r/src/compute-exec.cpp

// 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 "./arrow_types.h"

#if defined(ARROW_R_WITH_ACERO)

#include "./safe-call-into-r.h"

#include <arrow/acero/exec_plan.h>
#include <arrow/buffer.h>
#include <arrow/compute/api.h>
#include <arrow/compute/expression.h>
#include <arrow/table.h>
#include <arrow/util/async_generator.h>
#include <arrow/util/future.h>
#include <arrow/util/thread_pool.h>

#include <iostream>
#include <optional>

namespace acero = ::arrow::acero;
namespace compute = ::arrow::compute;

std::shared_ptr<compute::FunctionOptions> make_compute_options(std::string func_name,
                                                               cpp11::list options);

std::shared_ptr<arrow::KeyValueMetadata> strings_to_kvm(cpp11::strings metadata);

// [[acero::export]]
std::shared_ptr<acero::ExecPlan> ExecPlan_create(bool use_threads) {
  static compute::ExecContext threaded_context{gc_memory_pool(),
                                               arrow::internal::GetCpuThreadPool()};
  // TODO(weston) using gc_context() in this way is deprecated.  Once ordering has
  // been added we can probably entirely remove all reference to ExecPlan from R
  // in favor of DeclarationToXyz
  auto plan =
      ValueOrStop(acero::ExecPlan::Make(use_threads ? &threaded_context : gc_context()));

  return plan;
}

std::shared_ptr<acero::ExecNode> MakeExecNodeOrStop(
    const std::string& factory_name, acero::ExecPlan* plan,
    std::vector<acero::ExecNode*> inputs, const acero::ExecNodeOptions& options) {
  return std::shared_ptr<acero::ExecNode>(
      ValueOrStop(acero::MakeExecNode(factory_name, plan, std::move(inputs), options)),
      [](...) {
        // empty destructor: ExecNode lifetime is managed by an ExecPlan
      });
}

// This class is a special RecordBatchReader that holds a reference to the
// underlying exec plan so that (1) it can request that the ExecPlan *stop*
// producing when this object is deleted and (2) it can defer requesting
// the ExecPlan to *start* producing until the first batch has been pulled.
// This allows it to be transformed (e.g., using map_batches() or head())
// and queried (i.e., used as input to another ExecPlan), at the R level
// while maintaining the ability for the entire plan to be executed at once
// (e.g., to support user-defined functions) or never executed at all (e.g.,
// to support printing a nested ExecPlan without having to execute it).
class ExecPlanReader : public arrow::RecordBatchReader {
 public:
  enum ExecPlanReaderStatus { PLAN_NOT_STARTED, PLAN_RUNNING, PLAN_FINISHED };

  ExecPlanReader(const std::shared_ptr<arrow::acero::ExecPlan>& plan,
                 const std::shared_ptr<arrow::Schema>& schema,
                 arrow::AsyncGenerator<std::optional<compute::ExecBatch>> sink_gen)
      : schema_(schema),
        plan_(plan),
        sink_gen_(sink_gen),
        plan_status_(PLAN_NOT_STARTED),
        stop_token_(MainRThread::GetInstance().GetStopToken()) {}

  std::string PlanStatus() const {
    switch (plan_status_) {
      case PLAN_NOT_STARTED:
        return "PLAN_NOT_STARTED";
      case PLAN_RUNNING:
        return "PLAN_RUNNING";
      case PLAN_FINISHED:
        return "PLAN_FINISHED";
      default:
        return "UNKNOWN";
    }
  }

  std::shared_ptr<arrow::Schema> schema() const override { return schema_; }

  arrow::Status ReadNext(std::shared_ptr<arrow::RecordBatch>* batch_out) override {
    // If this is the first batch getting pulled, tell the exec plan to
    // start producing
    if (plan_status_ == PLAN_NOT_STARTED) {
      StartProducing();
    }

    // If we've closed the reader, keep sending nullptr
    // (consistent with what most RecordBatchReader subclasses do)
    if (plan_status_ == PLAN_FINISHED) {
      batch_out->reset();
      return arrow::Status::OK();
    }

    // Check for cancellation and stop the plan if we have a request. When
    // the ExecPlan supports passing a StopToken and handling this itself,
    // this will be redundant.
    if (stop_token_.IsStopRequested()) {
      StopProducing();
      return stop_token_.Poll();
    }

    auto out = sink_gen_().result();
    if (!out.ok()) {
      StopProducing();
      return out.status();
    }

    if (out.ValueUnsafe()) {
      auto batch_result = out.ValueUnsafe()->ToRecordBatch(schema_, gc_memory_pool());
      if (!batch_result.ok()) {
        StopProducing();
        return batch_result.status();
      }

