| ============================== |
| Next actions |
| ============================== |
| |
| - Membership coordinator logic and membership view Id. |
| |
| - Implement primary/secondary replication logic (fhanik) |
| Idea for implementation: Cookie stores info primary/backup node and hit counter. |
| Steps to implement: (coming soon) |
| Scenarios: |
| 1. Request in: new session dest: any |
| 2. Request in: existing session dest: primary server |
| 3. Request in: existing session dest: non-primary server (fail over) |
| 4. Request in: existing session dest: backup server |
| 5. Backup node crash |
| 6. Primary node crash (step 3) |
| 7. Sequence 6. then 2. (hit counter) |
| 8. Sequence 5. then 4. (swap) |
| 9. New member joins |
| 10. Backup and primary crash -> death! |
| 11. Sequence 2. in: invalid hit counter |
| 12. Sequence 4. in: invalid hit counter |
| 13. Session expires |
| 14. Sequence 3. then 13. on the old primary |
| |
| Notes: |
| Ability to add a custom "Session expired" redirect for global |
| redirects to never forward an invalid session into the server |
| when the session data has gone lost. |
| Ie, in a normal scenario, the session is viewed as "new" |
| but they may wanna display a session expired redirect. |
| Security: |
| Dont store any sensitive info in the new cookie. |
| only pointers to members. |
| |
| |
| - break out membership/messaging module and improve JMX |
| The clustering code has become very cluttered. |
| We need to break out the membership and the messaging |
| away from the logic of session handling and JMX. |
| The latter should be built on top of the clustering module, and not baked into it. |
| (maybe complete refactor for Tomcat 6) |
| Steps: |
| 1. Build a group object, to allow sending of messages |
| The group object will handle encompass membership and |
| messaging under one umbrella. |
| 2. Enable group RPC |
| Through the group object, enable RPC style messaging |
| This should be used for message like state transfers, instead of waiting for a timeout. |
| RCP strategies - first response, majority response, all response |
| 3. Group/Message interceptors - to perform logic such as fragmentation, ordering etc. |
| 4. JMX |
| JMX should be added on top of the core code, not inside the core code. |
| The idea of JMX as management should be added outside to the core components. |
| For example, SimpleTCPCluster should not contain JMX code, if I want to create |
| a different cluster class, I have to rewrite all the JMX code again. |
| Instead, create JMX beans that reference core components through interfaces. |
| |
| - Create stats object |
| Instead of keeping stats embedded in the code, make it reusable in |
| a shared component. |
| |
| - support CrossContext session replication |
| Idea: use endAccess to signal replicationValve that session from other app has changed. |
| refactor ReplicationValve |
| When session is serialzed from other app, set right app classloader |
| endaccess is also send at CoyoteAdpater for context session as recycle the request |
| - Add MbeanFactory to generate dynamic cluster at runtime. |
| Problem: How we can start those central services? |
| - StandardEngine support load Mbean from external file. |
| - when a lot of messages expire it comes to burst of messages |
| - all 60 Sec when ManagerBase#processExpires is called a lot of messages are send! |
| - Better is to transfer a spezial epxire message with an array of expired session messages. |
| - This reduce message transfer and reduce waits for acks. |
| - complicated implementation thing , sessions expires when call isValid :-( |
| - build a tool that receive the stage and present it as simple web app. |
| detect some problems |
| different active sessions |
| long queues |
| detect long wait acks |
| Idea: Wrote a ant script with the new jmx tasks! |
| Optimzied information access: |
| Create a MBean.attribute that store the complete cluster state |
| information inside one attribute from type TabularData and CompositeData! |
| Implement as SimpleTcpCluster operation |
| Cluster <CompositeData> |
| -- Attributes <SimpleType> |
| -- ClusterMembership |
| -- Attributes <SimpleType> |
| -- Members <CompositeType> |
| -- Attributes <SimpeType> |
| -- ClusterReceiver |
| -- Attributes <SimpleType> |
| -- ClusterSender |
| -- Attributes <SimpleType> |
| -- DataSenders <TabularData> |
| -- Sender <CompositeType> |
| -- Attributes <SimpeType> |
| -- (optional)Queue Stats <CompositeData> |
| -- Attributes <SimpeType> |
| - add cluster setup template (src) |
| - documentation |
| wrote a complete new how-to |
| add example configurations |
| add complete attribute descriptions |
| add JMX information |
| add deployment help |
| Need help (pero) |
