be/src/exec/analytic-eval-node.cc

// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

#include "exec/analytic-eval-node.h"

#include <gutil/strings/substitute.h>

#include "exec/exec-node-util.h"
#include "exprs/agg-fn-evaluator.h"
#include "exprs/agg-fn.h"
#include "exprs/scalar-expr-evaluator.h"
#include "exprs/scalar-expr.h"
#include "runtime/buffered-tuple-stream.inline.h"
#include "runtime/descriptors.h"
#include "runtime/mem-tracker.h"
#include "runtime/query-state.h"
#include "runtime/row-batch.h"
#include "runtime/runtime-state.h"
#include "udf/udf-internal.h"
#include "util/runtime-profile-counters.h"

#include "common/names.h"

static const int MAX_TUPLE_POOL_SIZE = 8 * 1024 * 1024; // 8MB
static const int MIN_REQUIRED_BUFFERS = 2;

using namespace strings;

namespace impala {

Status AnalyticEvalPlanNode::Init(const TPlanNode& tnode, RuntimeState* state) {
  RETURN_IF_ERROR(PlanNode::Init(tnode, state));

  const TAnalyticNode& analytic_node = tnode.analytic_node;
  TupleDescriptor* intermediate_tuple_desc =
      state->desc_tbl().GetTupleDescriptor(tnode.analytic_node.intermediate_tuple_id);
  TupleDescriptor* result_tuple_desc =
      state->desc_tbl().GetTupleDescriptor(tnode.analytic_node.output_tuple_id);
  bool has_lead_fn = false;

  for (int i = 0; i < analytic_node.analytic_functions.size(); ++i) {
    AggFn* analytic_fn;
    RETURN_IF_ERROR(AggFn::Create(analytic_node.analytic_functions[i],
        *children_[0]->row_descriptor_, *(intermediate_tuple_desc->slots()[i]),
        *(result_tuple_desc->slots()[i]), state, &analytic_fn));
    analytic_fns_.push_back(analytic_fn);
    DCHECK(!analytic_fn->is_merge());
    const TFunction& fn = analytic_node.analytic_functions[i].nodes[0].fn;
    const bool is_lead_fn = fn.name.function_name == "lead";
    is_lead_fn_.push_back(is_lead_fn);
    has_lead_fn |= is_lead_fn;
  }
  const TAnalyticWindow& window = tnode.analytic_node.window;
  DCHECK(!has_lead_fn || !window.__isset.window_start);
  DCHECK(window.__isset.window_end || !window.__isset.window_start)
      << "UNBOUNDED FOLLOWING is only supported with UNBOUNDED PRECEDING.";

  if (analytic_node.__isset.partition_by_eq || analytic_node.__isset.order_by_eq) {
    DCHECK(analytic_node.__isset.buffered_tuple_id);
    TupleDescriptor* buffered_tuple_desc =
        state->desc_tbl().GetTupleDescriptor(tnode.analytic_node.buffered_tuple_id);
    DCHECK(buffered_tuple_desc != nullptr);
    vector<TTupleId> tuple_ids;
    tuple_ids.push_back(children_[0]->row_descriptor_->tuple_descriptors()[0]->id());
    tuple_ids.push_back(buffered_tuple_desc->id());
    RowDescriptor cmp_row_desc(state->desc_tbl(), tuple_ids, vector<bool>(2, false));

    if (analytic_node.__isset.partition_by_eq) {
      RETURN_IF_ERROR(ScalarExpr::Create(
          analytic_node.partition_by_eq, cmp_row_desc, state, &partition_by_eq_expr_));
    }
    if (analytic_node.__isset.order_by_eq) {
      RETURN_IF_ERROR(ScalarExpr::Create(
          analytic_node.order_by_eq, cmp_row_desc, state, &order_by_eq_expr_));
    }
  }
  return Status::OK();
}

Status AnalyticEvalPlanNode::CreateExecNode(RuntimeState* state, ExecNode** node) const {
  ObjectPool* pool = state->obj_pool();
  *node = pool->Add(
      new AnalyticEvalNode(pool, *this, this->tnode_->analytic_node, state->desc_tbl()));
  return Status::OK();
}

AnalyticEvalNode::AnalyticEvalNode(ObjectPool* pool, const AnalyticEvalPlanNode& pnode,
      const TAnalyticNode& analytic_node, const DescriptorTbl& descs)
  : ExecNode(pool, pnode, descs),
    window_(analytic_node.window),
    intermediate_tuple_desc_(
        descs.GetTupleDescriptor(analytic_node.intermediate_tuple_id)),
    result_tuple_desc_(
        descs.GetTupleDescriptor(analytic_node.output_tuple_id)) {
  if (analytic_node.__isset.buffered_tuple_id) {
    buffered_tuple_desc_ =
        descs.GetTupleDescriptor(analytic_node.buffered_tuple_id);
  }
  if (!analytic_node.__isset.window) {
    fn_scope_ = AnalyticEvalNode::PARTITION;
  } else if (analytic_node.window.type == TAnalyticWindowType::RANGE) {
    fn_scope_ = AnalyticEvalNode::RANGE;
    DCHECK(!window_.__isset.window_start)
        << "RANGE windows must have UNBOUNDED PRECEDING";
    DCHECK(!window_.__isset.window_end
        || window_.window_end.type == TAnalyticWindowBoundaryType::CURRENT_ROW)
        << "RANGE window end bound must be CURRENT ROW or UNBOUNDED FOLLOWING";
  } else {
    DCHECK_EQ(analytic_node.window.type, TAnalyticWindowType::ROWS);
    fn_scope_ = AnalyticEvalNode::ROWS;
    if (window_.__isset.window_start) {
      TAnalyticWindowBoundary b = window_.window_start;
      if (b.__isset.rows_offset_value) {
        rows_start_offset_ = b.rows_offset_value;
        if (b.type == TAnalyticWindowBoundaryType::PRECEDING) rows_start_offset_ *= -1;
      } else {
        DCHECK_EQ(b.type, TAnalyticWindowBoundaryType::CURRENT_ROW);
        rows_start_offset_ = 0;
      }
    }
    if (window_.__isset.window_end) {
      TAnalyticWindowBoundary b = window_.window_end;
      if (b.__isset.rows_offset_value) {
        rows_end_offset_ = b.rows_offset_value;
        if (b.type == TAnalyticWindowBoundaryType::PRECEDING) rows_end_offset_ *= -1;
      } else {
        DCHECK_EQ(b.type, TAnalyticWindowBoundaryType::CURRENT_ROW);
        rows_end_offset_ = 0;
      }
    }
  }
  VLOG_FILE << id() << " Window=" << DebugWindowString();
  // Set the Exprs from the PlanNode
  partition_by_eq_expr_ = pnode.partition_by_eq_expr_;
  order_by_eq_expr_ = pnode.order_by_eq_expr_;
  analytic_fns_ = pnode.analytic_fns_;
  is_lead_fn_ = pnode.is_lead_fn_;
}

