1 | //===----------------------------------------------------------------------===// |
2 | // DuckDB |
3 | // |
4 | // duckdb/common/hive_partitioning.hpp |
5 | // |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #pragma once |
10 | |
11 | #include "duckdb/common/types/column/partitioned_column_data.hpp" |
12 | #include "duckdb/execution/expression_executor.hpp" |
13 | #include "duckdb/optimizer/filter_combiner.hpp" |
14 | #include "duckdb/optimizer/statistics_propagator.hpp" |
15 | #include "duckdb/planner/expression_iterator.hpp" |
16 | #include "duckdb/planner/table_filter.hpp" |
17 | #include "re2/re2.h" |
18 | |
19 | #include <iostream> |
20 | #include <sstream> |
21 | |
22 | namespace duckdb { |
23 | |
24 | class HivePartitioning { |
25 | public: |
26 | //! Parse a filename that follows the hive partitioning scheme |
27 | DUCKDB_API static std::map<string, string> Parse(const string &filename); |
28 | DUCKDB_API static std::map<string, string> Parse(const string &filename, duckdb_re2::RE2 ®ex); |
29 | //! Prunes a list of filenames based on a set of filters, can be used by TableFunctions in the |
30 | //! pushdown_complex_filter function to skip files with filename-based filters. Also removes the filters that always |
31 | //! evaluate to true. |
32 | DUCKDB_API static void ApplyFiltersToFileList(ClientContext &context, vector<string> &files, |
33 | vector<unique_ptr<Expression>> &filters, |
34 | unordered_map<string, column_t> &column_map, idx_t table_index, |
35 | bool hive_enabled, bool filename_enabled); |
36 | |
37 | //! Returns the compiled regex pattern to match hive partitions |
38 | DUCKDB_API static const string REGEX_STRING; |
39 | }; |
40 | |
41 | struct HivePartitionKey { |
42 | //! Columns by which we want to partition |
43 | vector<Value> values; |
44 | //! Precomputed hash of values |
45 | hash_t hash; |
46 | |
47 | struct Hash { |
48 | std::size_t operator()(const HivePartitionKey &k) const { |
49 | return k.hash; |
50 | } |
51 | }; |
52 | |
53 | struct Equality { |
54 | bool operator()(const HivePartitionKey &a, const HivePartitionKey &b) const { |
55 | if (a.values.size() != b.values.size()) { |
56 | return false; |
57 | } |
58 | for (idx_t i = 0; i < a.values.size(); i++) { |
59 | if (!Value::NotDistinctFrom(lvalue: a.values[i], rvalue: b.values[i])) { |
60 | return false; |
61 | } |
62 | } |
63 | return true; |
64 | } |
65 | }; |
66 | }; |
67 | |
68 | //! Maps hive partitions to partition_ids |
69 | typedef unordered_map<HivePartitionKey, idx_t, HivePartitionKey::Hash, HivePartitionKey::Equality> hive_partition_map_t; |
70 | |
71 | //! class shared between HivePartitionColumnData classes that synchronizes partition discovery between threads. |
72 | //! each HivePartitionedColumnData will hold a local copy of the key->partition map |
73 | class GlobalHivePartitionState { |
74 | public: |
75 | mutex lock; |
76 | hive_partition_map_t partition_map; |
77 | //! Used for incremental updating local copies of the partition map; |
78 | vector<hive_partition_map_t::const_iterator> partitions; |
79 | }; |
80 | |
81 | class HivePartitionedColumnData : public PartitionedColumnData { |
82 | public: |
83 | HivePartitionedColumnData(ClientContext &context, vector<LogicalType> types, vector<idx_t> partition_by_cols, |
84 | shared_ptr<GlobalHivePartitionState> global_state = nullptr) |
85 | : PartitionedColumnData(PartitionedColumnDataType::HIVE, context, std::move(types)), |
86 | global_state(std::move(global_state)), group_by_columns(std::move(partition_by_cols)), |
87 | hashes_v(LogicalType::HASH) { |
88 | InitializeKeys(); |
89 | } |
90 | HivePartitionedColumnData(const HivePartitionedColumnData &other); |
91 | void ComputePartitionIndices(PartitionedColumnDataAppendState &state, DataChunk &input) override; |
92 | |
93 | //! Reverse lookup map to reconstruct keys from a partition id |
94 | std::map<idx_t, const HivePartitionKey *> GetReverseMap(); |
95 | |
96 | protected: |
97 | //! Create allocators for all currently registered partitions |
98 | void GrowAllocators(); |
99 | //! Create append states for all currently registered partitions |
100 | void GrowAppendState(PartitionedColumnDataAppendState &state); |
101 | //! Create and initialize partitions for all currently registered partitions |
102 | void GrowPartitions(PartitionedColumnDataAppendState &state); |
103 | //! Register a newly discovered partition |
104 | idx_t RegisterNewPartition(HivePartitionKey key, PartitionedColumnDataAppendState &state); |
105 | //! Copy the newly added entries in the global_state.map to the local_partition_map (requires lock!) |
106 | void SynchronizeLocalMap(); |
107 | |
108 | private: |
109 | void InitializeKeys(); |
110 | |
111 | protected: |
112 | //! Shared HivePartitionedColumnData should always have a global state to allow parallel key discovery |
113 | shared_ptr<GlobalHivePartitionState> global_state; |
114 | //! Thread-local copy of the partition map |
115 | hive_partition_map_t local_partition_map; |
116 | //! The columns that make up the key |
117 | vector<idx_t> group_by_columns; |
118 | //! Thread-local pre-allocated vector for hashes |
119 | Vector hashes_v; |
120 | //! Thread-local pre-allocated HivePartitionKeys |
121 | vector<HivePartitionKey> keys; |
122 | }; |
123 | |
124 | } // namespace duckdb |
125 | |