| - implement fragmentation of large replication objects |
| (note this will be done in the Group object, see top.) |
| Compress at message level - this will go away! instead there will be a GZip interceptor |
| Splitting Messages ala FarmDeployer war handling |
| - add a message type to the message header. |
| - filtering at receiver that drop message before build Object |
| - define short type definition. |
| - Every Session message with differenz type |
| - type with String |
| - add test cluster project |
| functional testing a lot of szenarios |
| differen replication mode |
| restart after failure |
| crash under load |
| compress and uncompressed |
| Junit test ( started) |
| automated regression testing with some standard configs |
| wrote a test client that can get a JMX State to verify the different cluster testszenarios |
| - direct JSR 160 client - preferred option |
| - via mx4j http adaptor ( xml protocol) |
| - setup different szenarios |
| use new remote jmx ant tasks to grab information from mbeans |
| - filter all cluster messages |
| - Filtering that no all message send to all member. |
| - Now with domain mode registeration session message only send to |
| members from same domain. |
| - setup a Cluster LifecycleListener to send the cluster status! (monitoring) |
| - Make the compact state via JMX API avialable first. |
| - setup cluster Listener that send own state to spezial member (sender from a message like GET_ALL_SESSIONS) |
| - create a cluster status app (html/xml) |
| - attributes |
| - send message |
| - current members |
| - receive message |
| - active session at different clustered manager |
| - which app are clusters (registered managers) |
| - avg processing times |
| |
| - operations |
| - resetStats |
| - send a message (String) |
| - getDisplay state from other members?? |
| - stop queue to send |
| - queue receiver message from some members ( later send after redeploying) |
| - configure some send parameters |
| - keepalive |
| - compress |
| - wait for ack |
| - other things |
| - based on Cluster JMX API |
| - watch some values from complete cluster and display some graphs |
| - display the informations from all nodes |
| - display informations from other cluster domains |
| via XML documents and http |
| - display stats as xml |
| - operation via JMX (MX4J adaptor) |
| |
| ================ |
| problems |
| ================ |
| - How we can stop the request traffic when restart an application? |
| currently the jk 1.2.10 can only disable the complete loadbalancer, |
| but this detect only the new session request desicion. |
| Request with sessions marks send to tomcat. |
| Fix: jk > 1.2.11 has a stopped flag, but then all application stop traffic |
| and session transfer from other nodes not stopped!! |
| Stop MembershipService via JMX and after redeploying application, |
| start MembershipService and restart the application Session Manager or send GETALL Session from backup? |
| then enable apache worker again !! :-( |
| easier stop tomcats, deploy new app and start again... |
| - Can't stop message replication for a spezial member and application |
| - this need a spezial cluster message and send filter at SimpleTcpCluster |
| - Don't generate cluster message when no member is at cluster! |
| - Register DeltaManager as Cluster LifecycleListener and stop cresting and sending |
| - Reduce memory consume when only local node is active |
| - Important feature when nodes crashed, and only one server exists under load... |
| |
| ============================== |
| Nice to have: |
| ============================== |
| - Replication ContextAttributes |
| - Configure the McastService no accept every member. |
| - receive a secret key |
| - have a allow or deny list like RequestFilterValve |
| - Also receiver don't accept request from not allowed members |
| - ReplicationListener and SocketReplicationListener only accept data from cluster member (low level ip restriction) |
| - PooledSocketSender |
| Add more stats |
| check all Pooled sender checkKeepAlive |
| - Implement a NonSerializable interface for session attributes that do not |
| wish to be replicated |
| Then we must have ClusterNonSerializable at common classloader |
| - Extend StandardSession if possible |
| - Implement context attribute replication (?) |
| pero: |
| Also send Start/Stop messages from Context to complete cluster! |
| With 5.5.10 you can wrote a Cluster Lifecycle Manager that do this. |
| Register for AFTER_MANAGERREGISTER_EVENT and AFTER_MANAGERUNREGISTER_EVENT |
| and also to the context ServletContextAttributeEvent Listener |
| Access the Context |
| ((Manager)event.getData()).getContainer() |
| - Fix farm deployment for 5.5 |
| pero: |