AnalyticEvalNode::~AnalyticEvalNode() {
  // Check that we didn't leak any memory.
  DCHECK(input_stream_ == nullptr || input_stream_->is_closed());
}

Status AnalyticEvalNode::Prepare(RuntimeState* state) {
  SCOPED_TIMER(runtime_profile_->total_time_counter());
  RETURN_IF_ERROR(ExecNode::Prepare(state));
  DCHECK(child(0)->row_desc()->IsPrefixOf(*row_desc()));
  DCHECK_GE(resource_profile_.min_reservation,
      resource_profile_.spillable_buffer_size * MIN_REQUIRED_BUFFERS);
  state_ = state;
  curr_tuple_pool_.reset(new MemPool(mem_tracker()));
  prev_tuple_pool_.reset(new MemPool(mem_tracker()));
  prev_input_tuple_pool_.reset(new MemPool(mem_tracker()));
  evaluation_timer_ = ADD_TIMER(runtime_profile(), "EvaluationTime");

  DCHECK_EQ(result_tuple_desc_->slots().size(), analytic_fns_.size());
  RETURN_IF_ERROR(AggFnEvaluator::Create(analytic_fns_, state, pool_, expr_perm_pool(),
      expr_results_pool(), &analytic_fn_evals_));

  if (partition_by_eq_expr_ != nullptr) {
    RETURN_IF_ERROR(ScalarExprEvaluator::Create(*partition_by_eq_expr_, state, pool_,
        expr_perm_pool(), expr_results_pool(), &partition_by_eq_expr_eval_));
  }

  if (order_by_eq_expr_ != nullptr) {
    RETURN_IF_ERROR(ScalarExprEvaluator::Create(*order_by_eq_expr_, state, pool_,
        expr_perm_pool(), expr_results_pool(), &order_by_eq_expr_eval_));
  }

  // An intermediate tuple that is only allocated once and is reused. 'curr_tuple_' is
  // initialized in Open() before it is used.
  curr_tuple_ =
      Tuple::Create(intermediate_tuple_desc_->byte_size(), expr_perm_pool_.get());
  dummy_result_tuple_ =
      Tuple::Create(result_tuple_desc_->byte_size(), expr_perm_pool_.get());
  return Status::OK();
}

Status AnalyticEvalNode::Open(RuntimeState* state) {
  SCOPED_TIMER(runtime_profile_->total_time_counter());
  ScopedOpenEventAdder ea(this);
  RETURN_IF_ERROR(ExecNode::Open(state));
  RETURN_IF_CANCELLED(state);
  RETURN_IF_ERROR(QueryMaintenance(state));
  RETURN_IF_ERROR(child(0)->Open(state));

  // Claim reservation after the child has been opened to reduce the peak reservation
  // requirement.
  if (!buffer_pool_client()->is_registered()) {
    RETURN_IF_ERROR(ClaimBufferReservation(state));
  }
  DCHECK(input_stream_ == nullptr);
  input_stream_.reset(new BufferedTupleStream(state, child(0)->row_desc(),
      buffer_pool_client(), resource_profile_.spillable_buffer_size,
      resource_profile_.spillable_buffer_size));
  RETURN_IF_ERROR(input_stream_->Init(label(), true));
  bool success;
  RETURN_IF_ERROR(input_stream_->PrepareForReadWrite(true, &success));
  DCHECK(success) << "Had reservation: " << buffer_pool_client()->DebugString();

  for (int i = 0; i < analytic_fn_evals_.size(); ++i) {
    RETURN_IF_ERROR(analytic_fn_evals_[i]->Open(state));
    FunctionContext* agg_fn_ctx = analytic_fn_evals_[i]->agg_fn_ctx();
    if (!has_first_val_null_offset_ &&
        "first_value_rewrite" == analytic_fns_[i]->fn_name() &&
        agg_fn_ctx->GetNumArgs() == 2) {
      DCHECK(!has_first_val_null_offset_);
      first_val_null_offset_ =
          reinterpret_cast<BigIntVal*>(agg_fn_ctx->GetConstantArg(1))->val;
      VLOG_FILE << id() << " FIRST_VAL rewrite null offset: " << first_val_null_offset_;
      has_first_val_null_offset_ = true;
    }
  }

  if (partition_by_eq_expr_eval_ != nullptr) {
    RETURN_IF_ERROR(partition_by_eq_expr_eval_->Open(state));
  }

  if (order_by_eq_expr_eval_ != nullptr) {
    RETURN_IF_ERROR(order_by_eq_expr_eval_->Open(state));
  }

  // Initialize the tuple that was allocated in Prepare().
  // TODO: zeroing out curr_tuple_ shouldn't be strictly necessary.
  curr_tuple_->Init(intermediate_tuple_desc_->byte_size());
  AggFnEvaluator::Init(analytic_fn_evals_, curr_tuple_);
  curr_tuple_init_ = true;
  // Check for failures during AggFnEvaluator::Init().
  RETURN_IF_ERROR(state->GetQueryStatus());

  // Initialize state for the first partition.
  RETURN_IF_ERROR(InitNextPartition(state, 0));
  if (curr_child_batch_ == nullptr) {
    curr_child_batch_.reset(
        new RowBatch(child(0)->row_desc(), state->batch_size(), mem_tracker()));
  }
  return Status::OK();
}

string DebugWindowBoundString(const TAnalyticWindowBoundary& b) {
  if (b.type == TAnalyticWindowBoundaryType::CURRENT_ROW) {
    return "CURRENT_ROW";
  }
  stringstream ss;
  if (b.__isset.rows_offset_value) {
    ss << b.rows_offset_value;
  } else {
    // TODO: Return debug string when range offsets are supported
    DCHECK(false) << "Range offsets not yet implemented";
  }
  if (b.type == TAnalyticWindowBoundaryType::PRECEDING) {
    ss << " PRECEDING";
  } else {
    DCHECK_EQ(b.type, TAnalyticWindowBoundaryType::FOLLOWING);
    ss << " FOLLOWING";
  }
  return ss.str();
}

string AnalyticEvalNode::DebugWindowString() const {
  stringstream ss;
  if (fn_scope_ == PARTITION) {
    ss << "NO WINDOW";
    return ss.str();
  }
  ss << "{type=";
  if (fn_scope_ == RANGE) {
    ss << "RANGE";
  } else {
    ss << "ROWS";
  }
  ss << ", start=";
  if (window_.__isset.window_start) {
    ss << DebugWindowBoundString(window_.window_start);
  } else {
    ss << "UNBOUNDED_PRECEDING";
  }

  ss << ", end=";
  if (window_.__isset.window_end) {
    ss << DebugWindowBoundString(window_.window_end) << "}";
  } else {
    ss << "UNBOUNDED_FOLLOWING";
  }
  return ss.str();
}

string AnalyticEvalNode::DebugStateString(bool detailed = false) const {
  stringstream ss;
  ss << "num_returned=" << input_stream_->rows_returned()
     << " num_rows=" << input_stream_->num_rows()
     << " curr_partition_idx_=" << curr_partition_idx_
     << " last_result_idx=" << last_result_idx_;
  if (detailed) {
    ss << " result_tuples idx: [";
    for (const pair<int64_t, Tuple*>& result_tuple : result_tuples_) {
      ss << result_tuple.first;
      if (&result_tuple != &result_tuples_.back()) ss << ", ";
    }
    ss << "]";
    if (fn_scope_ == ROWS && window_.__isset.window_start) {
      ss << " window_tuples idx: [";
    for (const pair<int64_t, Tuple*>& window_tuple : window_tuples_) {
        ss << window_tuple.first;
        if (&window_tuple != &window_tuples_.back()) ss << ", ";
      }
      ss << "]";
    }
  } else {
    if (fn_scope_ == ROWS && window_.__isset.window_start) {
      if (window_tuples_.empty()) {
        ss << " window_tuples empty";
      } else {
        ss << " window_tuples idx range: (" << window_tuples_.front().first << ","
          << window_tuples_.back().first << ")";
      }
    }
    if (result_tuples_.empty()) {
      ss << " result_tuples empty";
    } else {
      ss << " result_tuples idx range: (" << result_tuples_.front().first << ","
        << result_tuples_.back().first << ")";
    }
  }
  return ss.str();
}

inline Status AnalyticEvalNode::AddRow(int64_t stream_idx, TupleRow* row) {
  if (fn_scope_ != ROWS || !window_.__isset.window_start ||
      stream_idx - rows_start_offset_ >= curr_partition_idx_) {
    VLOG_ROW << id() << " Update idx=" << stream_idx;
    AggFnEvaluator::Add(analytic_fn_evals_, row, curr_tuple_);
    if (window_.__isset.window_start) {
      VLOG_ROW << id() << " Adding tuple to window at idx=" << stream_idx;
      Tuple* tuple = row->GetTuple(0)->DeepCopy(
          *child(0)->row_desc()->tuple_descriptors()[0], curr_tuple_pool_.get());
      window_tuples_.emplace_back(stream_idx, tuple);
    }
  }

  Status status;
  // Buffer the entire input row to be returned later with the analytic eval results.
  if (UNLIKELY(!input_stream_->AddRow(row, &status))) {
    // AddRow returns false if an error occurs (available via status()) or there is
    // not enough memory (status() is OK). If there isn't enough memory, we unpin
    // the stream and continue writing/reading in unpinned mode.
    // TODO: Consider re-pinning later if the output stream is fully consumed.
    RETURN_IF_ERROR(status);
    RETURN_IF_ERROR(state_->StartSpilling(mem_tracker()));
    RETURN_IF_ERROR(
        input_stream_->UnpinStream(BufferedTupleStream::UNPIN_ALL_EXCEPT_CURRENT));
    VLOG_FILE << id() << " Unpin input stream while adding row idx=" << stream_idx;
    if (!input_stream_->AddRow(row, &status)) {
      // Rows should be added in unpinned mode unless an error occurs.
      RETURN_IF_ERROR(status);
      DCHECK(false);
    }
  }
  DCHECK(status.ok());
  return Status::OK();
}

Status AnalyticEvalNode::AddResultTuple(int64_t stream_idx) {
  VLOG_ROW << id() << " AddResultTuple idx=" << stream_idx;
  DCHECK(curr_tuple_ != nullptr);
  MemPool* curr_tuple_pool = curr_tuple_pool_.get();
  Tuple* result_tuple = Tuple::Create(result_tuple_desc_->byte_size(), curr_tuple_pool);

