OmniSciDB  c07336695a
JoinFilterPushDown.cpp File Reference
#include "JoinFilterPushDown.h"
#include "DeepCopyVisitor.h"
#include "RelAlgExecutor.h"
+ Include dependency graph for JoinFilterPushDown.cpp:

Go to the source code of this file.

Classes

class  anonymous_namespace{JoinFilterPushDown.cpp}::BindFilterToOutermostVisitor
 
class  anonymous_namespace{JoinFilterPushDown.cpp}::CollectInputColumnsVisitor
 

Namespaces

 anonymous_namespace{JoinFilterPushDown.cpp}
 

Functions

bool to_gather_info_for_filter_selectivity (const std::vector< InputTableInfo > &table_infos)
 
std::vector< PushedDownFilterInfofind_push_down_filters (const RelAlgExecutionUnit &ra_exe_unit, const std::vector< size_t > &input_permutation, const std::vector< size_t > &left_deep_join_input_sizes)
 

Function Documentation

◆ find_push_down_filters()

std::vector<PushedDownFilterInfo> find_push_down_filters ( const RelAlgExecutionUnit ra_exe_unit,
const std::vector< size_t > &  input_permutation,
const std::vector< size_t > &  left_deep_join_input_sizes 
)

Go through all tables involved in the relational algebra plan, and select potential candidates to be pushed down by calcite. For each filter we store a set of intermediate indices (previous, current, and next table) based on the column indices in their query string.

Definition at line 232 of file JoinFilterPushDown.cpp.

References CHECK_EQ, CHECK_GE, CHECK_LT, RelAlgExecutionUnit::input_descs, RelAlgExecutionUnit::join_quals, run-benchmark-import::result, and ScalarExprVisitor< T >::visit().

Referenced by RelAlgExecutor::selectFiltersToBePushedDown().

235  {
236  std::vector<PushedDownFilterInfo> result;
237  if (left_deep_join_input_sizes.empty()) {
238  return result;
239  }
240  std::vector<size_t> input_size_prefix_sums(left_deep_join_input_sizes.size());
241  std::partial_sum(left_deep_join_input_sizes.begin(),
242  left_deep_join_input_sizes.end(),
243  input_size_prefix_sums.begin());
244  std::vector<int> to_original_rte_idx(ra_exe_unit.input_descs.size(),
245  ra_exe_unit.input_descs.size());
246  if (!input_permutation.empty()) {
247  CHECK_EQ(to_original_rte_idx.size(), input_permutation.size());
248  for (size_t i = 0; i < input_permutation.size(); ++i) {
249  CHECK_LT(input_permutation[i], to_original_rte_idx.size());
250  CHECK_EQ(static_cast<size_t>(to_original_rte_idx[input_permutation[i]]),
251  to_original_rte_idx.size());
252  to_original_rte_idx[input_permutation[i]] = i;
253  }
254  } else {
255  std::iota(to_original_rte_idx.begin(), to_original_rte_idx.end(), 0);
256  }
257  std::unordered_map<int, std::vector<std::shared_ptr<Analyzer::Expr>>>
258  filters_per_nesting_level;
259  for (const auto& level_conditions : ra_exe_unit.join_quals) {
261  for (const auto& cond : level_conditions.quals) {
262  const auto rte_indices = visitor.visit(cond.get());
263  if (rte_indices.size() > 1) {
264  continue;
265  }
266  const int rte_idx = (!rte_indices.empty()) ? *rte_indices.cbegin() : 0;
267  if (!rte_idx) {
268  continue;
269  }
270  CHECK_GE(rte_idx, 0);
271  CHECK_LT(static_cast<size_t>(rte_idx), to_original_rte_idx.size());
272  filters_per_nesting_level[to_original_rte_idx[rte_idx]].push_back(cond);
273  }
274  }
275  for (const auto& kv : filters_per_nesting_level) {
276  CHECK_GE(kv.first, 0);
277  CHECK_LT(static_cast<size_t>(kv.first), input_size_prefix_sums.size());
278  size_t input_prev = (kv.first > 1) ? input_size_prefix_sums[kv.first - 2] : 0;
279  size_t input_start = kv.first ? input_size_prefix_sums[kv.first - 1] : 0;
280  size_t input_next = input_size_prefix_sums[kv.first];
281  result.emplace_back(
282  PushedDownFilterInfo{kv.second, input_prev, input_start, input_next});
283  }
284  return result;
285 }
#define CHECK_EQ(x, y)
Definition: Logger.h:195
#define CHECK_GE(x, y)
Definition: Logger.h:200
const std::vector< InputDescriptor > input_descs
const JoinQualsPerNestingLevel join_quals
#define CHECK_LT(x, y)
Definition: Logger.h:197
T visit(const Analyzer::Expr *expr) const
+ Here is the call graph for this function:
+ Here is the caller graph for this function:

◆ to_gather_info_for_filter_selectivity()

bool to_gather_info_for_filter_selectivity ( const std::vector< InputTableInfo > &  table_infos)

The main purpose of this function is to prevent going through extra overhead of computing required statistics for finding the right candidates and then the actual push-down, unless the problem is large enough that such effort is potentially helpful.

Definition at line 205 of file JoinFilterPushDown.cpp.

Referenced by RelAlgExecutor::selectFiltersToBePushedDown().

206  {
207  if (table_infos.size() < 2) {
208  return false;
209  }
210  // we currently do not support filter push down when there is a self-join involved:
211  // TODO(Saman): prevent Calcite from optimizing self-joins to remove this exclusion
212  std::unordered_set<int> table_ids;
213  for (auto ti : table_infos) {
214  if (table_ids.find(ti.table_id) == table_ids.end()) {
215  table_ids.insert(ti.table_id);
216  } else {
217  // a self-join is involved
218  return false;
219  }
220  }
221  // TODO(Saman): add some extra heuristics to avoid preflight count and push down if it
222  // is not going to be helpful.
223  return true;
224 }
+ Here is the caller graph for this function: