1 | // Licensed to the Apache Software Foundation (ASF) under one |
2 | // or more contributor license agreements. See the NOTICE file |
3 | // distributed with this work for additional information |
4 | // regarding copyright ownership. The ASF licenses this file |
5 | // to you under the Apache License, Version 2.0 (the |
6 | // "License"); you may not use this file except in compliance |
7 | // with the License. You may obtain a copy of the License at |
8 | // |
9 | // http://www.apache.org/licenses/LICENSE-2.0 |
10 | // |
11 | // Unless required by applicable law or agreed to in writing, |
12 | // software distributed under the License is distributed on an |
13 | // "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
14 | // KIND, either express or implied. See the License for the |
15 | // specific language governing permissions and limitations |
16 | // under the License. |
17 | |
18 | // This module contains the logical parquet-cpp types (independent of Thrift |
19 | // structures), schema nodes, and related type tools |
20 | |
21 | #ifndef PARQUET_SCHEMA_TYPES_H |
22 | #define PARQUET_SCHEMA_TYPES_H |
23 | |
24 | #include <cstdint> |
25 | #include <memory> |
26 | #include <ostream> |
27 | #include <string> |
28 | #include <unordered_map> |
29 | #include <vector> |
30 | |
31 | #include "parquet/types.h" |
32 | #include "parquet/util/macros.h" |
33 | #include "parquet/util/visibility.h" |
34 | |
35 | namespace parquet { |
36 | |
37 | class SchemaDescriptor; |
38 | |
39 | namespace schema { |
40 | |
41 | class Node; |
42 | |
43 | // List encodings: using the terminology from Impala to define different styles |
44 | // of representing logical lists (a.k.a. ARRAY types) in Parquet schemas. Since |
45 | // the converted type named in the Parquet metadata is ConvertedType::LIST we |
46 | // use that terminology here. It also helps distinguish from the *_ARRAY |
47 | // primitive types. |
48 | // |
49 | // One-level encoding: Only allows required lists with required cells |
50 | // repeated value_type name |
51 | // |
52 | // Two-level encoding: Enables optional lists with only required cells |
53 | // <required/optional> group list |
54 | // repeated value_type item |
55 | // |
56 | // Three-level encoding: Enables optional lists with optional cells |
57 | // <required/optional> group bag |
58 | // repeated group list |
59 | // <required/optional> value_type item |
60 | // |
61 | // 2- and 1-level encoding are respectively equivalent to 3-level encoding with |
62 | // the non-repeated nodes set to required. |
63 | // |
64 | // The "official" encoding recommended in the Parquet spec is the 3-level, and |
65 | // we use that as the default when creating list types. For semantic completeness |
66 | // we allow the other two. Since all types of encodings will occur "in the |
67 | // wild" we need to be able to interpret the associated definition levels in |
68 | // the context of the actual encoding used in the file. |
69 | // |
70 | // NB: Some Parquet writers may not set ConvertedType::LIST on the repeated |
71 | // SchemaElement, which could make things challenging if we are trying to infer |
72 | // that a sequence of nodes semantically represents an array according to one |
73 | // of these encodings (versus a struct containing an array). We should refuse |
74 | // the temptation to guess, as they say. |
75 | struct ListEncoding { |
76 | enum type { ONE_LEVEL, TWO_LEVEL, THREE_LEVEL }; |
77 | }; |
78 | |
79 | struct DecimalMetadata { |
80 | bool isset; |
81 | int32_t scale; |
82 | int32_t precision; |
83 | }; |
84 | |
85 | class PARQUET_EXPORT ColumnPath { |