  AggFnEvaluator::GetValue(analytic_fn_evals_, curr_tuple_, result_tuple);
  // Copy any string data in 'result_tuple' into 'curr_tuple_pool'. The var-len data
  // returned by GetValue() may be backed by an allocation from
  // 'expr_results_pool_' that will be recycled so it must be copied out.
  for (const SlotDescriptor* slot_desc : result_tuple_desc_->string_slots()) {
    if (result_tuple->IsNull(slot_desc->null_indicator_offset())) continue;
    StringValue* sv = result_tuple->GetStringSlot(slot_desc->tuple_offset());
    if (sv->len == 0) continue;
    char* new_ptr = reinterpret_cast<char*>(
        curr_tuple_pool->TryAllocateUnaligned(sv->len));
    if (UNLIKELY(new_ptr == nullptr)) {
      return curr_tuple_pool->mem_tracker()->MemLimitExceeded(nullptr,
          "Failed to allocate memory for analytic function's result.", sv->len);
    }
    memcpy(new_ptr, sv->ptr, sv->len);
    sv->ptr = new_ptr;
  }

  DCHECK_GT(stream_idx, last_result_idx_);
  result_tuples_.emplace_back(stream_idx, result_tuple);
  last_result_idx_ = stream_idx;
  VLOG_ROW << id() << " Added result tuple, final state: " << DebugStateString(true);
  return Status::OK();
}

inline Status AnalyticEvalNode::TryAddResultTupleForPrevRow(
    const TupleRow* child_tuple_cmp_row, bool next_partition, int64_t stream_idx) {
  // The analytic fns are finalized after the previous row if we found a new partition
  // or the window is a RANGE and the order by exprs changed. For ROWS windows we do not
  // need to compare the current row to the previous row.
  VLOG_ROW << id() << " TryAddResultTupleForPrevRow partition=" << next_partition
           << " idx=" << stream_idx;
  if (fn_scope_ != ROWS && (next_partition
        || (fn_scope_ == RANGE && window_.__isset.window_end
            && !(order_by_eq_expr_eval_ == nullptr ||
                 PrevRowCompare(order_by_eq_expr_eval_, child_tuple_cmp_row))))) {
    RETURN_IF_ERROR(AddResultTuple(stream_idx - 1));
  }
  return Status::OK();
}

inline Status AnalyticEvalNode::TryAddResultTupleForCurrRow(int64_t stream_idx) {
  VLOG_ROW << id() << " TryAddResultTupleForCurrRow idx=" << stream_idx;
  // We only add results at this point for ROWS windows (unless unbounded following)
  // Nothing to add if the end offset is before the start of the partition.
  if (fn_scope_ == ROWS && window_.__isset.window_end &&
      stream_idx - rows_end_offset_ >= curr_partition_idx_) {
    RETURN_IF_ERROR(AddResultTuple(stream_idx - rows_end_offset_));
  }
  return Status::OK();
}

inline void AnalyticEvalNode::TryRemoveRowsBeforeWindow(int64_t stream_idx) {
  if (fn_scope_ != ROWS || !window_.__isset.window_start) return;
  // The start of the window may have been before the current partition, in which case
  // there is no tuple to remove in window_tuples_. Check the index of the row at which
  // tuples from window_tuples_ should begin to be removed.
  int64_t remove_idx = stream_idx - rows_end_offset_ +
      min<int64_t>(rows_start_offset_, 0) - 1;
  if (remove_idx < curr_partition_idx_) return;
  VLOG_ROW << id() << " Remove idx=" << remove_idx << " stream_idx=" << stream_idx;
  DCHECK(!window_tuples_.empty()) << DebugStateString(true);
  DCHECK_EQ(remove_idx + max<int64_t>(rows_start_offset_, 0),
      window_tuples_.front().first) << DebugStateString(true);
  TupleRow* remove_row = reinterpret_cast<TupleRow*>(&window_tuples_.front().second);
  AggFnEvaluator::Remove(analytic_fn_evals_, remove_row, curr_tuple_);
  window_tuples_.pop_front();
}

inline Status AnalyticEvalNode::TryAddRemainingResults(int64_t partition_idx,
    int64_t prev_partition_idx) {
  DCHECK_LT(prev_partition_idx, partition_idx);
  // For PARTITION, RANGE, or ROWS with UNBOUNDED PRECEDING: add a result tuple for the
  // remaining rows in the partition that do not have an associated result tuple yet.
  if (fn_scope_ != ROWS || !window_.__isset.window_end) {
    if (last_result_idx_ < partition_idx - 1) {
      RETURN_IF_ERROR(AddResultTuple(partition_idx - 1));
    }
    return Status::OK();
  }

  // lead() is re-written to a ROWS window with an end bound FOLLOWING. Any remaining
  // results need the default value (set by Init()). If this is the case, the start bound
  // is UNBOUNDED PRECEDING (DCHECK in Init()).
  for (int i = 0; i < analytic_fn_evals_.size(); ++i) {
    if (is_lead_fn_[i]) {
      // Needs to call Finalize() to release resources.
      analytic_fn_evals_[i]->Finalize(curr_tuple_, dummy_result_tuple_);
      analytic_fn_evals_[i]->Init(curr_tuple_);
    }
  }

