[
https://issues.apache.org/jira/browse/IGNITE-24323?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17922026#comment-17922026
]
Aleksey Plekhanov commented on IGNITE-24323:
--------------------------------------------
Implementation details:
Regular {{IgniteThreadPoolExecutor}} is used instead of striped pool executor.
But with custom {{BlockingQueue}}. This blocking queue implemented as wrapper
over the new data structure {{QueryTasksQueue}}. On {{poll}} from this queue
query is blocked by ID and no more task with the same query ID can be polled
until query is unblocked explicitely. Query is unblocked in {{finally}} block
after the task completes execution. So no tasks can be executed concurrently
for the same query and we still provide initial tasks ordering in scope of one
query.
> Calcite engine. Query execution inside UDF leads to deadlock
> ------------------------------------------------------------
>
> Key: IGNITE-24323
> URL: https://issues.apache.org/jira/browse/IGNITE-24323
> Project: Ignite
> Issue Type: Bug
> Reporter: Aleksey Plekhanov
> Assignee: Aleksey Plekhanov
> Priority: Major
> Labels: calcite, ise
> Time Spent: 0.5h
> Remaining Estimate: 0h
>
> We have striped executor to execute query tasks. Thread for query/fragment on
> each node is fixed and defined by hash of query ID and fragment ID. If user
> created user defined function and try to execute some query inside it, it can
> lead to deadlock, since task execution thread blocked by new synchronous
> query execution request and the new query can be assigned to the same thread.
> For H2-based SQL engine threads are no binded to the query and there is no
> such problem.
> For Calcite-based SQL engine we use non thread safe datastructures in tasks.
> To workaround this problem we can use regular IgniteThreadPoolExecutor, but
> need to ensure that there are no tasks concurrently executing for the same
> query/fragment.
> Also, perhaps this approach can affect performance and should be enabled by
> some property explicitly only for users that require this mode.
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
