DPL Analysis: shortcircuit creation of empty slices

Framework/Core/include/Framework/ASoA.h

@@ -1517,6 +1517,11 @@ struct PreslicePolicySorted : public PreslicePolicyBase {
 
   SliceInfoPtr sliceInfo;
   std::shared_ptr<arrow::Table> getSliceFor(int value, std::shared_ptr<arrow::Table> const& input, uint64_t& offset) const;
+  // One-slot cache for the empty (0-row) slice, so that empty groups do not
+  // slice every column only to produce 0 rows (the common case for sparse
+  // grouping, e.g. candidates per collision). Keyed by the input table, which
+  // changes with every dataframe.
+  mutable std::pair<arrow::Table const*, std::shared_ptr<arrow::Table>> emptySlice{nullptr, nullptr};
 };
 
 struct PreslicePolicyGeneral : public PreslicePolicyBase {
@@ -1731,7 +1736,10 @@ auto doSliceByCached(T const* table, framework::expressions::BindingNode const&
   auto localCache = cache.ptr->getCacheFor({"", originReplacement(cache.ptr->newOrigin)(o2::soa::getMatcherFromTypeForKey<T>(node.name)),
                                             node.name});
   auto [offset, count] = localCache.getSliceFor(value);
-  auto t = typename T::self_t({table->asArrowTable()->Slice(static_cast<uint64_t>(offset), count)}, static_cast<uint64_t>(offset));
+  // Empty group: reuse a cached empty (0-row) table instead of slicing every column.
+  auto slice = count == 0 ? cache.ptr->getEmptySliceFor(table->asArrowTable())
+                          : table->asArrowTable()->Slice(static_cast<uint64_t>(offset), count);
+  auto t = typename T::self_t({slice}, static_cast<uint64_t>(offset));
   if (t.tableSize() != 0) {
     table->copyIndexBindings(t);
   }
@@ -1744,7 +1752,9 @@ auto doFilteredSliceByCached(T const* table, framework::expressions::BindingNode
   auto localCache = cache.ptr->getCacheFor({"", originReplacement(cache.ptr->newOrigin)(o2::soa::getMatcherFromTypeForKey<T>(node.name)),
                                             node.name});
   auto [offset, count] = localCache.getSliceFor(value);
-  auto slice = table->asArrowTable()->Slice(static_cast<uint64_t>(offset), count);
+  // Empty group: reuse a cached empty (0-row) table instead of slicing every column.
+  auto slice = count == 0 ? cache.ptr->getEmptySliceFor(table->asArrowTable())
+                          : table->asArrowTable()->Slice(static_cast<uint64_t>(offset), count);
   return prepareFilteredSlice(table, slice, offset);
 }
 

Framework/Core/include/Framework/ArrowTableSlicingCache.h

@@ -107,6 +107,12 @@ struct ArrowTableSlicingCache {
   SliceInfoPtr getCacheForPos(int pos) const;
   SliceInfoUnsortedPtr getCacheUnsortedForPos(int pos) const;
 
+  // get a cached empty (0-row) slice of the given table, so that empty groups
+  // do not slice every column only to produce 0 rows (the common case for
+  // sparse grouping). One-slot cache keyed by the table pointer.
+  std::shared_ptr<arrow::Table> getEmptySliceFor(std::shared_ptr<arrow::Table> const& table);
+  std::pair<arrow::Table const*, std::shared_ptr<arrow::Table>> emptySlice{nullptr, nullptr};
+
   static void validateOrder(Entry const& bindingKey, std::shared_ptr<arrow::Table> const& input);
 };
 } // namespace o2::framework

Framework/Core/include/Framework/GroupSlicer.h

@@ -218,10 +218,16 @@ struct GroupSlicer {
       auto oc = sliceInfos[index].getSliceFor(pos);
       uint64_t offset = oc.first;
       auto count = oc.second;
-      auto groupedElementsTable = originalTable.asArrowTable()->Slice(offset, count);
       if (count == 0) {
-        return std::decay_t<A1>{{groupedElementsTable}, soa::SelectionVector{}};
+        // Empty group: avoid slicing every column only to discard it. Cache one
+        // empty (0-row) table per associated table and reuse it. This is the
+        // common case for sparse grouping (e.g. collisions with no candidates).
+        if (!emptyTables[index]) {
+          emptyTables[index] = originalTable.asArrowTable()->Slice(0, 0);
+        }
+        return std::decay_t<A1>{{emptyTables[index]}, soa::SelectionVector{}};
       }
+      auto groupedElementsTable = originalTable.asArrowTable()->Slice(offset, count);
 
       // for each grouping element we need to slice the selection vector
       auto start_iterator = std::lower_bound(starts[index], selections[index]->end(), offset);
@@ -275,6 +281,9 @@ struct GroupSlicer {
     std::span<int64_t const> groupSelection;
     std::array<std::span<int64_t const> const*, sizeof...(A)> selections;
     std::array<std::span<int64_t const>::iterator, sizeof...(A)> starts;
+    // Cached empty (0-row) table per associated table, lazily built and reused
+    // for empty groups so we do not slice every column on each empty group.
+    std::array<std::shared_ptr<arrow::Table>, sizeof...(A)> emptyTables{};
 
     std::array<SliceInfoPtr, sizeof...(A)> sliceInfos;
     std::array<SliceInfoUnsortedPtr, sizeof...(A)> sliceInfosUnsorted;

Framework/Core/src/ASoA.cxx

@@ -317,9 +317,16 @@ void PreslicePolicyGeneral::updateSliceInfo(SliceInfoUnsortedPtr&& si)
 std::shared_ptr<arrow::Table> PreslicePolicySorted::getSliceFor(int value, std::shared_ptr<arrow::Table> const& input, uint64_t& offset) const
 {
   auto [offset_, count] = this->sliceInfo.getSliceFor(value);
-  auto output = input->Slice(offset_, count);
   offset = static_cast<int64_t>(offset_);
-  return output;
+  if (count == 0) {
+    // Empty group: avoid slicing every column only to discard it. Cache one
+    // empty (0-row) table per input table and reuse it (see GroupSlicer).
+    if (emptySlice.first != input.get()) {
+      emptySlice = {input.get(), input->Slice(0, 0)};
+    }
+    return emptySlice.second;
+  }
+  return input->Slice(offset_, count);
 }
 
 std::span<const int64_t> PreslicePolicyGeneral::getSliceFor(int value) const

Framework/Core/src/ArrowTableSlicingCache.cxx

@@ -269,6 +269,14 @@ SliceInfoUnsortedPtr ArrowTableSlicingCache::getCacheUnsortedForPos(int pos) con
   };
 }
 
+std::shared_ptr<arrow::Table> ArrowTableSlicingCache::getEmptySliceFor(std::shared_ptr<arrow::Table> const& table)
+{
+  if (emptySlice.first != table.get()) {
+    emptySlice = {table.get(), table->Slice(0, 0)};
+  }
+  return emptySlice.second;
+}
+
 void ArrowTableSlicingCache::validateOrder(Entry const& bindingKey, const std::shared_ptr<arrow::Table>& input)
 {
   auto const& [target, matcher, key, enabled] = bindingKey;