[IMPALA-10233] Hit DCHECK in DmlExecState::AddPartition when inserting to a partitioned table with zorder

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: Impala 4.0.0
• Fix Version/s: Impala 4.0.0
• Component/s: Backend
• Labels: crash
• Target Version: Impala 4.0.0
• Epic Color: ghx-label-13
• Assignee: Quanlong Huang
• Reporter: Quanlong Huang
• Created: 12/Oct/20 07:13
• Updated: 20/Oct/20 07:30
• Resolved: 19/Oct/20 00:35

Description

Hit the DCHECK when inserting to a partitioned parquet table with zorder. I'm on master branch (commit=b8a2b75).

F1012 15:04:27.726274  3868 dml-exec-state.cc:432] a6479cc4725101fd:b86db2a100000003] Check failed: per_partition_status_.find(name) == per_partition_status_.end() 
*** Check failure stack trace: *** 
    @          0x51ff3cc  google::LogMessage::Fail()
    @          0x5200cbc  google::LogMessage::SendToLog()
    @          0x51fed2a  google::LogMessage::Flush()
    @          0x5202928  google::LogMessageFatal::~LogMessageFatal()
    @          0x234ba18  impala::DmlExecState::AddPartition()
    @          0x2817786  impala::HdfsTableSink::GetOutputPartition()
    @          0x2813151  impala::HdfsTableSink::WriteClusteredRowBatch()
    @          0x28156c4  impala::HdfsTableSink::Send()
    @          0x23139dd  impala::FragmentInstanceState::ExecInternal()
    @          0x230fe10  impala::FragmentInstanceState::Exec()
    @          0x227bb79  impala::QueryState::ExecFInstance()
    @          0x2279f7b  _ZZN6impala10QueryState15StartFInstancesEvENKUlvE_clEv
    @          0x227e2c2  _ZN5boost6detail8function26void_function_obj_invoker0IZN6impala10QueryState15StartFInstancesEvEUlvE_vE6invokeERNS1_15function_bufferE
    @          0x2137699  boost::function0<>::operator()()
    @          0x2715d7d  impala::Thread::SuperviseThread()
    @          0x271dd1a  boost::_bi::list5<>::operator()<>()
    @          0x271dc3e  boost::_bi::bind_t<>::operator()()
    @          0x271dbff  boost::detail::thread_data<>::run()
    @          0x3f05f01  thread_proxy
    @     0x7fb18bebb6b9  start_thread
    @     0x7fb188a474dc  clone 

It seems the zorder sort node doesn't keep the rows sorted by partition keys. Thus violates the assumption of HdfsTableSink::WriteClusteredRowBatch() that input must be ordered by the partition key expressions. So a partition key was deleted and then inserted again to the partition_keys_to_output_partitions_ map.

  /// Maps all rows in 'batch' to partitions and appends them to their temporary Hdfs
  /// files. The input must be ordered by the partition key expressions.
  Status WriteClusteredRowBatch(RuntimeState* state, RowBatch* batch) WARN_UNUSED_RESULT;

The key got removed here: https://github.com/apache/impala/blob/b8a2b754669eb7f8d164e8112e594ac413e436ef/be/src/exec/hdfs-table-sink.cc#L334 when processing a new partition key.
It got reinserted here: https://github.com/apache/impala/blob/b8a2b754669eb7f8d164e8112e594ac413e436ef/be/src/exec/hdfs-table-sink.cc#L590 so hit the DCHECK.

Issue Links

• relates to: IMPALA-8755 Implement Z-ordering for Impala
• links to: Code Review