  // If the start bound is not UNBOUNDED PRECEDING and there are still rows in the
  // partition for which we need to produce result tuples, we need to continue removing
  // input tuples at the start of the window from each row that we're adding results for.
  VLOG_ROW << id() << " TryAddRemainingResults prev_partition_idx=" << prev_partition_idx
           << " " << DebugStateString(true);
  for (int64_t next_result_idx = last_result_idx_ + 1; next_result_idx < partition_idx;
      ++next_result_idx) {
    if (window_tuples_.empty()) break;
    if (next_result_idx + rows_start_offset_ > window_tuples_.front().first) {
      DCHECK_EQ(next_result_idx + rows_start_offset_ - 1, window_tuples_.front().first);
      // For every tuple that is removed from the window: Remove() from the evaluators
      // and add the result tuple at the next index.
      VLOG_ROW << id() << " Remove window_row_idx=" << window_tuples_.front().first
               << " for result row at idx=" << next_result_idx;
      TupleRow* remove_row = reinterpret_cast<TupleRow*>(&window_tuples_.front().second);
      AggFnEvaluator::Remove(analytic_fn_evals_, remove_row, curr_tuple_);
      window_tuples_.pop_front();
    }
    RETURN_IF_ERROR(AddResultTuple(last_result_idx_ + 1));
  }

  // If there are still rows between the row with the last result (AddResultTuple() may
  // have updated last_result_idx_) and the partition boundary, add the current results
  // for the remaining rows with the same result tuple (curr_tuple_ is not modified).
  if (last_result_idx_ < partition_idx - 1) {
    RETURN_IF_ERROR(AddResultTuple(partition_idx - 1));
  }
  return Status::OK();
}

inline Status AnalyticEvalNode::InitNextPartition(RuntimeState* state,
    int64_t stream_idx) {
  VLOG_FILE << id() << " InitNextPartition idx=" << stream_idx;
  DCHECK_LT(curr_partition_idx_, stream_idx);
  int64_t prev_partition_stream_idx = curr_partition_idx_;
  curr_partition_idx_ = stream_idx;

  // If the window has an end bound preceding the current row, we will have output tuples
  // for rows beyond the previous partition, so they should be removed.  Because
  // result_tuples_ is a sparse structure, the last result tuple of the previous
  // partition may have been added to result_tuples_ with a stream index equal to or
  // beyond curr_partition_idx_. So the last entry in result_tuples_ with a stream index
  // >= curr_partition_idx_ is the last result tuple of the previous partition.  Adding
  // the last result tuple to result_tuples_ with a stream index curr_partition_idx_ - 1
  // ensures that all rows in the previous partition have corresponding analytic results.
  Tuple* prev_partition_last_result_tuple = nullptr;
  while (!result_tuples_.empty() && result_tuples_.back().first >= curr_partition_idx_) {
    DCHECK(fn_scope_ == ROWS && window_.__isset.window_end &&
        window_.window_end.type == TAnalyticWindowBoundaryType::PRECEDING);
    VLOG_ROW << id() << " Removing result past partition idx: "
             << result_tuples_.back().first;
    prev_partition_last_result_tuple = result_tuples_.back().second;
    result_tuples_.pop_back();
  }
  if (prev_partition_last_result_tuple != nullptr) {
    if (result_tuples_.empty() ||
        result_tuples_.back().first < curr_partition_idx_ - 1) {
      // prev_partition_last_result_tuple was the last result tuple in the partition, add
      // it back with the index of the last row in the partition so that all output rows
      // in this partition get the correct value.
      result_tuples_.emplace_back(
          curr_partition_idx_ - 1, prev_partition_last_result_tuple);
    }
    DCHECK(!result_tuples_.empty());
    last_result_idx_ = result_tuples_.back().first;
    VLOG_ROW << id() << " After removing results past partition: "
             << DebugStateString(true);
    DCHECK_EQ(last_result_idx_, curr_partition_idx_ - 1);
    DCHECK_LE(input_stream_->rows_returned(), last_result_idx_);
  }
  DCHECK(result_tuples_.empty() || (last_result_idx_ == result_tuples_.back().first));

  if (fn_scope_ == ROWS && stream_idx > 0 && (!window_.__isset.window_end ||
        window_.window_end.type == TAnalyticWindowBoundaryType::FOLLOWING)) {
    RETURN_IF_ERROR(TryAddRemainingResults(stream_idx, prev_partition_stream_idx));
  }
  window_tuples_.clear();

  VLOG_ROW << id() << " Reset curr_tuple";
  // Call finalize to release resources; result is not needed but the dst tuple must be
  // a tuple described by result_tuple_desc_.
  DCHECK(curr_tuple_init_);
  AggFnEvaluator::Finalize(analytic_fn_evals_, curr_tuple_, dummy_result_tuple_);
  // Re-initialize curr_tuple_.
  // TODO: zeroing out curr_tuple_ shouldn't be strictly necessary.
  curr_tuple_->Init(intermediate_tuple_desc_->byte_size());
  AggFnEvaluator::Init(analytic_fn_evals_, curr_tuple_);
  // Check for errors in AggFnEvaluator::Init().
  RETURN_IF_ERROR(state->GetQueryStatus());

  // Add a result tuple containing values set by Init() (e.g. nullptr for sum(), 0 for
  // count()) for output rows that have no input rows in the window. We need to add this
  // result tuple before any input rows are consumed and the evaluators are updated.
  if (fn_scope_ == ROWS && window_.__isset.window_end &&
      window_.window_end.type == TAnalyticWindowBoundaryType::PRECEDING) {
    if (has_first_val_null_offset_) {
      // Special handling for FIRST_VALUE which has the window rewritten in the FE
      // in order to evaluate the fn efficiently with a trivial agg fn implementation.
      // This occurs when the original analytic window has a start bound X PRECEDING. In
      // that case, the window is rewritten to have an end bound X PRECEDING which would
      // normally mean we add the newly Init()'d result tuple X rows down (so that those
      // first rows have the initial value because they have no rows in their windows).
      // However, the original query did not actually have X PRECEDING so we need to do
      // one of the following:
      // 1) Do not insert the initial result tuple with at all, indicated by
      //    first_val_null_offset_ == -1. This happens when the original end bound was
      //    actually CURRENT ROW or Y FOLLOWING.
      // 2) Insert the initial result tuple at first_val_null_offset_. This happens when
      //    the end bound was actually Y PRECEDING.
      if (first_val_null_offset_ != -1) {
        RETURN_IF_ERROR(AddResultTuple(curr_partition_idx_ + first_val_null_offset_ - 1));
      }
    } else {
      RETURN_IF_ERROR(AddResultTuple(curr_partition_idx_ - rows_end_offset_ - 1));
    }
  }
  return Status::OK();
}

inline bool AnalyticEvalNode::PrevRowCompare(
   ScalarExprEvaluator* pred_eval, const TupleRow* child_tuple_cmp_row) {
  DCHECK(pred_eval != nullptr);
  BooleanVal result = pred_eval->GetBooleanVal(child_tuple_cmp_row);
  DCHECK(!result.is_null);
  return result.val;
}

