bringing sites offline and online

For example, assume this scenario:

  1. Site site2 is entirely offline from the site1 perspective. This means that all {jdgserver_name} servers on site2 are off or the network between site1 and site2 is broken.

  2. You run {project_name} servers and {jdgserver_name} server jdg1 in site site1

  3. Someone logs in on a {project_name} server on site1.

  4. The {project_name} server from site1 will try to write the session to the remote cache on jdg1 server, which is supposed to backup data to the jdg2 server in the site2. See Communication details for more information.

  5. Server jdg2 is offline or unreachable from jdg1. So the backup from jdg1 to jdg2 will fail.

  6. The exception is thrown in jdg1 log and the failure will be propagated from jdg1 server to {project_name} servers as well because the default FAIL backup failure policy is configured. See Backup failure policy for details around the backup policies.

  7. The error will happen on {project_name} side too and user may not be able to finish his login.

According to your environment, it may be more or less probable that the network between sites is unavailable or temporarily broken (split-brain). In case this happens, it is good that {jdgserver_name} servers on site1 are aware of the fact that {jdgserver_name} servers on site2 are unavailable, so they will stop trying to reach the servers in the jdg2 site and the backup failures won’t happen. This is called Take site offline .

Take site offline

There are 2 ways to take the site offline.

Manually by admin - Admin can use the jconsole or other tool and run some JMX operations to manually take the particular site offline. This is useful especially if the outage is planned. With jconsole or CLI, you can connect to the jdg1 server and take the site2 offline. More details about this are available in the JDG documentation.

These steps usually need to be done for all the {project_name} caches mentioned in SYNC or ASYNC backups.

Automatically - After some amount of failed backups, the site2 will usually be taken offline automatically. This is done due the configuration of take-offline element inside the cache configuration as configured in {jdgserver_name} server setup.

<take-offline min-wait="60000" after-failures="3" />

This example shows that the site will be taken offline automatically for the particular single cache if there are at least 3 subsequent failed backups and there is no any successful backup within 60 seconds.

Automatically taking a site offline is useful especially if the broken network between sites is unplanned. The disadvantage is that there will be some failed backups until the network outage is detected, which could also mean failures on the application side. For example, there will be failed logins for some users or big login timeouts. Especially if failure-policy with value FAIL is used.

The tracking of whether a site is offline is tracked separately for every cache.
Take site online

Once your network is back and site1 and site2 can talk to each other, you may need to put the site online. This needs to be done manually through JMX or CLI in similar way as taking a site offline. Again, you may need to check all the caches and bring them online.

Once the sites are put online, it’s usually good to:

State transfer

State transfer is a required, manual step. {jdgserver_name} server does not do this automatically, for example during split-brain, it is only the admin who may decide which site has preference and hence if state transfer needs to be done bidirectionally between both sites or just unidirectionally, as in only from site1 to site2, but not from site2 to site1.

A bidirectional state transfer will ensure that entities which were created after split-brain on site1 will be transferred to site2. This is not an issue as they do not yet exist on site2. Similarly, entities created after split-brain on site2 will be transferred to site1. Possibly problematic parts are those entities which exist before split-brain on both sites and which were updated during split-brain on both sites. When this happens, one of the sites will win and will overwrite the updates done during split-brain by the second site.

Unfortunately, there is no any universal solution to this. Split-brains and network outages are just state, which is usually impossible to be handled 100% correctly with 100% consistent data between sites. In the case of {project_name}, it typically is not a critical issue. In the worst case, users will need to re-login again to their clients, or have the improper count of loginFailures tracked for brute force protection. See the {jdgserver_name}/JGroups documentation for more tips how to deal with split-brain.

The state transfer can be also done on the {jdgserver_name} server side through JMX. The operation name is pushState. There are few other operations to monitor status, cancel push state, and so on. More info about state transfer is available in the {jdgserver_name} docs.

Clear caches

After split-brain it is safe to manually clear caches in the {project_name} admin console. This is because there might be some data changed in the database on site1 and because of the event, that the cache should be invalidated wasn’t transferred during split-brain to site2. Hence {project_name} nodes on site2 may still have some stale data in their caches.