86 | public: |
87 | ColumnPath() : path_() {} |
88 | explicit ColumnPath(const std::vector<std::string>& path) : path_(path) {} |
89 | explicit ColumnPath(std::vector<std::string>&& path) : path_(path) {} |
90 | |
91 | static std::shared_ptr<ColumnPath> FromDotString(const std::string& dotstring); |
92 | static std::shared_ptr<ColumnPath> FromNode(const Node& node); |
93 | |
94 | std::shared_ptr<ColumnPath> extend(const std::string& node_name) const; |
95 | std::string ToDotString() const; |
96 | const std::vector<std::string>& ToDotVector() const; |
97 | |
98 | protected: |
99 | std::vector<std::string> path_; |
100 | }; |
101 | |
102 | class GroupNode; |
103 | |
104 | // Base class for logical schema types. A type has a name, repetition level, |
105 | // and optionally a logical type (ConvertedType in Parquet metadata parlance) |
106 | class PARQUET_EXPORT Node { |
107 | public: |
108 | enum type { PRIMITIVE, GROUP }; |
109 | |
110 | Node(Node::type type, const std::string& name, Repetition::type repetition, |
111 | LogicalType::type logical_type = LogicalType::NONE, int id = -1) |
112 | : type_(type), |
113 | name_(name), |
114 | repetition_(repetition), |
115 | logical_type_(logical_type), |
116 | id_(id), |
117 | parent_(NULLPTR) {} |
118 | |
119 | virtual ~Node() {} |
120 | |
121 | bool is_primitive() const { return type_ == Node::PRIMITIVE; } |
122 | |
123 | bool is_group() const { return type_ == Node::GROUP; } |
124 | |
125 | bool is_optional() const { return repetition_ == Repetition::OPTIONAL; } |
126 | |
127 | bool is_repeated() const { return repetition_ == Repetition::REPEATED; } |
128 | |
129 | bool is_required() const { return repetition_ == Repetition::REQUIRED; } |
130 | |
131 | virtual bool Equals(const Node* other) const = 0; |
132 | |
133 | const std::string& name() const { return name_; } |
134 | |
135 | Node::type node_type() const { return type_; } |
136 | |
137 | Repetition::type repetition() const { return repetition_; } |
138 | |
139 | LogicalType::type logical_type() const { return logical_type_; } |
140 | |
141 | int id() const { return id_; } |
142 | |
143 | const Node* parent() const { return parent_; } |
144 | |
145 | const std::shared_ptr<ColumnPath> path() const; |
146 | |
147 | // ToParquet returns an opaque void* to avoid exporting |
148 | // parquet::SchemaElement into the public API |
149 | virtual void ToParquet(void* opaque_element) const = 0; |
150 | |
151 | // Node::Visitor abstract class for walking schemas with the visitor pattern |
152 | class Visitor { |
153 | public: |
154 | virtual ~Visitor() {} |
155 | |
156 | virtual void Visit(Node* node) = 0; |
157 | }; |
158 | class ConstVisitor { |
159 | public: |
160 | virtual ~ConstVisitor() {} |
161 | |
162 | virtual void Visit(const Node* node) = 0; |
163 | }; |
164 | |
165 | virtual void Visit(Visitor* visitor) = 0; |
166 | virtual void VisitConst(ConstVisitor* visitor) const = 0; |
167 | |
168 | protected: |
169 | friend class GroupNode; |
170 | |
171 | Node::type type_; |
172 | std::string name_; |
173 | Repetition::type repetition_; |
174 | LogicalType::type logical_type_; |
175 | int id_; |
176 | // Nodes should not be shared, they have a single parent. |
177 | const Node* parent_; |
178 | |
179 | bool EqualsInternal(const Node* other) const; |
180 | void SetParent(const Node* p_parent); |
181 | |
182 | private: |
183 | PARQUET_DISALLOW_COPY_AND_ASSIGN(Node); |
184 | }; |
185 | |
186 | // Save our breath all over the place with these typedefs |
187 | typedef std::shared_ptr<Node> NodePtr; |
188 | typedef std::vector<NodePtr> NodeVector; |
189 | |
190 | // A type that is one of the primitive Parquet storage types. In addition to |
191 | // the other type metadata (name, repetition level, logical type), also has the |
192 | // physical storage type and their type-specific metadata (byte width, decimal |
193 | // parameters) |
194 | class PARQUET_EXPORT PrimitiveNode : public Node { |
195 | public: |
196 | // FromParquet accepts an opaque void* to avoid exporting |
197 | // parquet::SchemaElement into the public API |
198 | static std::unique_ptr<Node> FromParquet(const void* opaque_element, int id); |
199 | |
200 | static inline NodePtr Make(const std::string& name, Repetition::type repetition, |
201 | Type::type type, |
202 | LogicalType::type logical_type = LogicalType::NONE, |
203 | int length = -1, int precision = -1, int scale = -1) { |
204 | return NodePtr(new PrimitiveNode(name, repetition, type, logical_type, length, |
205 | precision, scale)); |
206 | } |
207 | |
208 | bool Equals(const Node* other) const override; |
209 | |
210 | Type::type physical_type() const { return physical_type_; } |
211 | |
212 | ColumnOrder column_order() const { return column_order_; } |
213 | |
214 | void SetColumnOrder(ColumnOrder column_order) { column_order_ = column_order; } |
215 | |
216 | int32_t type_length() const { return type_length_; } |
217 | |
218 | const DecimalMetadata& decimal_metadata() const { return decimal_metadata_; } |
219 | |
220 | void ToParquet(void* opaque_element) const override; |
221 | void Visit(Visitor* visitor) override; |
222 | void VisitConst(ConstVisitor* visitor) const override; |
223 | |
224 | private: |
225 | PrimitiveNode(const std::string& name, Repetition::type repetition, Type::type type, |
226 | LogicalType::type logical_type = LogicalType::NONE, int length = -1, |
227 | int precision = -1, int scale = -1, int id = -1); |
228 | |
229 | Type::type physical_type_; |
230 | int32_t type_length_; |
231 | DecimalMetadata decimal_metadata_; |
232 | ColumnOrder column_order_; |
233 | |
234 | // For FIXED_LEN_BYTE_ARRAY |
235 | void SetTypeLength(int32_t length) { type_length_ = length; } |
236 | |
237 | // For Decimal logical type: Precision and scale |
238 | void SetDecimalMetadata(int32_t scale, int32_t precision) { |
239 | decimal_metadata_.scale = scale; |
240 | decimal_metadata_.precision = precision; |
241 | } |
242 | |
243 | bool EqualsInternal(const PrimitiveNode* other) const; |
244 | |
245 | FRIEND_TEST(TestPrimitiveNode, Attrs); |
246 | FRIEND_TEST(TestPrimitiveNode, Equals); |
247 | FRIEND_TEST(TestPrimitiveNode, PhysicalLogicalMapping); |
248 | FRIEND_TEST(TestPrimitiveNode, FromParquet); |
249 | }; |
250 | |
251 | class PARQUET_EXPORT GroupNode : public Node { |
252 | public: |
253 | // Like PrimitiveNode, GroupNode::FromParquet accepts an opaque void* to avoid exporting |
254 | // parquet::SchemaElement into the public API |
255 | static std::unique_ptr<Node> FromParquet(const void* opaque_element, int id, |
256 | const NodeVector& fields); |
257 | |
258 | static inline NodePtr Make(const std::string& name, Repetition::type repetition, |
259 | const NodeVector& fields, |
260 | LogicalType::type logical_type = LogicalType::NONE) { |
261 | return NodePtr(new GroupNode(name, repetition, fields, logical_type)); |
262 | } |
263 | |
264 | bool Equals(const Node* other) const override; |
265 | |
266 | NodePtr field(int i) const { return fields_[i]; } |
267 | // Get the index of a field by its name, or negative value if not found. |
268 | // If several fields share the same name, it is unspecified which one |
269 | // is returned. |