Status AnalyticEvalNode::ProcessChildBatches(RuntimeState* state) {
  // Consume child batches until eos or there are enough rows to return more than an
  // output batch. Ensuring there is at least one more row left after returning results
  // allows us to simplify the logic dealing with last_result_idx_ and result_tuples_.
  while (!input_eos_ && NumOutputRowsReady() < state->batch_size() + 1) {
    RETURN_IF_CANCELLED(state);
    if (has_partition_or_order_by_expr_eval() && prev_input_tuple_ != nullptr
        && curr_child_batch_->num_rows() > 0) {
      // 'prev_input_tuple_' is from the last row in  'curr_child_batch_' and is needed
      // by subsequent calls to ProcessChildBatch(). Deep copy it so that we can safely
      // free the memory backing it. 'prev_input_tuple_pool_' is not backing the previous
      // input tuple any more - it is from the last row in 'curr_child_batch_' and
      // therefore backed by 'curr_child_batch_'.
      prev_input_tuple_pool_->Clear();
      prev_input_tuple_ = prev_input_tuple_->DeepCopy(
          *child(0)->row_desc()->tuple_descriptors()[0], prev_input_tuple_pool_.get());
    }
    curr_child_batch_->Reset();
    RETURN_IF_ERROR(child(0)->GetNext(state, curr_child_batch_.get(), &input_eos_));
    RETURN_IF_ERROR(QueryMaintenance(state));
    RETURN_IF_ERROR(ProcessChildBatch(state));
    // TODO: DCHECK that the size of result_tuples_ is bounded. It shouldn't be larger
    // than 2x the batch size unless the end bound has an offset preceding, in which
    // case it may be slightly larger (proportional to the offset but still bounded).
  }
  return Status::OK();
}

Status AnalyticEvalNode::ProcessChildBatch(RuntimeState* state) {
  // TODO: DCHECK input is sorted (even just first row vs prev_input_tuple_)
  VLOG_FILE << id() << " ProcessChildBatch: " << DebugStateString()
            << " input batch size:" << curr_child_batch_->num_rows()
            << " tuple pool size:" << curr_tuple_pool_->total_allocated_bytes();
  SCOPED_TIMER(evaluation_timer_);

  // BufferedTupleStream::num_rows() returns the total number of rows that have been
  // inserted into the stream (it does not decrease when we read rows), so the index of
  // the next input row that will be inserted will be the current size of the stream.
  int64_t stream_idx = input_stream_->num_rows();

  // The very first row in the stream is handled specially because there is no previous
  // row to compare and we cannot rely on PrevRowCompare() returning true even for the
  // same row pointers if there are NaN values.
  int batch_idx = 0;
  if (UNLIKELY(stream_idx == 0 && curr_child_batch_->num_rows() > 0)) {
    TupleRow* row = curr_child_batch_->GetRow(0);
    RETURN_IF_ERROR(AddRow(0, row));
    RETURN_IF_ERROR(TryAddResultTupleForCurrRow(0));
    prev_input_tuple_ = row->GetTuple(0);
    ++batch_idx;
    ++stream_idx;
  }
  Tuple* child_tuple_cmp_row_tuples[2] = {nullptr, nullptr};
  TupleRow* child_tuple_cmp_row =
      reinterpret_cast<TupleRow*>(child_tuple_cmp_row_tuples);
  for (; batch_idx < curr_child_batch_->num_rows(); ++batch_idx, ++stream_idx) {
    TupleRow* row = curr_child_batch_->GetRow(batch_idx);
    if (has_partition_or_order_by_expr_eval()) {
      // Only set the tuples in child_tuple_cmp_row if there are partition exprs or
      // order by exprs that require comparing the current and previous rows. If there
      // aren't partition or order by exprs (i.e. empty OVER() clause), there was no
      // sort and there could be nullable tuples (whereas the sort node does not produce
      // them), see IMPALA-1562.
      child_tuple_cmp_row->SetTuple(0, prev_input_tuple_);
      child_tuple_cmp_row->SetTuple(1, row->GetTuple(0));
    }
    TryRemoveRowsBeforeWindow(stream_idx);

    // Every row is compared against the previous row to determine if (a) the row
    // starts a new partition or (b) the row does not share the same values for the
    // ordering exprs. When either of these occurs, the analytic_fn_evals_ are
    // finalized and the result tuple is added to result_tuples_ so that it may be
    // added to output rows in GetNextOutputBatch(). When a new partition is found
    // (a), a new, empty result tuple is created and initialized over the
    // analytic_fn_evals_. If the row has different values for the ordering
    // exprs (b), then a new tuple is created but copied from curr_tuple_ because the
    // original is used for one or more previous row(s) but the incremental state still
    // applies to the current row.
    bool next_partition = false;
    if (partition_by_eq_expr_eval_ != nullptr) {
      // partition_by_eq_expr_eval_ checks equality over the predicate exprs
      next_partition = !PrevRowCompare(partition_by_eq_expr_eval_, child_tuple_cmp_row);
    }
    RETURN_IF_ERROR(TryAddResultTupleForPrevRow(
          child_tuple_cmp_row, next_partition, stream_idx));
    if (next_partition) RETURN_IF_ERROR(InitNextPartition(state, stream_idx));

