From: "Scott Rossillo" <srossillo@smartling.com>
To: "Marek Posolda" <mposolda@redhat.com>, afield@redhat.com
Cc: "keycloak-user" <keycloak-user@lists.jboss.org>, "Bill Burke" <bburke@redhat.com>
Sent: Monday, December 14, 2015 6:31:30 PM
Subject: Re: [keycloak-user] Replace use of Infinispan with User Sessions SPI ?There are two issues:1. Infinispan relies on JGroups, which is difficult to configure correctly with the various ping techniques that aren’t UDP multicast. I can elaborate on each one that we tested but it’s just generally complex to get right. That’s not to say it’s impossible or the biggest reason this is complicated on ECS or _insert container service here_, see #2 for that.
<stack name="s3">
<transport type="TCP" socket-binding="jgroups-tcp"/>
<protocol type="S3_PING">
<property name="location">${jgroups.s3.bucket:}</property>
<property name="access_key">${jgroups.s3.access_key:}</property>
<property name="secret_access_key">${jgroups.s3.secret_access_key:}</property>
<property name="pre_signed_delete_url">${jgroups.s3.pre_signed_delete_url:}</property>
<property name="pre_signed_put_url">${jgroups.s3.pre_signed_put_url:}</property>
<property name="prefix">${jgroups.s3.prefix:}</property>
</protocol>
<protocol type="MERGE3"/>
<protocol type="FD_SOCK" socket-binding="jgroups-tcp-fd"/>
<protocol type="FD_ALL"/>
<protocol type="VERIFY_SUSPECT"/>
<protocol type="pbcast.NAKACK2">
<property name="use_mcast_xmit">false</property>
</protocol>
<protocol type="UNICAST3"/>
<protocol type="pbcast.STABLE"/>
<protocol type="pbcast.GMS"/>
<protocol type="MFC"/>
<protocol type="FRAG2"/>
</stack>
With this in the configuration file, you can start the server with the following system properties defined:
bin/clustered.sh -Djboss.node.name=node0 -Djboss.socket.binding.port-offset=0 -Djboss.default.jgroups.stack=s3 -Djgroups.s3.bucket=<s3_bucket_name> -Djgroups.s3.access_key=<access_key> -Djgroups.s3.secret_access_key=<secret_access_key> -Djboss.bind.address=$IP -Djboss.bind.address.management=$IP
This will cause the server to start and the nodes will write to a file in the S3 bucket to allow the nodes to discover each other. I do not see this stack defined in the configuration used by WildFly 9, but it should work there as well. It is also possible to use the JGroups Gossip Router for discovery, but it requires running a separate process that all of the nodes contact during the discovery phase. I have the following in my .bashrc to set the IP environment variable:
export IP=`GET http://169.254.169.254/latest/meta-data/local-ipv4`
I have verified that I can cluster plain EC2 instances on the internal IP addresses: (i.e. 172.31.4.165 and 172.31.18.207) These addresses are not publically accessible though, but this cluster can be a cache for applications running in EC2.
2. It is difficult to do discovery correctly with JGroups and Docker. Non-privileged Docker instances - the default and recommend type - do not implicitly know their host’s IP. This causes IP mismatches between what JGroups thinks the machine’s IP is and what it actually is when connecting to hosts on different machines. This is the main issue and it’s not the fault of JGroups per se, but there’s no simple work around.Take for example a simple 2 node cluster:Node 1 comes up on the docker0 interface of host A with the IP address 172.16.0.4. The host A IP is 10.10.0.100.Node 2 comes up on the docker0 interface of host B with the IP address 172.16.0.8. The host B IP is 10.10.0.108.The 172.16 network is not routable between hosts (by design). Docker does port forwarding for ports we wish to expose to this works fine for HTTP/HTTPS but not the cluster traffic.
