[
https://issues.apache.org/jira/browse/TS-2184?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13911428#comment-13911428
]
Zhao Yongming commented on TS-2184:
-----------------------------------
when Cluster is designed, the origin goal is to have ONLY one single valid
content in the cluster, that is a good idea when you have very huge volume
contents, and we have continued on this target, make sure even when some of the
machines flapping in the cluster , to ensure that at anytime there is only one
valid content in the cluster.
but consider of the ICP protocol and others, those may have multiple same
content in the ICP cluster, and if you make it complex, it may have
multi-version content in the cluster at sometime. so, ICP alike protocols is
consider as a not so cool(safe) protocol if you need to enforce of the
consistency of the contents you provide to the user agents.
back to this requirement, we can make the cluster act like ICP, first write to
the cluster hashing machine, and the second or later read poll that content
from the cluster and write to the local, but it will introduce the consistency
problem here, you don't know who have that content local in the cluster when it
is updated on the origin side, within the freshness. in most case, all write to
the cache haveto broadcast to all the machines in the cluster, to enforce that
change.
the proxy.config.http.cache.cluster_cache_local enabled is a directive to
disable cluster hashing in the cluster mode, our origin target is to use it to
put some very hot hostnames(or urls) in the local, to reducing the
intro-cluster traffic. proxy.config.http.cache.cluster_cache_local enabled is
override-able and we have the same directive in the cache.config too. if it is
in active, Cluster may consider as mode=3, the single host mod.
so, if we want to achive ICP alike feature in cluster, mostly we should:
1, write the content on the hashing machine, if it is a miss in the cluster
2, read the cluster if it is missing in the local machine
3, write the local if it is a hit in the cluster
4, broadcast the change to all the machines in the cluster, if it is a
overwrite(ie, revalidating etc)
5, purge on the hashing machine and broadcast the purge to all the machines in
the cluster
it will be a very big change in the Cluster and http transaction
cc [~zwoop]
> Fetch from cluster with proxy.config.http.cache.cluster_cache_local enabled
> ---------------------------------------------------------------------------
>
> Key: TS-2184
> URL: https://issues.apache.org/jira/browse/TS-2184
> Project: Traffic Server
> Issue Type: Improvement
> Components: Cache, Clustering
> Reporter: Scott Harris
> Assignee: Bin Chen
> Fix For: 6.0.0
>
>
> With proxy.config.http.cache.cluster_cache_local enabled I would like cluster
> nodes to store content locally but try to retrieve content from the cluster
> first (if not cached locally) and if no cluster nodes have content cached
> then retrieve from origin.
> Example - 2 Cluster nodes in Full cluster mode.
> 1. Node1 and Node2 are both empty.
> 2. Request to Node1 for "http://www.example.com/foo.html";.
> 3. Query Cluster for object
> 4. Not cached in cluster so retrieve from orgin, serve to client, object now
> cached on Node1.
> 5. Request comes to Node2 for "http://www.example.com/foo.html";.
> 6. Node2 retrieves cached version from Node1, serves to client, stores
> locally.
> 7. Subsequent request comes to Node1 or Node2 for
> "http://www.example.com/foo.html";, object is served to client from local
> cache.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
