| 1 | #pragma once |
| 2 | |
| 3 | #include <boost/noncopyable.hpp> |
| 4 | |
| 5 | #include <Common/HyperLogLogCounter.h> |
| 6 | #include <Common/HashTable/SmallTable.h> |
| 7 | |
| 8 | |
| 9 | namespace DB |
| 10 | { |
| 11 | |
| 12 | |
| 13 | /** For a small number of keys - an array of fixed size "on the stack". |
| 14 | * For large, HyperLogLog is allocated. |
| 15 | * See also the more practical implementation in CombinedCardinalityEstimator.h, |
| 16 | * where a hash table is also used for medium-sized sets. |
| 17 | */ |
| 18 | template |
| 19 | < |
| 20 | typename Key, |
| 21 | UInt8 small_set_size, |
| 22 | UInt8 K, |
| 23 | typename Hash = IntHash32<Key>, |
| 24 | typename DenominatorType = double> |
| 25 | class HyperLogLogWithSmallSetOptimization : private boost::noncopyable |
| 26 | { |
| 27 | private: |
| 28 | using Small = SmallSet<Key, small_set_size>; |
| 29 | using Large = HyperLogLogCounter<K, Hash, UInt32, DenominatorType>; |
| 30 | |
| 31 | Small small; |
| 32 | Large * large = nullptr; |
| 33 | |
| 34 | bool isLarge() const |
| 35 | { |
| 36 | return large != nullptr; |
| 37 | } |
| 38 | |
| 39 | void toLarge() |
| 40 | { |
| 41 | /// At the time of copying data from `tiny`, setting the value of `large` is still not possible (otherwise it will overwrite some data). |
| 42 | Large * tmp_large = new Large; |
| 43 | |
| 44 | for (const auto & x : small) |
| 45 | tmp_large->insert(x.getValue()); |
| 46 | |
| 47 | large = tmp_large; |
| 48 | } |
| 49 | |
| 50 | public: |
| 51 | using value_type = Key; |
| 52 | |
| 53 | ~HyperLogLogWithSmallSetOptimization() |
| 54 | { |
| 55 | if (isLarge()) |
| 56 | delete large; |
| 57 | } |
| 58 | |
| 59 | /// ALWAYS_INLINE is required to have better code layout for uniqHLL12 function |
| 60 | void ALWAYS_INLINE insert(Key value) |
| 61 | { |
| 62 | if (!isLarge()) |
| 63 | { |
| 64 | if (small.find(value) == small.end()) |
| 65 | { |
| 66 | if (!small.full()) |
| 67 | small.insert(value); |
| 68 | else |
| 69 | { |
| 70 | toLarge(); |
| 71 | large->insert(value); |
| 72 | } |
| 73 | } |
| 74 | } |
| 75 | else |
| 76 | large->insert(value); |
| 77 | } |
| 78 | |
| 79 | UInt64 size() const |
| 80 | { |
| 81 | return !isLarge() ? small.size() : large->size(); |
| 82 | } |
| 83 | |
| 84 | void merge(const HyperLogLogWithSmallSetOptimization & rhs) |
| 85 | { |
| 86 | if (rhs.isLarge()) |
| 87 | { |
| 88 | if (!isLarge()) |
| 89 | toLarge(); |
| 90 | |
| 91 | large->merge(*rhs.large); |
| 92 | } |
| 93 | else |
| 94 | { |
| 95 | for (const auto & x : rhs.small) |
| 96 | insert(x.getValue()); |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | /// You can only call for an empty object. |
| 101 | void read(DB::ReadBuffer & in) |
| 102 | { |
| 103 | bool is_large; |
| 104 | readBinary(is_large, in); |
| 105 | |
| 106 | if (is_large) |
| 107 | { |
| 108 | toLarge(); |
| 109 | large->read(in); |
| 110 | } |
| 111 | else |
| 112 | small.read(in); |
| 113 | } |
| 114 | |
| 115 | void readAndMerge(DB::ReadBuffer & in) |
| 116 | { |
| 117 | bool is_rhs_large; |
| 118 | readBinary(is_rhs_large, in); |
| 119 | |
| 120 | if (!isLarge() && is_rhs_large) |
| 121 | toLarge(); |
| 122 | |
| 123 | if (!is_rhs_large) |
| 124 | { |
| 125 | typename Small::Reader reader(in); |
| 126 | while (reader.next()) |
| 127 | insert(reader.get()); |
| 128 | } |
| 129 | else |
| 130 | large->readAndMerge(in); |
| 131 | } |
| 132 | |
| 133 | void write(DB::WriteBuffer & out) const |
| 134 | { |
| 135 | writeBinary(isLarge(), out); |
| 136 | |
| 137 | if (isLarge()) |
| 138 | large->write(out); |
| 139 | else |
| 140 | small.write(out); |
| 141 | } |
| 142 | }; |
| 143 | |
| 144 | |
| 145 | } |
| 146 | |