So Node 1 will advertise itself as having IP 172.16.0.4 while Node 2 advertises 172.16.0.8. The two cannot talk to each other by default. However, using the hard coded IPs and TCP PING, we can set external_addr on Node 1 to 10.10.0.100 and external_addr on Node 2 to 10.10.0.108 and set initial_hosts to 10.10.0.100, 10.10.0.108. This will cause the nodes to discover each other. However, they will not form a cluster. The nodes will reject the handshake thinking they’re not actually 10.10.0.100 or 10.10.0.108 respectively.I’d like to discuss further and I can share where we’ve gotten so far with workarounds to this but it may be better to get into the weeds on another list.Let me know what you think.
Best,ScottOn Dec 14, 2015, at 5:32 PM, Marek Posolda <mposolda@redhat.com> wrote:CCing Alan Field from RH Infinispan team and forwarding his question:I'd like to know which configuration files you are using and why is is harder to use with Amazon’s Docker service (ECS) or Beanstalk. I'd also be interested in how big a cluster you are using in AWS.
On 14/12/15 22:24, Scott Rossillo wrote:AWS was why we didn’t use Infinispan to begin with. That and it’s even more complicated when you deploy using Amazon’s Docker service (ECS) or Beanstalk.It’s too bad Infinispan / JGroups are beasts when the out of the box configuration can’t be used. I’m planning to document this as we fix but I’d avoid S3_PING and use JDBC_PING. You already need JDBC for the Keycloak DB, unless you’re using Mongo and it’s easier to test locally.TCPPING will bite you on AWS if Amazon decides to replace one of your instances (which it does occasionally w/ECS or Beanstalk).Best,ScottOn Dec 14, 2015, at 10:59 AM, Marek Posolda <mposolda@redhat.com> wrote:On 14/12/15 16:55, Marek Posolda wrote:On 14/12/15 15:58, Bill Burke wrote:Btv. if TCPPING or S3_PING is an issue, there is also AWS_PINGOn 12/14/2015 5:01 AM, Niko Köbler wrote:+1Hi Marek,seems to me it would be much easier to get Infinispan working on AWSAm 14.12.2015 um 08:50 schrieb Marek Posolda <mposolda@redhat.comMy current customer is running his apps in AWS. As known, multicast is
<mailto:mposolda@redhat.com>>:
Btv. what's your motivation to not use infinispan? If you afraid of
cluster communication, you don't need to worry much about it, because
if you run single keycloak through standalone.xml, the infinispan
automatically works in LOCAL mode and there is no any cluster
communication at all.
not available in cloud infrastructures. Wildfly/Infinispan Cluster works
pretty well with multicast w/o having to know too much about JGroups
config. S3_PING seams to be a viable way to get a cluster running in AWS.
But additionally, my customer doesn’t have any (deep) knowledge about
JBoss infrastructures and so I’m looking for a way to be able to run
Keycloak in a cluster in AWS without the need to build up deeper
knowlegde of JGroups config, for example in getting rid of Infinispan.
But I do understand all the concerns in doing this.
I still have to test S3_PING, if it works as easy as multicast. If yes,
we can use it, if no… I don’t know yet. But this gets offtopic for
Keycloak mailinglist, it’s more related to pure Wildfly/Infinispan.
than to write and maintain an entire new caching mechanism and hope we
don't refactor the cache SPI.
I am sure infinispan/JGroups has possibility to run in non-multicast
environment. You may just need to figure how exactly to configure it. So
I agree that this issue is more related to Wildfly/Infinispan itself
than to Keycloak.
You may need to use jgroups protocols like TCP instead of default UDP
and maybe TCPPING (this requires to manually list all your cluster
nodes. But still, it's much better option IMO than rewriting UserSession
SPI)
http://www.jgroups.org/manual-3.x/html/protlist.html#d0e5100 , but it's
not official part of jgroups.
Marek
Marek
_______________________________________________
keycloak-user mailing list
keycloak-user@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/keycloak-user
_______________________________________________
keycloak-user mailing list
keycloak-user@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/keycloak-user