270 | int FieldIndex(const std::string& name) const; |
271 | // Get the index of a field by its node, or negative value if not found. |
272 | int FieldIndex(const Node& node) const; |
273 | |
274 | int field_count() const { return static_cast<int>(fields_.size()); } |
275 | |
276 | void ToParquet(void* opaque_element) const override; |
277 | void Visit(Visitor* visitor) override; |
278 | void VisitConst(ConstVisitor* visitor) const override; |
279 | |
280 | private: |
281 | GroupNode(const std::string& name, Repetition::type repetition, |
282 | const NodeVector& fields, LogicalType::type logical_type = LogicalType::NONE, |
283 | int id = -1) |
284 | : Node(Node::GROUP, name, repetition, logical_type, id), fields_(fields) { |
285 | field_name_to_idx_.clear(); |
286 | auto field_idx = 0; |
287 | for (NodePtr& field : fields_) { |
288 | field->SetParent(this); |
289 | field_name_to_idx_.emplace(field->name(), field_idx++); |
290 | } |
291 | } |
292 | |
293 | NodeVector fields_; |
294 | bool EqualsInternal(const GroupNode* other) const; |
295 | |
296 | // Mapping between field name to the field index |
297 | std::unordered_multimap<std::string, int> field_name_to_idx_; |
298 | |
299 | FRIEND_TEST(TestGroupNode, Attrs); |
300 | FRIEND_TEST(TestGroupNode, Equals); |
301 | FRIEND_TEST(TestGroupNode, FieldIndex); |
302 | FRIEND_TEST(TestGroupNode, FieldIndexDuplicateName); |
303 | }; |
304 | |
305 | // ---------------------------------------------------------------------- |
306 | // Convenience primitive type factory functions |
307 | |
308 | #define PRIMITIVE_FACTORY(FuncName, TYPE) \ |
309 | static inline NodePtr FuncName(const std::string& name, \ |
310 | Repetition::type repetition = Repetition::OPTIONAL) { \ |
311 | return PrimitiveNode::Make(name, repetition, Type::TYPE); \ |
312 | } |
313 | |
314 | PRIMITIVE_FACTORY(Boolean, BOOLEAN); |
315 | PRIMITIVE_FACTORY(Int32, INT32); |
316 | PRIMITIVE_FACTORY(Int64, INT64); |
317 | PRIMITIVE_FACTORY(Int96, INT96); |
318 | PRIMITIVE_FACTORY(Float, FLOAT); |
319 | PRIMITIVE_FACTORY(Double, DOUBLE); |
320 | PRIMITIVE_FACTORY(ByteArray, BYTE_ARRAY); |
321 | |
322 | void PARQUET_EXPORT PrintSchema(const schema::Node* schema, std::ostream& stream, |
323 | int indent_width = 2); |
324 | |
325 | } // namespace schema |
326 | |
327 | // The ColumnDescriptor encapsulates information necessary to interpret |
328 | // primitive column data in the context of a particular schema. We have to |
329 | // examine the node structure of a column's path to the root in the schema tree |
330 | // to be able to reassemble the nested structure from the repetition and |
331 | // definition levels. |
332 | class PARQUET_EXPORT ColumnDescriptor { |
333 | public: |
334 | ColumnDescriptor(const schema::NodePtr& node, int16_t max_definition_level, |
335 | int16_t max_repetition_level, |
336 | const SchemaDescriptor* schema_descr = NULLPTR); |
337 | |
338 | bool Equals(const ColumnDescriptor& other) const; |
339 | |
340 | int16_t max_definition_level() const { return max_definition_level_; } |
341 | |
342 | int16_t max_repetition_level() const { return max_repetition_level_; } |
343 | |
344 | Type::type physical_type() const { return primitive_node_->physical_type(); } |
345 | |
346 | LogicalType::type logical_type() const { return primitive_node_->logical_type(); } |
347 | |
348 | ColumnOrder column_order() const { return primitive_node_->column_order(); } |
349 | |
350 | SortOrder::type sort_order() const { |
351 | return GetSortOrder(logical_type(), physical_type()); |
352 | } |
353 | |
354 | const std::string& name() const { return primitive_node_->name(); } |