| Every start all application are deployed only to all running cluster nodes |
| New registered nodes don't get the applications! |
| Deploy must send a GET_ALL_APP to all other Deployers. |
| FH: Correction, you should not send GET_ALL_APP to all deployers, only |
| to the main one. Which could be another property of the Member object, it |
| would not make sense to transfer the same webapp over and over again. |
| Only watchEnabled Deployer send this member all deployed application. |
| pero: |
| Yes very true, but currently we distribute also all wars from watch node at begining. |
| Waiting to start other nodes is only change to not got these war's! |
| I have made some experiments to register war deployment at new memberAdded to cluster. |
| Add JMX Support |
| Resend Deployed Applications to all or one cluster node. |
| Show all watch Resource |
| Processing Time |
| Change fileMessage Buffersize. |
| Start/Stop Cluster wide application |
| Deployer and Watcher sync with engine background thread! |
| Fixed! |
| Last FileMessage fragment need longe ackTimeout |
| <Cluster ..> <Sender ... ackTimeout="60000"/> </Cluster> |
| - Change the cluster protocol that developer can add there own data serialzable/deserialzable format (high risk) |
| Note from Filip: Lets not do this, last attempt caused a broken clustering piece for 4 versions. |
| Currently |
| header 6 bytes (FLT2002) |
| compressflag 4 byte |
| data.length 4 bytes |
| data, |
| end header 6 bytes (TLF2003) |
| |
| Optimized to |
| header 2 bytes (TC) |
| type 1 byte |
| compressflag 1 byte |
| data.length 4 bytes, |
| data | <real uncompressed data.length (4 bytes)> data |
| |
| "type" means user defined type and receiver extract bytes and type and sende it to callback |
| s. ObjectReader or SocketObjectReader |
| |
| compress 1 |
| first data 4 data bytes are the real uncompressed data length. ( Is for better memory management atr recevier side, S. XByteBuffer) |
| |
| change at DataSender.writeData and XByteBuffer and add flexible handling to ClusterSender and ClusterReceiver |
| |
| - Add single sign on support |
| |
| ============================== |
| COMPLETED |
| ============================== |
| 5.5.16 (fhanik) |
| - Extensive logging when a member crashes |
| When a member crashes and you use pooled socket mode then |
| the logs are spewing out info. |
| Instead, now we only report one IOException if it happens, and |
| mark the member as suspect. |
| |
| 5.5.16 (fhanik) |
| - Average message size |
| For my test application, the average message size has increased from |
| 526 bytes per message to over 750 bytes per message from version 5.0.28 to 5.5.15. |
| Need to investigate why this is the case and where the 200+ extra bytes are coming from. |
| Fix: The default compress value had changed. The new default is that no compression is used. |
| To re-enable compression, <Sender compress="true" in the sender element. |
| |
| 5.5.15 (pero) |
| - Optimized getMembers access from McastMembership |
| Member list was copied at every access. (Used inside ReplicationValve.invoke) |
| SimpleTcpCluster need a alive time sorted array |
| |
| 5.5.11 (pero) |
| - MemoryUser principal from UserDatabaseRealm not handled to replicated |
| - look inside DeltaRequest.setPrincipal(Principal,boolean) |
| detected by Dirk de Kok (tomdev 16.8.2005) |
| - only GenericPrincipal from all other realms are handled well. |
| - change JAASRealm and UserDatabaseRealm that also used GenericPrincipal |
| 5.5.10 (pero) |
| - Cluster config at engine level (user request 06/05) |
| Register a cluster infrastructur for many vhosts |
| configure backup systems! |
| Add Cluster Element to digister |
| - add mapping sender mapping properties file (IDataSenderFactory) |
| - let advanced people eaiser implemented there own sender mode |
| - We register different application with same name from different host? |
| SimpleTcpManager register manager with app name + hostname when Cluster is configured as Engine element. |
| - Configured DeltaManager inside context |
| - SimpleTcpCluster setProperty and transferproperty reflect changes only to defaultMode managers |
| - Look inside SimpleTcpCluster.addManager and DeltaManager.start? |
| - Session serialization eat memory but now we can send session messages with blocks... |
| When all sessions serialze after GET_ALL_SESSION is received following works |
| - find all sessions |
| - serialize a block or all sessions as byte array |
| - serialize the complete SessionMessageImpl to transfer message |
| - WaitForAck mode and resend probleme |
| - Now message creator can configure resend and compress mode! |
| - Add a default simple cluster config with good defaults and only |
| one cluster element inside server.xml. Setup with fastasyncmode. |