      *batch_out = batch_result.ValueUnsafe();
    } else {
      batch_out->reset();
      plan_status_ = PLAN_FINISHED;
      return plan_->finished().status();
    }

    return arrow::Status::OK();
  }

  arrow::Status Close() override {
    StopProducing();
    return arrow::Status::OK();
  }

  const std::shared_ptr<arrow::acero::ExecPlan>& Plan() const { return plan_; }

  ~ExecPlanReader() { StopProducing(); }

 private:
  std::shared_ptr<arrow::Schema> schema_;
  std::shared_ptr<arrow::acero::ExecPlan> plan_;
  arrow::AsyncGenerator<std::optional<compute::ExecBatch>> sink_gen_;
  ExecPlanReaderStatus plan_status_;
  arrow::StopToken stop_token_;

  void StartProducing() {
    plan_->StartProducing();
    plan_status_ = PLAN_RUNNING;
  }

  void StopProducing() {
    if (plan_status_ == PLAN_RUNNING) {
      // We're done with the plan, but it may still need some time
      // to finish and clean up after itself. To do this, we give a
      // callable with its own copy of the shared_ptr<ExecPlan> so
      // that it can delete itself when it is safe to do so.
      std::shared_ptr<arrow::acero::ExecPlan> plan(plan_);
      bool not_finished_yet = plan_->finished().TryAddCallback(
          [&plan] { return [plan](const arrow::Status&) {}; });

      if (not_finished_yet) {
        plan_->StopProducing();
      }
    }

    plan_status_ = PLAN_FINISHED;
    // A previous version of this called plan_.reset() and reset
    // sink_gen_ to an empty generator; however, this caused
    // crashes on some platforms.
  }
};

// [[acero::export]]
cpp11::list ExecPlanReader__batches(
    const std::shared_ptr<arrow::RecordBatchReader>& reader) {
  auto result = RunWithCapturedRIfPossible<arrow::RecordBatchVector>(
      [&]() { return reader->ToRecordBatches(); });
  return arrow::r::to_r_list(ValueOrStop(result));
}

// [[acero::export]]
std::shared_ptr<arrow::Table> Table__from_ExecPlanReader(
    const std::shared_ptr<arrow::RecordBatchReader>& reader) {
  auto result = RunWithCapturedRIfPossible<std::shared_ptr<arrow::Table>>(
      [&]() { return reader->ToTable(); });

  return ValueOrStop(result);
}

// [[acero::export]]
std::shared_ptr<acero::ExecPlan> ExecPlanReader__Plan(
    const std::shared_ptr<ExecPlanReader>& reader) {
  if (reader->PlanStatus() == "PLAN_FINISHED") {
    cpp11::stop("Can't extract ExecPlan from a finished ExecPlanReader");
  }

  return reader->Plan();
}

// [[acero::export]]
std::string ExecPlanReader__PlanStatus(const std::shared_ptr<ExecPlanReader>& reader) {
  return reader->PlanStatus();
}

// [[acero::export]]
std::shared_ptr<ExecPlanReader> ExecPlan_run(
    const std::shared_ptr<acero::ExecPlan>& plan,
    const std::shared_ptr<acero::ExecNode>& final_node, cpp11::strings metadata) {
  // For now, don't require R to construct SinkNodes.
  // Instead, just pass the node we should collect as an argument.
  arrow::AsyncGenerator<std::optional<compute::ExecBatch>> sink_gen;

  MakeExecNodeOrStop("sink", plan.get(), {final_node.get()},
                     acero::SinkNodeOptions{&sink_gen});

  StopIfNotOk(plan->Validate());

  // Attach metadata to the schema
  auto out_schema = final_node->output_schema();
  if (metadata.size() > 0) {
    auto kv = strings_to_kvm(metadata);
    out_schema = out_schema->WithMetadata(kv);
  }

  return std::make_shared<ExecPlanReader>(plan, out_schema, sink_gen);
}

// [[acero::export]]
std::string ExecPlan_ToString(const std::shared_ptr<acero::ExecPlan>& plan) {
  return plan->ToString();
}

// [[acero::export]]
void ExecPlan_UnsafeDelete(const std::shared_ptr<acero::ExecPlan>& plan) {
  auto& plan_unsafe = const_cast<std::shared_ptr<acero::ExecPlan>&>(plan);
  plan_unsafe.reset();
}

// [[acero::export]]
std::shared_ptr<arrow::Schema> ExecNode_output_schema(
    const std::shared_ptr<acero::ExecNode>& node) {
  return node->output_schema();
}

// [[acero::export]]
bool ExecNode_has_ordered_batches(const std::shared_ptr<acero::ExecNode>& node) {
  return !node->ordering().is_unordered();
}

#if defined(ARROW_R_WITH_DATASET)

#include <arrow/dataset/file_base.h>
#include <arrow/dataset/plan.h>
#include <arrow/dataset/scanner.h>

// [[dataset::export]]
std::shared_ptr<acero::ExecNode> ExecNode_Scan(
    const std::shared_ptr<acero::ExecPlan>& plan,
    const std::shared_ptr<ds::Dataset>& dataset,
    const std::shared_ptr<compute::Expression>& filter, cpp11::list projection) {
  arrow::dataset::internal::Initialize();

  // TODO: pass in FragmentScanOptions
  auto options = std::make_shared<ds::ScanOptions>();

  options->use_threads = arrow::r::GetBoolOption("arrow.use_threads", true);
  options->dataset_schema = dataset->schema();

  // This filter is only used for predicate pushdown;
  // you still need to pass it to a FilterNode after to handle any other components
  options->filter = *filter;

  // ScanNode needs to know which fields to materialize.
  // It will pull them from this projection to prune the scan,
  // but you still need to Project after
  std::vector<compute::Expression> exprs;
  for (SEXP expr : projection) {
    auto expr_ptr = cpp11::as_cpp<std::shared_ptr<compute::Expression>>(expr);
    exprs.push_back(*expr_ptr);
  }
  cpp11::strings field_names(projection.attr(R_NamesSymbol));
  options->projection = call(
      "make_struct", std::move(exprs),
      compute::MakeStructOptions{cpp11::as_cpp<std::vector<std::string>>(field_names)});

  return MakeExecNodeOrStop("scan", plan.get(), {},
                            ds::ScanNodeOptions{dataset, options});
}

// [[dataset::export]]
void ExecPlan_Write(const std::shared_ptr<acero::ExecPlan>& plan,
                    const std::shared_ptr<acero::ExecNode>& final_node,
                    const std::shared_ptr<arrow::Schema>& schema,
                    const std::shared_ptr<ds::FileWriteOptions>& file_write_options,
                    const std::shared_ptr<fs::FileSystem>& filesystem,
                    std::string base_dir,
                    const std::shared_ptr<ds::Partitioning>& partitioning,
                    std::string basename_template,
                    arrow::dataset::ExistingDataBehavior existing_data_behavior,
                    int max_partitions, uint32_t max_open_files,
                    uint64_t max_rows_per_file, uint64_t min_rows_per_group,
                    uint64_t max_rows_per_group, bool create_directory) {
  arrow::dataset::internal::Initialize();

  // TODO(ARROW-16200): expose FileSystemDatasetWriteOptions in R
  // and encapsulate this logic better
  ds::FileSystemDatasetWriteOptions opts;
  opts.file_write_options = file_write_options;
  opts.existing_data_behavior = existing_data_behavior;
  opts.filesystem = filesystem;
  opts.base_dir = base_dir;
  opts.partitioning = partitioning;
  opts.basename_template = basename_template;
  opts.max_partitions = max_partitions;
  opts.max_open_files = max_open_files;
  opts.max_rows_per_file = max_rows_per_file;
  opts.min_rows_per_group = min_rows_per_group;
  opts.max_rows_per_group = max_rows_per_group;
  opts.create_dir = create_directory;

  ds::WriteNodeOptions options(std::move(opts));
  options.custom_schema = std::move(schema);

  MakeExecNodeOrStop("write", final_node->plan(), {final_node.get()}, std::move(options));

  StopIfNotOk(plan->Validate());

  arrow::Status result = RunWithCapturedRIfPossibleVoid([&]() {
    plan->StartProducing();
    return plan->finished().status();
  });

  StopIfNotOk(result);
}

#endif

// [[acero::export]]
std::shared_ptr<acero::ExecNode> ExecNode_Filter(
    const std::shared_ptr<acero::ExecNode>& input,
    const std::shared_ptr<compute::Expression>& filter) {
  return MakeExecNodeOrStop("filter", input->plan(), {input.get()},
                            acero::FilterNodeOptions{*filter});
}