To clear the caches, see {adminguide_clearcache_link}[{adminguide_clearcache_name}].

When the network is back, it is sufficient to clear the cache just on one {project_name} node on any random site. The cache invalidation event will be sent to all the other {project_name} nodes in all sites. However, it needs to be done for all the caches (realms, users, keys). See {adminguide_clearcache_name} for more information.

Tuning the JDG cache configuration

This section contains tips and options for configuring your JDG cache.

Backup failure policy

By default, the configuration of backup failure-policy in the Infinispan cache configuration in the JDG clustered.xml file is configured as FAIL. You may change it to WARN or IGNORE, as you prefer.

The difference between FAIL and WARN is that when FAIL is used and the {jdgserver_name} server tries to back data up to the other site and the backup fails then the failure will be propagated back to the caller (the {project_name} server). The backup might fail because the second site is temporarily unreachable or there is a concurrent transaction which is trying to update same entity. In this case, the {project_name} server will then retry the operation a few times. However, if the retry fails, then the user might see the error after a longer timeout.

When using WARN, the failed backups are not propagated from the {jdgserver_name} server to the {project_name} server. The user won’t see the error and the failed backup will be just ignored. There will be a shorter timeout, typically 10 seconds as that’s the default timeout for backup. It can be changed by the attribute timeout of backup element. There won’t be retries. There will just be a WARNING message in the {jdgserver_name} server log.

The potential issue is, that in some cases, there may be just some a short network outage between sites, where the retry (usage of the FAIL policy) may help, so with WARN (without retry), there will be some data inconsistencies across sites. This can also happen if there is an attempt to update the same entity concurrently on both sites.

How bad are these inconsistencies? Usually only means that a user will need to re-authenticate.

When using the WARN policy, it may happen that the single-use cache, which is provided by the actionTokens cache and which handles that particular key is really single use, but may "successfully" write the same key twice. But, for example, the OAuth2 specification mentions that code must be single-use. With the WARN policy, this may not be strictly guaranteed and the same code could be written twice if there is an attempt to write it concurrently in both sites.

If there is a longer network outage or split-brain, then with both FAIL and WARN, the other site will be taken offline after some time and failures as described in Bringing sites offline and online. With the default 1 minute timeout, it is usually 1-3 minutes until all the involved caches are taken offline. After that, all the operations will work fine from an end user perspective. You only need to manually restore the site when it is back online as mentioned in Bringing sites offline and online.

In summary, if you expect frequent, longer outages between sites and it is acceptable for you to have some data inconsistencies and a not 100% accurate single-use cache, but you never want end-users to see the errors and long timeouts, then switch to WARN.

The difference between WARN and IGNORE is, that with IGNORE warnings are not written in the JDG log. See more details in the Infinispan documentation.

Lock acquisition timeout

The default configuration is using transaction in NON_DURABLE_XA mode with acquire timeout 0. This means that transaction will fail-fast if there is another transaction in progress for the same key.

The reason to switch this to 0 instead of default 10 seconds was to avoid possible deadlock issues. With {project_name}, it can happen that the same entity (typically session entity or loginFailure) is updated concurrently from both sites. This can cause deadlock under some circumstances, which will cause the transaction to be blocked for 10 seconds. See this JIRA report for details.

With timeout 0, the transaction will immediately fail and then will be retried from {project_name} if backup failure-policy with the value FAIL is configured. As long as the second concurrent transaction is finished, the retry will usually be successful and the entity will have applied updates from both concurrent transactions.

We see very good consistency and results for concurrent transaction with this configuration, and it is recommended to keep it.

The only (non-functional) problem is the exception in the {jdgserver_name} server log, which happens every time when the lock is not immediately available.

