Hello Igniters !
In the process of Ignite usage i found that some part of Compute functionality
are thread unsafe and seems was designed with such limitations initially.
Example : one (client, but it doesn`t matter at all) instance is shared between
numerous of fabric, all of them calls something like :
IgniteCompute#execute(ComputeTask<T,R>, T)
or
IgniteCompute#execute(java.lang.Class<? extends ComputeTask<T,R>>, T)
and appropriate «async» methods — what kind of instance will be called is
nondeterministic for now and as a confirmation of my words — i found no tests
covered multi thread usage of Computing i also found nothing on documentation
page [1].
We have all necessary info for correct processing of such cases:
from initiator (ignite.compute(...) starter) side we have Class or it instance
and appropriate class loader which will be wired by class loader id from
execution side.
I create a fix and seems all work perfectly well besides one place, this
functionality :
/**
* Executes given task within the cluster group. For step-by-step explanation
of task execution process
* refer to {@link ComputeTask} documentation.
* <p>
* If task for given name has not been deployed yet, then {@code taskName} will
be
* used as task class name to auto-deploy the task (see {@link
#localDeployTask(Class, ClassLoader)} method).
*/
public <T, R> R execute(String taskName, T arg) throws IgniteException;
and attendant
/**
* Finds class loader for the given class.
*
* @param rsrcName Class name or class alias to find class loader for.
* @return Deployed class loader, or {@code null} if not deployed.
*/
public DeploymentResource findResource(String rsrcName);
is thread unsafe by default, no guarantee that concurrent call of
localDeployTask and execute will bring expected result.
My proposal is to deprecate (or probably annotate [2], as a minimal —
additionally document it) this methods and to append additional :
public DeploymentResource findResource(String rsrcName, ClassLoader clsLdr);
Only one problem i can observe here, if someone creates new class loaders and
appropriate class instances in loop (i don`t know the purpose) and doesn`t
undeploy them then he will get possibly OOM here.
Such approach will give a possibility to use compute in concurrent scenario. If
there is no objections here i will mark this methods and publish my PR, of
course with additional tests.
What do you think ?
[1] https://ignite.apache.org/docs/latest/code-deployment/peer-class-loading
[2] https://jcip.net/annotations/doc/net/jcip/annotations/NotThreadSafe.html