    // The analytic_fn_evals_ are updated with the current row.
    RETURN_IF_ERROR(AddRow(stream_idx, row));

    RETURN_IF_ERROR(TryAddResultTupleForCurrRow(stream_idx));
    prev_input_tuple_ = row->GetTuple(0);
  }

  if (UNLIKELY(input_eos_ && stream_idx > curr_partition_idx_)) {
    // We need to add the results for the last row(s).
    RETURN_IF_ERROR(TryAddRemainingResults(stream_idx, curr_partition_idx_));
  }

  // Transfer resources to prev_tuple_pool_ when enough resources have accumulated
  // and the prev_tuple_pool_ has already been transfered to an output batch.
  if (curr_tuple_pool_->total_allocated_bytes() > MAX_TUPLE_POOL_SIZE &&
      (prev_pool_last_result_idx_ == -1 || prev_pool_last_window_idx_ == -1)) {
    prev_tuple_pool_->AcquireData(curr_tuple_pool_.get(), false);
    prev_pool_last_result_idx_ = last_result_idx_;
    if (window_tuples_.size() > 0) {
      prev_pool_last_window_idx_ = window_tuples_.back().first;
    } else {
      prev_pool_last_window_idx_ = -1;
    }
    VLOG_FILE << id() << " Transfer resources from curr to prev pool at idx: "
              << stream_idx << ", stores tuples with last result idx: "
              << prev_pool_last_result_idx_ << " last window idx: "
              << prev_pool_last_window_idx_;
  }
  return Status::OK();
}

Status AnalyticEvalNode::GetNextOutputBatch(
    RuntimeState* state, RowBatch* output_batch, bool* eos) {
  SCOPED_TIMER(evaluation_timer_);
  VLOG_FILE << id() << " GetNextOutputBatch: " << DebugStateString()
            << " tuple pool size:" << curr_tuple_pool_->total_allocated_bytes();
  if (input_stream_->rows_returned() == input_stream_->num_rows()) {
    *eos = true;
    return Status::OK();
  }

  const int num_child_tuples = child(0)->row_desc()->tuple_descriptors().size();
  RowBatch input_batch(child(0)->row_desc(), output_batch->capacity(), mem_tracker());
  int64_t stream_idx = input_stream_->rows_returned();
  RETURN_IF_ERROR(input_stream_->GetNext(&input_batch, eos));
  for (int i = 0; i < input_batch.num_rows(); ++i) {
    if (ReachedLimit()) break;
    DCHECK(!output_batch->AtCapacity());
    DCHECK(!result_tuples_.empty());
    VLOG_ROW << id() << " Output row idx=" << stream_idx << " " << DebugStateString(true);

    // CopyRow works as expected: input_batch tuples form a prefix of output_batch
    // tuples.
    TupleRow* dest = output_batch->GetRow(output_batch->AddRow());
    input_batch.CopyRow(input_batch.GetRow(i), dest);
    dest->SetTuple(num_child_tuples, result_tuples_.front().second);
    output_batch->CommitLastRow();
    IncrementNumRowsReturned(1);

    // Remove the head of result_tuples_ if all rows using that evaluated tuple
    // have been returned.
    DCHECK_LE(stream_idx, result_tuples_.front().first);
    if (stream_idx >= result_tuples_.front().first) result_tuples_.pop_front();
    ++stream_idx;
  }
  input_batch.TransferResourceOwnership(output_batch);
  if (ReachedLimit()) *eos = true;
  return Status::OK();
}

inline int64_t AnalyticEvalNode::NumOutputRowsReady() const {
  if (result_tuples_.empty()) return 0;
  int64_t rows_to_return = last_result_idx_ - input_stream_->rows_returned();
  if (last_result_idx_ > input_stream_->num_rows()) {
    // This happens when we were able to add a result tuple before consuming child rows,
    // e.g. initializing a new partition with an end bound that is X preceding. The first
    // X rows get the default value and we add that tuple to result_tuples_ before
    // consuming child rows. It's possible the result is negative, and that's fine
    // because this result is only used to determine if the number of rows to return
    // is at least as big as the batch size.
    rows_to_return -= last_result_idx_ - input_stream_->num_rows();
  } else {
    DCHECK_GE(rows_to_return, 0);
  }
  return rows_to_return;
}

Status AnalyticEvalNode::GetNext(RuntimeState* state, RowBatch* row_batch, bool* eos) {
  SCOPED_TIMER(runtime_profile_->total_time_counter());
  ScopedGetNextEventAdder ea(this, eos);
  RETURN_IF_ERROR(ExecDebugAction(TExecNodePhase::GETNEXT, state));
  RETURN_IF_CANCELLED(state);
  RETURN_IF_ERROR(QueryMaintenance(state));
  VLOG_FILE << id() << " GetNext: " << DebugStateString();
  DCHECK(!input_stream_->is_closed()); // input_stream_ is closed if we already hit eos.

  if (ReachedLimit()) {
    // TODO: This transfer is simple and correct, but not necessarily efficient. We
    // should optimize the use/transfer of memory to better amortize allocations
    // over multiple Reset()/Open()/GetNext()* cycles.
    row_batch->tuple_data_pool()->AcquireData(prev_tuple_pool_.get(), false);
    row_batch->tuple_data_pool()->AcquireData(curr_tuple_pool_.get(), false);
    DCHECK(!input_stream_->is_closed());
    // Flush resources in case we are in a subplan and need to allocate more blocks
    // when the node is reopened.
    input_stream_->Close(row_batch, RowBatch::FlushMode::FLUSH_RESOURCES);
    *eos = true;
    return Status::OK();
  } else {
    *eos = false;
  }