| Service Elements |
| ReplicationTransmitter, |
| SocketReplicationListener, |
| McastService, |
| ClusterSessionListener |
| ReplicationValve |
| You can change property setting with SimpleTcpCluster prefix "sender.XXX, receiver.XXX, valve.XXX, listener.XXX, service.XXX" |
| - Fix resend GET_ALL_SESSIONS when wait ACK failed at receiver side |
| - Fix that ClusterValve not remove when cluster stops |
| - Set timestamp only at first time inside SessionMessageImpl |
| - Set timestamp from findsessions when handling GET_ALL_SESSION |
| - Set this timestamp to all SEND_SESSION_DATA and TRANSFER Complete messages |
| - Drop all received message inside GET_ALL_SESSION message queue (DeltaManager) |
| |
| - Mcast Service as JMX MBean (change cluster domain at runtime) |
| - send cluster domain with mcast ping |
| - With sendClusterDomainOnly=true only session message from same domain are received |
| - Session only replicated to members from same domain, with sendClusterDomainOnly=false |
| at Sender (ReplicationTransmitter) session messages send to all members. |
| - GET ALL Session send to first member inside same cluster domain |
| - better restart szenario at DeltaManager after failure restart (java service wrapper). |
| queue all other session events |
| as STATE Transfer Complete is received, dequeue all received sessions messages. |
| - restructure methods at DeltaManager |
| - extract handleXXS methods for better DeltaManager subclassing. |
| - split big get all sessions from one server into blocks of sessions and separate STATE Transfer message! |
| - no complete sync sessions when GET ALL Sessions event is received. |
| - add JMX API for ClusteRreceiver |
| - ClusterReceiver is now Callback when message is received |
| - SimpleTcpCluster only receive ClusterMessage (API change) |
| Redesign SimpleTcpCluster message receiving to ClusterReceiverBase: |
| - optimized data uncompressed |
| - better extendablity |
| - XByteBuffer only buffer bytes and don't uncompress. |
| - Add receiver JMX stats with new attribute 'Receiver.doReceivedProcessingStats' |
| - optimized createManager and addManager that also can configured normal StandardManager |
| to use cluster message transfer without replication. |
| - add support to dynamic property transfer from SimpleTcpCluster to the Manager |
| like ReplicationTransmitter |
| All manager attributes can be configured: |
| - expireSessionsOnShutdown (false) |
| - notifySessionListenersOnReplication (true) |
| - notifyListenersOnReplication (true) |
| - maxActiveSessions (-1) |
| - timeoutAllSession (60 sec) |
| - sessionIdLength (16) |
| - processExpiresFrequency (6 - exipre all six engine background periodic event (60 sec)) |
| - algorithm (MD5) |
| - entropy |
| - randomFile |
| - randomClass |
| - Setup the cluster without SessionReplication Manager |
| Only a message bus. |
| Configure the bus with you ClusterListener and valves and a StandardManager |
| - reduce cpu and memory consume (Receiver) |
| - set new compress sender flag at default=false ( < CPU usage) |
| - Make compact algo |
| currently message receive data is split at XbyteBuffer#extractPackage and |
| SimpleTcpCluster#messageDataReceived |
| - reduce memory and cpu consume (send message) |
| - set new compress sender flag at default=false ( < CPU usage) |
| - don't copy the buffer to add message header |
| transfer this from SimpleTcpCluster to DataSender pushMessage |
| successfull refactored |
| - make it possible that a subclass crypt the transfered messages |
| sub class ReplicationTransmitter and override createMessageData |
| - don't copy START and END Header for every message, instead send dirctly and DataSender.writeData. |
| - Add a flag for replicated attribute events, to enable or disable them |
| Now can configued with notifyListenersOnReplication=false at SimpleTCPCluster |
| Also can drop HttpSessionLsitener events |
| can configued with notifySessionListenersOnReplication=false at SimpleTCPCluster |
| - Refactoring DeltaManager |
| - Transfer attributes from Cluster config to DeltaManager |
| - Fix at 5.5.9 cluster hang bug! |
| - Add more Valve to direct cluster config |
| - Add Lifecycle Listener support to direct cluster config |
| - Add ClusterListener support to direct cluster config |
| - Add new SocketReplicationListener |
| - Add Stats to DeltaManager |
| |
| 5.5.9 (pero) |
| - JMX friendly |
| pero: Add some MBeanSupport to SimpleTCPCluster, ReplicationTransmitter and Senders |
| - Add Keep Alive and WaitForAck at async mode implementation. |
| Make this feature configurable to Sender element at server.xml |
| Is include with 5.5.8 |
| - Add support to new Async Mode from Rainer Jung |
| Integrated with 5.5.9 |
| fastasyncqueue mode |
| |