SYNC or ASYNC backups

An important part of the backup element is the strategy attribute. You must decide whether it needs to be SYNC or ASYNC. We have 7 caches which might be Cross-Datacenter Replication aware, and these can be configured in 3 different modes regarding cross-dc:

  1. SYNC backup

  2. ASYNC backup

  3. No backup at all

If the SYNC backup is used, then the backup is synchronous and operation is considered finished on the caller ({project_name} server) side once the backup is processed on the second site. This has worse performance than ASYNC, but on the other hand, you are sure that subsequent reads of the particular entity, such as user session, on site2 will see the updates from site1. Also, it is needed if you want data consistency. As with ASYNC the caller is not notified at all if backup to the other site failed.

For some caches, it is even possible to not backup at all and completely skip writing data to the {jdgserver_name} server. To set this up, do not use the remote-store element for the particular cache on the {project_name} side (file KEYCLOAK_HOME/standalone/configuration/standalone-ha.xml) and then the particular replicated-cache element is also not needed on the {jdgserver_name} server side.

By default, all 7 caches are configured with SYNC backup, which is the safest option. Here are a few things to consider:

  • If you are using active/passive mode (all {project_name} servers are in single site site1 and the {jdgserver_name} server in site2 is used purely as backup. See Modes for more details), then it is usually fine to use ASYNC strategy for all the caches to save the performance.

  • The work cache is used mainly to send some messages, such as cache invalidation events, to the other site. It is also used to ensure that some special events, such as userStorage synchronizations, happen only on single site. It is recommended to keep this set to SYNC.

  • The actionTokens cache is used as single-use cache to track that some tokens/tickets were used just once. For example action tokens or OAuth2 codes. It is possible to set this to ASYNC to slightly improved performance, but then it is not guaranteed that particular ticket is really single-use. For example, if there is concurrent request for same ticket in both sites, then it is possible that both requests will be successful with the ASYNC strategy. So what you set here will depend on whether you prefer better security (SYNC strategy) or better performance (ASYNC strategy).

  • The loginFailures cache may be used in any of the 3 modes. If there is no backup at all, it means that count of login failures for a user will be counted separately for every site (See Infinispan caches for details). This has some security implications, however it has some performance advantages. Also it mitigates the possible risk of denial of service (DoS) attacks. For example, if an attacker simulates 1000 concurrent requests using the username and password of the user on both sites, it will mean lots of messages being passed between the sites, which may result in network congestion. The ASYNC strategy might be even worse as the attacker requests won’t be blocked by waiting for the backup to the other site, resulting in potentially even more congested network traffic. The count of login failures also will not be accurate with the ASYNC strategy.

For the environments with slower network between data centers and probability of DoS, it is recommended to not backup the loginFailures cache at all.

  • It is recommended to keep the sessions and clientSessions caches in SYNC. Switching them to ASYNC is possible only if you are sure that user requests and backchannel requests (requests from client applications to {project_name} as described in Request processing) will be always processed on same site. This is true, for example, if:

    • You use active/passive mode as described Modes.

    • All your client applications are using the {project_name} JavaScript Adapter. The JavaScript adapter sends the backchannel requests within the browser and hence they participate on the browser sticky session and will end on same cluster node (hence on same site) as the other browser requests of this user.

    • Your load balancer is able to serve the requests based on client IP address (location) and the client applications are deployed on both sites.

      For example you have 2 sites LON and NYC. As long as your applications are deployed in both LON and NYC sites too, you can ensure that all the user requests from London users will be redirected to the applications in LON site and also to the {project_name} servers in LON site. Backchannel requests from the LON site client deployments will end on {project_name} servers in LON site too. On the other hand, for the American users, all the {project_name} requests, application requests and backchannel requests will be processed on NYC site.

  • For offlineSessions and offlineClientSessions it is similar, with the difference that you even don’t need to backup them at all if you never plan to use offline tokens for any of your client applications.

Generally, if you are in doubt and performance is not a blocker for you, it’s safer to keep the caches in SYNC strategy.

Regarding the switch to SYNC/ASYNC backup, make sure that you edit the strategy attribute of the backup element. For example like this:
<backup site="site2" failure-policy="FAIL" strategy="ASYNC" enabled="true">

Note the mode attribute of cache-configuration element.


The following tips are intended to assist you should you need to troubleshoot:

  • It is recommended to go through the Basic setup and have this one working first, so that you have some understanding of how things work. It is also wise to read this entire document to have some understanding of things.

