[IMPALA-10337] DCHECK hit at SpillableRowBatchQueue when row size exceed max reservation

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: Impala 3.4.0
• Fix Version/s: Impala 4.0.0
• Component/s: Backend
• Epic Color: ghx-label-4
• Assignee: Riza Suminto
• Reporter: Riza Suminto
• Created: 18/Nov/20 04:44
• Updated: 05/Dec/20 03:14
• Resolved: 05/Dec/20 03:14

Description

While working on IMPALA-9856, I found that the following DCHECK in SpillableRowBatchQueue::AddBatch consistently hit when result spooling is enabled and row size is larger than resource_profile_.max_reservation, causing impalad to crash.
https://github.com/apache/impala/blob/eea617b/be/src/runtime/spillable-row-batch-queue.cc#L97

We can reproduce this issue by adding the following query options in
query_test/test_insert.py::TestInsertQueries::test_insert_large_string

    self.client.set_configuration_option("spool_query_results", "1")
    self.client.set_configuration_option("max_row_size", "257mb")

Additionally, setting max_result_spooling_mem to 512MB will increase
resource_profile_.max_reservation to fit the large row and avoid this DCHECK.

Instead of DCHECK, I think impalad should return error status, suggesting that user need to set larger max_result_spooling_mem.

Another solution is to also consider max_row_size when computing maxMemReservationBytes in PlanRootSink.java.
https://github.com/apache/impala/blob/eea617b/fe/src/main/java/org/apache/impala/planner/PlanRootSink.java#L74

Issue Links

• is related to: IMPALA-9856 Enable result spooling by default