355 | |
356 | const std::shared_ptr<schema::ColumnPath> path() const; |
357 | |
358 | const schema::NodePtr& schema_node() const { return node_; } |
359 | |
360 | int type_length() const; |
361 | |
362 | int type_precision() const; |
363 | |
364 | int type_scale() const; |
365 | |
366 | private: |
367 | schema::NodePtr node_; |
368 | const schema::PrimitiveNode* primitive_node_; |
369 | |
370 | int16_t max_definition_level_; |
371 | int16_t max_repetition_level_; |
372 | }; |
373 | |
374 | // Container for the converted Parquet schema with a computed information from |
375 | // the schema analysis needed for file reading |
376 | // |
377 | // * Column index to Node |
378 | // * Max repetition / definition levels for each primitive node |
379 | // |
380 | // The ColumnDescriptor objects produced by this class can be used to assist in |
381 | // the reconstruction of fully materialized data structures from the |
382 | // repetition-definition level encoding of nested data |
383 | // |
384 | // TODO(wesm): this object can be recomputed from a Schema |
385 | class PARQUET_EXPORT SchemaDescriptor { |
386 | public: |
387 | SchemaDescriptor() {} |
388 | ~SchemaDescriptor() {} |
389 | |
390 | // Analyze the schema |
391 | void Init(std::unique_ptr<schema::Node> schema); |
392 | void Init(const schema::NodePtr& schema); |
393 | |
394 | const ColumnDescriptor* Column(int i) const; |
395 | |
396 | // Get the index of a column by its dotstring path, or negative value if not found. |
397 | // If several columns share the same dotstring path, it is unspecified which one |
398 | // is returned. |
399 | int ColumnIndex(const std::string& node_path) const; |
400 | // Get the index of a column by its node, or negative value if not found. |
401 | int ColumnIndex(const schema::Node& node) const; |
402 | |
403 | bool Equals(const SchemaDescriptor& other) const; |
404 | |
405 | // The number of physical columns appearing in the file |
406 | int num_columns() const { return static_cast<int>(leaves_.size()); } |
407 | |
408 | const schema::NodePtr& schema_root() const { return schema_; } |
409 | |
410 | const schema::GroupNode* group_node() const { return group_node_; } |
411 | |
412 | // Returns the root (child of the schema root) node of the leaf(column) node |
413 | const schema::Node* GetColumnRoot(int i) const; |
414 | |
415 | const std::string& name() const { return group_node_->name(); } |
416 | |
417 | std::string ToString() const; |
418 | |
419 | void updateColumnOrders(const std::vector<ColumnOrder>& column_orders); |
420 | |
421 | private: |
422 | friend class ColumnDescriptor; |
423 | |
424 | // Root Node |
425 | schema::NodePtr schema_; |
426 | // Root Node |
427 | const schema::GroupNode* group_node_; |
428 | |
429 | void BuildTree(const schema::NodePtr& node, int16_t max_def_level, |
430 | int16_t max_rep_level, const schema::NodePtr& base); |
431 | |
432 | // Result of leaf node / tree analysis |
433 | std::vector<ColumnDescriptor> leaves_; |
434 | |
435 | // Mapping between leaf nodes and root group of leaf (first node |
436 | // below the schema's root group) |
437 | // |
438 | // For example, the leaf `a.b.c.d` would have a link back to `a` |
439 | // |
440 | // -- a <------ |
441 | // -- -- b | |
442 | // -- -- -- c | |
443 | // -- -- -- -- d |
444 | std::unordered_map<int, const schema::NodePtr> leaf_to_base_; |
445 | |
446 | // Mapping between ColumnPath DotString to the leaf index |
447 | std::unordered_multimap<std::string, int> leaf_to_idx_; |
448 | }; |
449 | |
450 | } // namespace parquet |
451 | |
452 | #endif // PARQUET_SCHEMA_TYPES_H |
453 | |