  • Check in jconsole cluster status (GMS) and the JGroups status (RELAY) of {jdgserver_name} as described in {jdgserver_name} server setup. If things do not look as expected, then the issue is likely in the setup of {jdgserver_name} servers.

  • For the {project_name} servers, you should see a message like this during the server startup:

    18:09:30,156 INFO  [org.keycloak.connections.infinispan.DefaultInfinispanConnectionProviderFactory] (ServerService Thread Pool -- 54)
    Node name: node11, Site name: site1

    Check that the site name and the node name looks as expected during the startup of {project_name} server.

  • Check that {project_name} servers are in cluster as expected, including that only the {project_name} servers from the same data center are in cluster with each other. This can be also checked in JConsole through the GMS view. See cluster troubleshooting for additional details.

  • If there are exceptions during startup of {project_name} server like this:

    17:33:58,605 ERROR [org.infinispan.client.hotrod.impl.operations.RetryOnFailureOperation] (ServerService Thread Pool -- 59) ISPN004007: Exception encountered. Retry 10 out of 10: org.infinispan.client.hotrod.exceptions.TransportException:: Could not fetch transport
    Caused by: org.infinispan.client.hotrod.exceptions.TransportException:: Could not connect to server:
    	at org.infinispan.client.hotrod.impl.transport.tcp.TcpTransport.<init>(

    it usually means that {project_name} server is not able to reach the {jdgserver_name} server in his own datacenter. Make sure that firewall is set as expected and {jdgserver_name} server is possible to connect.

  • If there are exceptions during startup of {project_name} server like this:

    16:44:18,321 WARN  [org.infinispan.client.hotrod.impl.protocol.Codec21] (ServerService Thread Pool -- 57) ISPN004005: Error received from the server: javax.transaction.RollbackException: ARJUNA016053: Could not commit transaction.

    then check the log of corresponding {jdgserver_name} server of your site and check if has failed to backup to the other site. If the backup site is unavailable, then it is recommended to switch it offline, so that {jdgserver_name} server won’t try to backup to the offline site causing the operations to pass successfully on {project_name} server side as well. See Administration of Cross DC deployment for more information.

  • Check the Infinispan statistics, which are available through JMX. For example, try to login and then see if the new session was successfully written to both {jdgserver_name} servers and is available in the sessions cache there. This can be done indirectly by checking the count of elements in the sessions cache for the MBean jboss.datagrid-infinispan:type=Cache,name="sessions(repl_sync)",manager="clustered",component=Statistics and attribute numberOfEntries. After login, there should be one more entry for numberOfEntries on both {jdgserver_name} servers on both sites.

  • Enable DEBUG logging as described {project_name} servers setup. For example, if you log in and you think that the new session is not available on the second site, it’s good to check the {project_name} server logs and check that listeners were triggered as described in the {project_name} servers setup. If you do not know and want to ask on keycloak-user mailing list, it is helpful to send the log files from {project_name} servers on both datacenters in the email. Either add the log snippets to the mails or put the logs somewhere and reference them in the email.

  • If you updated the entity, such as user, on {project_name} server on site1 and you do not see that entity updated on the {project_name} server on site2, then the issue can be either in the replication of the synchronous database itself or that {project_name} caches are not properly invalidated. You may try to temporarily disable the {project_name} caches as described here to nail down if the issue is at the database replication level. Also it may help to manually connect to the database and check if data are updated as expected. This is specific to every database, so you will need to consult the documentation for your database.

  • Sometimes you may see the exceptions related to locks like this in {jdgserver_name} server log:

    (HotRodServerHandler-6-35) ISPN000136: Error executing command ReplaceCommand,
    writing keys [[B0x033E243034396234..[39]]: org.infinispan.util.concurrent.TimeoutException: ISPN000299: Unable to acquire lock after
    0 milliseconds for key [B0x033E243034396234..[39] and requestor GlobalTx:jdg1:4353. Lock is held by GlobalTx:jdg1:4352

    Those exceptions are not necessarily an issue. They may happen anytime when a concurrent edit of the same entity is triggered on both DCs. This is common in a deployment. Usually the {project_name} server is notified about the failed operation and will retry it, so from the user’s point of view, there is usually not any issue.

