[ 
https://issues.apache.org/jira/browse/OAK-2843?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15035683#comment-15035683
 ] 

Thomas Mueller commented on OAK-2843:
-------------------------------------

I could now verify the cache works as expected. My test was: 

* Two cluster nodes, using the MongoDB document store.
* Delete the persistent cache files.
* Using the persistent cache setting is follows (OSGi configuration):
{noformat}
persistentCache="crx-quickstart/repository/cache,size\=1024,binary\=0,broadcast\=tcp:key
 123"
{noformat}
* Read all nodes of the repository (called "traversal check" in our 
application).
* This took 20 seconds (because it had to load all nodes from MongoDB).
* Do the same on the other cluster node, which only took 5 seconds.

I ran the same test without the broadcasting cache enabled, that is just with 
{noformat}
persistentCache="crx-quickstart/repository/cache,size\=1024,binary\=0"
{noformat}
The first time it took 24 seconds on _each_ cluster node (because both cluster 
nodes have to load all data from MongoDB, if the persistent cache is empty). 
The second time it took 5 seconds. After a restart (but without deleting the 
local persistent cache), it also took 5 seconds.

> Broadcasting cache
> ------------------
>
>                 Key: OAK-2843
>                 URL: https://issues.apache.org/jira/browse/OAK-2843
>             Project: Jackrabbit Oak
>          Issue Type: Improvement
>          Components: mongomk
>            Reporter: Thomas Mueller
>            Assignee: Thomas Mueller
>             Fix For: 1.3.12
>
>
> In a cluster environment, we could speed up reading if the cache(s) broadcast 
> data to other instances. This would avoid bottlenecks at the storage layer 
> (MongoDB, RDBMs).
> The configuration metadata (IP addresses and ports of where to send data to, 
> a unique identifier of the repository and the cluster nodes, possibly 
> encryption key) rarely changes and can be stored in the same place as we 
> store cluster metadata (cluster info collection). That way, in many cases no 
> manual configuration is needed. We could use TCP/IP and / or UDP.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to