  RETURN_IF_ERROR(ProcessChildBatches(state));

  bool output_eos = false;
  RETURN_IF_ERROR(GetNextOutputBatch(state, row_batch, &output_eos));
  if (input_eos_ && output_eos) {
    // Transfer the ownership of all row-backing resources on eos for simplicity.
    // TODO: This transfer is simple and correct, but not necessarily efficient. We
    // should optimize the use/transfer of memory to better amortize allocations
    // over multiple Reset()/Open()/GetNext()* cycles.
    row_batch->tuple_data_pool()->AcquireData(prev_tuple_pool_.get(), false);
    row_batch->tuple_data_pool()->AcquireData(curr_tuple_pool_.get(), false);
    // Flush resources in case we are in a subplan and need to allocate more blocks
    // when the node is reopened.
    input_stream_->Close(row_batch, RowBatch::FlushMode::FLUSH_RESOURCES);
    *eos = true;
  } else if (prev_pool_last_result_idx_ != -1 &&
      prev_pool_last_result_idx_ < input_stream_->rows_returned() &&
      prev_pool_last_window_idx_ < window_tuples_.front().first) {
    // Transfer resources to the output row batch if enough have accumulated and they're
    // no longer needed by output rows to be returned later.
    VLOG_FILE << id() << " Transfer prev pool to output batch, "
              << " pool size: " << prev_tuple_pool_->total_allocated_bytes()
              << " last result idx: " << prev_pool_last_result_idx_
              << " last window idx: " << prev_pool_last_window_idx_;
    row_batch->tuple_data_pool()->AcquireData(prev_tuple_pool_.get(), !*eos);
    prev_pool_last_result_idx_ = -1;
    prev_pool_last_window_idx_ = -1;
  }

  COUNTER_SET(rows_returned_counter_, rows_returned());
  return Status::OK();
}

Status AnalyticEvalNode::Reset(RuntimeState* state, RowBatch* row_batch) {
  result_tuples_.clear();
  window_tuples_.clear();
  last_result_idx_ = -1;
  curr_partition_idx_ = -1;
  prev_pool_last_result_idx_ = -1;
  prev_pool_last_window_idx_ = -1;
  input_eos_ = false;
  // Transfer the ownership of all row-backing resources.
  row_batch->tuple_data_pool()->AcquireData(prev_tuple_pool_.get(), false);
  row_batch->tuple_data_pool()->AcquireData(curr_tuple_pool_.get(), false);
  // Call Finalize() to clear evaluator allocations, but do not Close() them,
  // so we can keep evaluating them.
  if (curr_tuple_init_) {
    AggFnEvaluator::Finalize(analytic_fn_evals_, curr_tuple_, dummy_result_tuple_);
    curr_tuple_init_ = false;
  }
  // The following members will be re-created in Open().
  // input_stream_ should have been closed by last GetNext() call.
  if (input_stream_ != nullptr && !input_stream_->is_closed()) {
    input_stream_->Close(row_batch, RowBatch::FlushMode::FLUSH_RESOURCES);
  }
  input_stream_.reset();
  prev_input_tuple_ = nullptr;
  prev_input_tuple_pool_->Clear();
  curr_child_batch_->Reset();
  return ExecNode::Reset(state, row_batch);
}

void AnalyticEvalNode::Close(RuntimeState* state) {
  if (is_closed()) return;
  // We may need to clean up input_stream_ if an error occurred at some point.
  if (input_stream_ != nullptr) {
    input_stream_->Close(nullptr, RowBatch::FlushMode::NO_FLUSH_RESOURCES);
  }

  // Close all evaluators. If an error occurred in Init or Prepare there may be fewer
  // be fewer evaluators than analytic fns. We also need to Finalize if curr_tuple_ was
  // created in Open().
  DCHECK_LE(analytic_fn_evals_.size(), analytic_fns_.size());
  DCHECK(curr_tuple_ == nullptr ||
      analytic_fn_evals_.size() == analytic_fns_.size());
  // Need to make sure finalize is called in case there is any state to clean up.
  if (curr_tuple_init_)  {
    AggFnEvaluator::Finalize(analytic_fn_evals_, curr_tuple_, dummy_result_tuple_);
  }
  AggFnEvaluator::Close(analytic_fn_evals_, state);
  AggFn::Close(analytic_fns_);

  if (partition_by_eq_expr_ != nullptr) {
    if (partition_by_eq_expr_eval_ != nullptr) {
      partition_by_eq_expr_eval_->Close(state);
    }
    partition_by_eq_expr_->Close();
  }
  if (order_by_eq_expr_ != nullptr) {
    if (order_by_eq_expr_eval_ != nullptr) order_by_eq_expr_eval_->Close(state);
    order_by_eq_expr_->Close();
  }
  if (curr_child_batch_.get() != nullptr) curr_child_batch_.reset();
  if (curr_tuple_pool_.get() != nullptr) curr_tuple_pool_->FreeAll();
  if (prev_tuple_pool_.get() != nullptr) prev_tuple_pool_->FreeAll();
  if (prev_input_tuple_pool_.get() != nullptr) prev_input_tuple_pool_->FreeAll();
  ExecNode::Close(state);
}

void AnalyticEvalNode::DebugString(int indentation_level, stringstream* out) const {
  *out << string(indentation_level * 2, ' ');
  *out << "AnalyticEvalNode("
       << " window=" << DebugWindowString();
  if (partition_by_eq_expr_ != nullptr) {
    *out << " partition_exprs=" << partition_by_eq_expr_->DebugString();
  }
  if (order_by_eq_expr_ != nullptr) {
    *out << " order_by_exprs=" << order_by_eq_expr_->DebugString();
  }
  *out << AggFn::DebugString(analytic_fns_);
  ExecNode::DebugString(indentation_level, out);
  *out << ")";
}
}