  • If there are exceptions during startup of {project_name} server, like this:

    16:44:18,321 WARN  [org.infinispan.client.hotrod.impl.protocol.Codec21] (ServerService Thread Pool -- 55) ISPN004005: Error received from the server: java.lang.SecurityException: ISPN000287: Unauthorized access: subject 'Subject with principal(s): []' lacks 'READ' permission

    These log entries are the result of {project_name} automatically detecting whether authentication is required on {jdgserver_name} and mean that authentication is necessary. At this point you will notice that either the server starts successfully and you can safely ignore these or that the server fails to start. If the server fails to start, ensure that {jdgserver_name} has been configured properly for authentication as described in {jdgserver_name} server setup. To prevent this log entry from being included, you can force authentication by setting remoteStoreSecurityEnabled property to true in spi=connectionsInfinispan/provider=default configuration:

    <subsystem xmlns="urn:jboss:domain:keycloak-server:1.1">
        <spi name="connectionsInfinispan">
            <provider name="default" enabled="true">
                    <property name="remoteStoreSecurityEnabled" value="true"/>
  • If you try to authenticate with {project_name} to your application, but authentication fails with an infinite number of redirects in your browser and you see the errors like this in the {project_name} server log:

    2017-11-27 14:50:31,587 WARN  [] (default task-17) type=LOGIN_ERROR, realmId=master, clientId=null, userId=null,, error=expired_code, restart_after_timeout=true

    it probably means that your load balancer needs to be set to support sticky sessions. Make sure that the provided route name used during startup of {project_name} server (Property contains the correct name used by the load balancer server to identify the current server.

  • If the {jdgserver_name} work cache grows indefinitely, you may be experiencing this {jdgserver_name} issue, which is caused by cache items not being properly expired. In that case, update the cache declaration with an empty <expiration /> tag like this:

        <replicated-cache name="work" configuration="sessions-cfg">
            <expiration />
  • If you see Warnings in the {jdgserver_name} server log like:

    18:06:19,687 WARN  [org.infinispan.server.hotrod.Decoder2x] (HotRod-ServerWorker-7-12) ISPN006011: Operation 'PUT_IF_ABSENT' forced to
      return previous value should be used on transactional caches, otherwise data inconsistency issues could arise under failure situations
    18:06:19,700 WARN  [org.infinispan.server.hotrod.Decoder2x] (HotRod-ServerWorker-7-10) ISPN006010: Conditional operation 'REPLACE_IF_UNMODIFIED' should
      be used with transactional caches, otherwise data inconsistency issues could arise under failure situations

    you can just ignore them. To avoid the warning, the caches on {jdgserver_name} server side could be changed to transactional caches, but this is not recommended as it can cause some other issues caused by the bug So for now, the warnings just need to be ignored.

  • If you see errors in the {jdgserver_name} server log like:

    12:08:32,921 ERROR [org.infinispan.server.hotrod.CacheDecodeContext] (HotRod-ServerWorker-7-11) ISPN005003: Exception reported: org.infinispan.server.hotrod.InvalidMagicIdException: Error reading magic byte or message id: 7
    	at org.infinispan.server.hotrod.HotRodDecoder.readHeader(
    	at org.infinispan.server.hotrod.HotRodDecoder.decodeHeader(
    	at org.infinispan.server.hotrod.HotRodDecoder.decode(
    	at io.netty.handler.codec.ByteToMessageDecoder.callDecode(
    	at io.netty.handler.codec.ByteToMessageDecoder.channelRead(

    and you see some similar errors in the {project_name} log, it can indicate that there are incompatible versions of the HotRod protocol being used. This is likely happen when you try to use {project_name} with the JDG 7.2 server or an old version of the Infinispan server. It will help if you add the protocolVersion property as an additional property to the remote-store element in the {project_name} configuration file. For example:

    <property name="protocolVersion">2.6</property>