[infinispan-issues] [JBoss JIRA] (ISPN-1586) inconsistent cache data in replication cluster with local (not shared) cache store
Dan Berindei (Commented) (JIRA)
jira-events at lists.jboss.org
Thu Dec 8 14:55:40 EST 2011
[ https://issues.jboss.org/browse/ISPN-1586?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12649406#comment-12649406 ]
Dan Berindei commented on ISPN-1586:
------------------------------------
{quote}To enforce purgeOnStartup on when fetchPersistentState is enabled will NOT work for use cases where cache stores are used as a data persistent mechanism (as in my case).
But, you might want to do purgeOnStartup ONLY on nodes (not the first node in the cluster) joining with other nodes. In this case, the data from the first node's cache store is loaded to cache and used as the "golden" copy. Indeed, this was the approach I first tried to work around the problem.{quote}
You're right, it would be reasonable to load from the cache store only on the first node to start. Actually that was the behaviour before ISPN-1171 (see ISPN-370). It could probably work the same in replicated mode and in distributed mode.
Note that sometimes the two nodes can start as two separate clusters and then merge to form a single cluster. JGroups 3.0.1 significantly reduces the possibility of a merge during startup, but it doesn't eliminate it completely. So until we handle merges better (ISPN-263), there will always be a chance of stale/missing entries.
If you have the option, I would recommend using a shared cache store for persistence instead.
{quote}Regrading your comment on holding/passing state to cache loaders, we already holde the whole state in StateTransferManager right now.{quote}
No, we don't: even in replicated mode, a joiner will receive state from all the existing members, and {{StateTransferManager.applyState(state)}} will be called multiple times with different state information. With chunked state transfer even the state from a single existing member will arive as multiple {{applyState}} calls.
The state received in those calls is directly applied to the data container (and indirectly saved to the cache store). But we don't keep a copy of that state in {{StateTransferManager}} to use during future {{applystate}} calls.
> inconsistent cache data in replication cluster with local (not shared) cache store
> ----------------------------------------------------------------------------------
>
> Key: ISPN-1586
> URL: https://issues.jboss.org/browse/ISPN-1586
> Project: Infinispan
> Issue Type: Bug
> Components: Core API
> Affects Versions: 5.0.0.FINAL, 5.1.0.CR1
> Environment: ISPN 5.0.0. Final and ISPN 5.1 sanpshot
> Java 1.7
> Linux Cent OS
> Reporter: dex chen
> Assignee: Dan Berindei
> Fix For: 5.1.0.CR2, 5.1.0.FINAL
>
>
> I rerun my test (an embedded ISPN cluser) with ISPN 5.0.0. final and 5.1 Sanpshot code.
> It is configured in "replication", using local cache store, and preload=true, purgeOnStartup=false .. (see the whole config below).
> I will get the inconsistent data among the nodes in the following scenario:
> 1) start 2 node cluster
> 2) after the cluster is formed, add some data to the cache
> k1-->v1
> k2-->v2
> I will see the data replication working perfectly at this point.
> 3) bring node 2 down
> 4) delete entry k1-->v1 through node1
> Note: At this point, on the local (persistent) cache store on the node2 have 2 entries.
> 5) start node2, and wait to join the cluster
> 6) after state merging, you will see now that node1 has 1 entry and nod2 has 2 entries.
> I am expecting that the data should be consistent across the cluster.
> Here is the infinispan config:
> <infinispan
> xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
> xsi:schemaLocation="urn:infinispan:config:5.0 http://www.infinispan.org/schemas/infinispan-config-5.0.xsd"
> xmlns="urn:infinispan:config:5.0">
> <global>
> <transport clusterName="demoCluster"
> machineId="node1"
> rackId="r1" nodeName="dexlaptop"
> >
> <properties>
> <property name="configurationFile" value="./jgroups-tcp.xml" />
> </properties>
> </transport>
> <globalJmxStatistics enabled="true"/>
> </global>
> <default>
> <locking
> isolationLevel="READ_COMMITTED"
> lockAcquisitionTimeout="20000"
> writeSkewCheck="false"
> concurrencyLevel="5000"
> useLockStriping="false"
> />
> <jmxStatistics enabled="true"/>
> <clustering mode="replication">
> <stateRetrieval
> timeout="240000"
> fetchInMemoryState="true"
> alwaysProvideInMemoryState="false"
> />
> <!--
> Network calls are synchronous.
