[IMPALA-4554] Memory corruption of nested collection with MT_DOP > 0.

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Impala 2.8.0
• Fix Version/s: Impala 2.8.0
• Component/s: Backend
• Labels: crash hang
• Target Version: Impala 2.8.0
• Assignee: Tim Armstrong
• Reporter: Alexander Behm
• Created: 29/Nov/16 21:36
• Updated: 06/Feb/17 17:50
• Resolved: 30/Nov/16 17:19

Description

A query that explodes and aggregates a nested collection appears to hang with mt_dop>0.

The reason is that the the in-memory collection slot has a "garbage" size that leads to infinite unnesting. It looks like the memory gets corrupted at some point, but I have not been able to make out where exactly yet.

Start Impala like this:

bin/start-impala-cluster.py -s 1 --impalad_args="--default_query_options=mt_dop=1"

Query to repro:

select id, cnt from functional_parquet.complextypestbl t,
(select count(item) cnt from t.int_array) v;

Relevant code snippet from unnest-node.cc

  Tuple* tuple = containing_subplan_->current_input_row_->GetTuple(coll_tuple_idx_);
  if (tuple != NULL) {
    // Retrieve the collection value to be unnested directly from the tuple. We purposely
    // ignore the null bit of the slot because we may have set it in a previous Open() of
    // this same unnest node for projection.
    coll_value_ = reinterpret_cast<const CollectionValue*>(
        tuple->GetSlot(coll_slot_desc_->tuple_offset()));
    // Projection: Set the slot containing the collection value to NULL.
    tuple->SetNull(coll_slot_desc_->null_indicator_offset());
  } else {
    coll_value_ = &EMPTY_COLLECTION_VALUE;
    DCHECK_EQ(coll_value_->num_tuples, 0);
  }

At this point coll_value_->num_tuples appears to be garbage.