// [[acero::export]]
std::shared_ptr<acero::ExecNode> ExecNode_Project(
    const std::shared_ptr<acero::ExecNode>& input,
    const std::vector<std::shared_ptr<compute::Expression>>& exprs,
    std::vector<std::string> names) {
  // We have shared_ptrs of expressions but need the Expressions
  std::vector<compute::Expression> expressions;
  for (auto expr : exprs) {
    expressions.push_back(*expr);
  }
  return MakeExecNodeOrStop(
      "project", input->plan(), {input.get()},
      acero::ProjectNodeOptions{std::move(expressions), std::move(names)});
}

// [[acero::export]]
std::shared_ptr<acero::ExecNode> ExecNode_Aggregate(
    const std::shared_ptr<acero::ExecNode>& input, cpp11::list options,
    std::vector<std::string> key_names) {
  std::vector<arrow::compute::Aggregate> aggregates;

  for (cpp11::list name_opts : options) {
    auto function = cpp11::as_cpp<std::string>(name_opts["fun"]);
    auto opts = make_compute_options(function, name_opts["options"]);
    auto target_names = cpp11::as_cpp<std::vector<std::string>>(name_opts["targets"]);
    auto name = cpp11::as_cpp<std::string>(name_opts["name"]);

    std::vector<arrow::FieldRef> targets;
    for (auto&& target : target_names) {
      targets.emplace_back(std::move(target));
    }
    aggregates.push_back(arrow::compute::Aggregate{std::move(function), opts,
                                                   std::move(targets), std::move(name)});
  }

  std::vector<arrow::FieldRef> keys;
  for (auto&& name : key_names) {
    keys.emplace_back(std::move(name));
  }
  return MakeExecNodeOrStop(
      "aggregate", input->plan(), {input.get()},
      acero::AggregateNodeOptions{std::move(aggregates), std::move(keys)});
}

// [[acero::export]]
std::shared_ptr<acero::ExecNode> ExecNode_Join(
    const std::shared_ptr<acero::ExecNode>& input, acero::JoinType join_type,
    const std::shared_ptr<acero::ExecNode>& right_data,
    std::vector<std::string> left_keys, std::vector<std::string> right_keys,
    std::vector<std::string> left_output, std::vector<std::string> right_output,
    std::string output_suffix_for_left, std::string output_suffix_for_right,
    bool na_matches) {
  std::vector<arrow::FieldRef> left_refs, right_refs, left_out_refs, right_out_refs;
  std::vector<acero::JoinKeyCmp> key_cmps;
  for (auto&& name : left_keys) {
    left_refs.emplace_back(std::move(name));
    // Populate key_cmps in this loop, one for each key
    // Note that Acero supports having different values for each key, but dplyr
    // only supports one value for all keys, so we're only going to support that
    // for now.
    key_cmps.emplace_back(na_matches ? acero::JoinKeyCmp::IS : acero::JoinKeyCmp::EQ);
  }
  for (auto&& name : right_keys) {
    right_refs.emplace_back(std::move(name));
  }
  for (auto&& name : left_output) {
    left_out_refs.emplace_back(std::move(name));
  }
  // dplyr::semi_join => LEFT_SEMI; dplyr::anti_join => LEFT_ANTI
  // So ignoring RIGHT_SEMI and RIGHT_ANTI here because dplyr doesn't implement them.
  if (join_type != acero::JoinType::LEFT_SEMI &&
      join_type != acero::JoinType::LEFT_ANTI) {
    // Don't include out_refs in semi/anti join
    for (auto&& name : right_output) {
      right_out_refs.emplace_back(std::move(name));
    }
  }

  return MakeExecNodeOrStop(
      "hashjoin", input->plan(), {input.get(), right_data.get()},
      acero::HashJoinNodeOptions{join_type, std::move(left_refs), std::move(right_refs),
                                 std::move(left_out_refs), std::move(right_out_refs),
                                 std::move(key_cmps), compute::literal(true),
                                 std::move(output_suffix_for_left),
                                 std::move(output_suffix_for_right)});
}

// [[acero::export]]
std::shared_ptr<acero::ExecNode> ExecNode_Union(
    const std::shared_ptr<acero::ExecNode>& input,
    const std::shared_ptr<acero::ExecNode>& right_data) {
  return MakeExecNodeOrStop("union", input->plan(), {input.get(), right_data.get()}, {});
}

// [[acero::export]]
std::shared_ptr<acero::ExecNode> ExecNode_Fetch(
    const std::shared_ptr<acero::ExecNode>& input, int64_t offset, int64_t limit) {
  return MakeExecNodeOrStop("fetch", input->plan(), {input.get()},
                            acero::FetchNodeOptions{offset, limit});
}