> -->
> <sync replTimeout="20000"/>
> </clustering>
> <loaders
> passivation="false"
> shared="false"
> preload="true">
> <loader
> class="org.infinispan.loaders.jdbc.stringbased.JdbcStringBasedCacheStore"
> fetchPersistentState="true"
> purgeOnStartup="false">
> <!-- set to true for not first node in the cluster in testing/demo -->
> <properties>
> <property name="stringsTableNamePrefix" value="ISPN_STRING_TABLE"/>
> <property name="idColumnName" value="ID_COLUMN"/>
> <property name="dataColumnName" value="DATA_COLUMN"/>
> <property name="timestampColumnName" value="TIMESTAMP_COLUMN"/>
> <property name="timestampColumnType" value="BIGINT"/>
> <property name="connectionFactoryClass" value="org.infinispan.loaders.jdbc.connectionfactory.PooledConnectionFactory"/>
> <property name="connectionUrl" value="jdbc:h2:file:/var/tmp/h2cachestore;DB_CLOSE_DELAY=-1"/>
> <property name="userName" value="sa"/>
> <property name="driverClass" value="org.h2.Driver"/>
> <property name="idColumnType" value="VARCHAR(255)"/>
> <property name="dataColumnType" value="BINARY"/>
> <property name="dropTableOnExit" value="false"/>
> <property name="createTableOnStart" value="true"/>
> </properties>
> <!--
> <async enabled="false" />
> -->
> </loader>
> </loaders>
> </default>
> </infinispan>
> Basically, current ISPN implementation in state transfer will result in data insistence among nodes in replication mode and each node has local cache store.
> I found code BaseStateTransferManagerImpl's applyState code does not remove stale data in the local cache store and result in inconsistent data when joins a cluster:
> Here is the code snipt of applyState():
> public void applyState(Collection<InternalCacheEntry> state,
> Address sender, int viewId) throws InterruptedException {
> .....
>
> for (InternalCacheEntry e : state) {
> InvocationContext ctx = icc.createInvocationContext(false, 1);
> // locking not necessary as during rehashing we block all transactions
> ctx.setFlags(CACHE_MODE_LOCAL, SKIP_CACHE_LOAD, SKIP_REMOTE_LOOKUP, SKIP_SHARED_CACHE_STORE, SKIP_LOCKING,
> SKIP_OWNERSHIP_CHECK);
> try {
> PutKeyValueCommand put = cf.buildPutKeyValueCommand(e.getKey(), e.getValue(), e.getLifespan(), e.getMaxIdle(), ctx.getFlags());
> interceptorChain.invoke(ctx, put);
> } catch (Exception ee) {
> log.problemApplyingStateForKey(ee.getMessage(), e.getKey());
> }
> }
>
> ...
> }
> As we can see that the code bascically try to add all data entryies got from the cluster (other node). Hence, it does not know any previous entries were deleted from the cluster which exist in its local cache store. This is exactly my test case (my confiuration is that each node has its own cache store and in replication mode).
> To fix this, we need to delete any entries from the local cache/cache store which no longer exist in the new state.
> I modified the above method by adding the following code before put loop, and it fixed the problem in my configuration:
> //Remove entries which no loger exist in the new state from local cache/cache store
> for (InternalCacheEntry ie: dataContainer.entrySet()) {
>
> if (!state.contains(ie)) {
> log.debug("Try to delete local store entry no loger exists in the new state: " + ie.getKey());
> InvocationContext ctx = icc.createInvocationContext(false, 1);
> // locking not necessary as during rehashing we block all transactions
> ctx.setFlags(CACHE_MODE_LOCAL, SKIP_CACHE_LOAD, SKIP_REMOTE_LOOKUP, SKIP_SHARED_CACHE_STORE, SKIP_LOCKING,
> SKIP_OWNERSHIP_CHECK);
> try {
> RemoveCommand remove = cf.buildRemoveCommand(ie.getKey(), ie.getValue(), ctx.getFlags());
> interceptorChain.invoke(ctx, remove);
> dataContainer.remove(ie.getKey());
> } catch (Exception ee) {
> log.error("failed to delete local store entry", ee);
> }
> }
> }
> ...
> Obvious, the above "fix" is based on assumption/configure that dataContainer will have all local entries, i.e., preload=true, no enviction replication.
> The real fix, I think, we need delegate the syncState(state) to cache store impl, where we can check the configurations and do the right thing.
> For example, in the cache store impl, we can calculate the changes based on local data and new state, and apply the changes there.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.jboss.org/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
More information about the infinispan-issues
mailing list