[infinispan-issues] [JBoss JIRA] (ISPN-2956) putIfAbsent on Hot Rod Java client doesn't reliably fulfil contract

[Galder Zamarreño (JIRA)]

    [ https://issues.jboss.org/browse/ISPN-2956?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12970541#comment-12970541 ] 

Galder Zamarreño commented on ISPN-2956:
----------------------------------------

I've been looking into this last week trying to come up with a solution that would make these kind of operations reliable. For that, a change of protocol is required and as we near the completion of Hot Rod 2.0 version, it's probably a good time to address this. 

To support these kind of use cases, Hot Rod needs to move from the server producing the entry's version, to having the Hot Rod clients produce the version information and shipping it along with each modification. If the operation fails and the client receives the failure, it would be able to retry the operation. The retry would essentially involve sending the same operation, with a new message ID and a new RETRY flag. When the server receives a retry operation, it would check whether the entry is already present in the cache (cluster-wide), and check if the version provided by the client is the same as the entry in memory. If it's not, in the case of putIfAbsent, it would mean that entry has been overriden by someone else since the retry, and putIfAbsent would just return false. It would be the responsibility of the Hot Rod client implementations to generate the version and hence it would be transparent for the Hot Rod users.

All operations making modifications, except clear, would ship the version to be applied with the entry, and operations such as replaceIfUnmodified would ship both, the version to compare the entry with, and the version to be used if the replace succeeds. Even operations such as put() would need to ship to the server the version because if the clients provide the flag to force to return the previous value, they need to provide atomicity guarantees.

> putIfAbsent on Hot Rod Java client doesn't reliably fulfil contract
> -------------------------------------------------------------------
>
>                 Key: ISPN-2956
>                 URL: https://issues.jboss.org/browse/ISPN-2956
>             Project: Infinispan
>          Issue Type: Bug
>          Components: Remote Protocols
>            Reporter: Galder Zamarreño
>            Assignee: Galder Zamarreño
>            Priority: Minor
>              Labels: hotrod-java-client, remote-clients
>             Fix For: 7.0.0.Beta1
>
>
> Hot Rod's putIfAbsent might have issues on some edge cases:
> {quote}I want to know whether the putting entry already exists in the remote
> cache cluster, or not.
> I thought that RemoteCache.putIfAbsent() would be useful for that
> purpose, i.e.,
> {code}
> if (remoteCache.putIfAbsent(k,v) == null) {
>    // new entry.
> } else {
>    // k already exists.
> }
> {code}
> But no.
> The putIfAbsent() for new entry may return non-null value, if one of the
> server crushed while putting.
> The behavior is like the following:
> 1. Client do putIfAbsent(k,v).
> 2. The server receives the request and sends replication requests to
> other servers. If the server crushed before completing replication, some
> servers own that (k,v), but others not.
> 3. Client receives the error. The putIfAbsent() internally retries the
> same request to the next server in the cluster server list.
> 4. If the next server owns the (k,v), the putIfAbsent() returns the
> replicated (k,v) at the step 2, without any error.
> So, putIfAbsent() is not reliable for knowing whether the putting entry
> is *exactly* new or not.
> Does anyone have any idea/workaround for this purpose?{quote}
> A workaround is to do this:
> {quote}We got a simple solution, which can be applied to our customer's application.
> If each value part of putting (k,v) is unique or contains unique value,
> the client can do *double check* wether the entry is new.
> {code}
> val = System.nanoTime(); // or uuid is also useful.
> if ((ret = cache.putIfAbsent(key, val)) == null
>     || ret.equals(val)) {
>    // new entry, if the return value is just the same.
> } else {
>    // key already exists.
> }
> {code}
> We are proposing this workaround which almost works fine.{quote}
> However, this is a bit of a cludge.
> Hot Rod should be improved with an operation that allows a version to be passed when entry is created, instead of relying on the client generating it.



--
This message was sent by Atlassian JIRA
(v6.2.3#6260)