// [[acero::export]]
std::shared_ptr<acero::ExecNode> ExecNode_OrderBy(
    const std::shared_ptr<acero::ExecNode>& input, cpp11::list sort_options) {
  return MakeExecNodeOrStop(
      "order_by", input->plan(), {input.get()},
      acero::OrderByNodeOptions{std::dynamic_pointer_cast<compute::SortOptions>(
                                    make_compute_options("sort_indices", sort_options))
                                    ->AsOrdering()});
}

// [[acero::export]]
std::shared_ptr<acero::ExecNode> ExecNode_SourceNode(
    const std::shared_ptr<acero::ExecPlan>& plan,
    const std::shared_ptr<arrow::RecordBatchReader>& reader) {
  arrow::acero::RecordBatchReaderSourceNodeOptions options{reader};
  return MakeExecNodeOrStop("record_batch_reader_source", plan.get(), {}, options);
}

// [[acero::export]]
std::shared_ptr<acero::ExecNode> ExecNode_TableSourceNode(
    const std::shared_ptr<acero::ExecPlan>& plan,
    const std::shared_ptr<arrow::Table>& table) {
  arrow::acero::TableSourceNodeOptions options{/*table=*/table,
                                               // TODO: make batch_size configurable
                                               /*batch_size=*/1048576};

  return MakeExecNodeOrStop("table_source", plan.get(), {}, options);
}

#endif

#if defined(ARROW_R_WITH_SUBSTRAIT)

#include <arrow/engine/substrait/api.h>

// Just for example usage until a C++ method is available that implements
// a RecordBatchReader output (ARROW-15849)
class AccumulatingConsumer : public acero::SinkNodeConsumer {
 public:
  const std::vector<std::shared_ptr<arrow::RecordBatch>>& batches() { return batches_; }

  arrow::Status Init(const std::shared_ptr<arrow::Schema>& schema,
                     acero::BackpressureControl* backpressure_control,
                     acero::ExecPlan* exec_plan) override {
    schema_ = schema;
    return arrow::Status::OK();
  }

  arrow::Status Consume(compute::ExecBatch batch) override {
    auto record_batch = batch.ToRecordBatch(schema_);
    ARROW_RETURN_NOT_OK(record_batch);
    batches_.push_back(record_batch.ValueUnsafe());

    return arrow::Status::OK();
  }

  arrow::Future<> Finish() override { return arrow::Future<>::MakeFinished(); }

 private:
  std::shared_ptr<arrow::Schema> schema_;
  std::vector<std::shared_ptr<arrow::RecordBatch>> batches_;
};

// Expose these so that it's easier to write tests

// [[substrait::export]]
std::string substrait__internal__SubstraitToJSON(
    const std::shared_ptr<arrow::Buffer>& serialized_plan) {
  return ValueOrStop(arrow::engine::internal::SubstraitToJSON("Plan", *serialized_plan));
}

// [[substrait::export]]
std::shared_ptr<arrow::Buffer> substrait__internal__SubstraitFromJSON(
    std::string substrait_json) {
  return ValueOrStop(arrow::engine::internal::SubstraitFromJSON("Plan", substrait_json));
}

// [[substrait::export]]
std::shared_ptr<arrow::Table> ExecPlan_run_substrait(
    const std::shared_ptr<acero::ExecPlan>& plan,
    const std::shared_ptr<arrow::Buffer>& serialized_plan) {
  std::vector<std::shared_ptr<AccumulatingConsumer>> consumers;

  std::function<std::shared_ptr<acero::SinkNodeConsumer>()> consumer_factory = [&] {
    consumers.emplace_back(new AccumulatingConsumer());
    return consumers.back();
  };

  arrow::Result<std::vector<acero::Declaration>> maybe_decls =
      ValueOrStop(arrow::engine::DeserializePlans(*serialized_plan, consumer_factory));
  std::vector<acero::Declaration> decls = std::move(ValueOrStop(maybe_decls));

  // For now, the Substrait plan must include a 'read' that points to
  // a Parquet file (instead of using a source node create in Arrow)
  for (const acero::Declaration& decl : decls) {
    auto node = decl.AddToPlan(plan.get());
    StopIfNotOk(node.status());
  }

  StopIfNotOk(plan->Validate());
  plan->StartProducing();
  StopIfNotOk(plan->finished().status());

  std::vector<std::shared_ptr<arrow::RecordBatch>> all_batches;
  for (const auto& consumer : consumers) {
    for (const auto& batch : consumer->batches()) {
      all_batches.push_back(batch);
    }
  }

  return ValueOrStop(arrow::Table::FromRecordBatches(std::move(all_batches)));
}

#endif