[IMPALA-4415] DCHECK failure in SimpleScheduler::CreateScanInstances()

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: Impala 2.8.0
• Fix Version/s: Impala 2.8.0
• Component/s: Backend
• Target Version: Impala 2.8.0
• Assignee: Henry Robinson
• Reporter: Henry Robinson
• Created: 01/Nov/16 23:25
• Updated: 02/Nov/16 05:29
• Resolved: 02/Nov/16 05:29

Description

Due to a floating-point rounding error, CreateScanInstances() can fail to assign the last scan range:

    // BUG: avg_bytes_per_instance * num_instances may be < total_size
    float avg_bytes_per_instance = static_cast<float>(total_size) / num_instances;
    int64_t total_assigned_bytes = 0;
    int params_idx = 0;  // into params_list
    for (int i = 0; i < num_instances; ++i) {
      fragment_params->instance_exec_params.emplace_back(
          schedule->GetNextInstanceId(), host, i, *fragment_params);
      FInstanceExecParams& instance_params = fragment_params->instance_exec_params.back();

      // Threshold beyond which we want to assign to the next instance.
      // Bug: when i+1 = num_instances, threshold_total_bytes may be < total_size
      int64_t threshold_total_bytes = avg_bytes_per_instance * (i + 1);

      // Assign each scan range in params_list. When the per-instance threshold is
      // reached, move onto the next instance.
      // Bug: loop may exit before all bytes have been assigned for last instance
      while (params_idx < params_list.size() && total_assigned_bytes < threshold_total_bytes) {
        const TScanRangeParams& scan_range_params = params_list[params_idx];
        instance_params.per_node_scan_ranges[leftmost_scan_id].push_back(
            scan_range_params);
        if (scan_range_params.scan_range.__isset.hdfs_file_split) {
          total_assigned_bytes += scan_range_params.scan_range.hdfs_file_split.length;
        } else {
          // for Kudu and Hbase every split has length 1
          ++total_assigned_bytes;
        }
        ++params_idx;
      }
      if (params_idx == params_list.size()) break; // nothing left to assign
    }
    // Bug: loop exits before params_list.size() ranges have been assigned.
    DCHECK_EQ(params_idx, params_list.size());  // everything got assigned