-
Notifications
You must be signed in to change notification settings - Fork 28k
/
ParquetFilters.scala
378 lines (339 loc) · 18.1 KB
/
ParquetFilters.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
/*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You under the Apache License, Version 2.0
* (the "License"); you may not use this file except in compliance with
* the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.spark.sql.execution.datasources.parquet
import java.lang.{Long => JLong}
import java.sql.{Date, Timestamp}
import scala.collection.JavaConverters.asScalaBufferConverter
import org.apache.parquet.filter2.predicate._
import org.apache.parquet.filter2.predicate.FilterApi._
import org.apache.parquet.io.api.Binary
import org.apache.parquet.schema.{DecimalMetadata, MessageType, OriginalType, PrimitiveComparator}
import org.apache.parquet.schema.OriginalType._
import org.apache.parquet.schema.PrimitiveType.PrimitiveTypeName
import org.apache.parquet.schema.PrimitiveType.PrimitiveTypeName._
import org.apache.spark.sql.catalyst.util.DateTimeUtils
import org.apache.spark.sql.catalyst.util.DateTimeUtils.SQLDate
import org.apache.spark.sql.sources
import org.apache.spark.unsafe.types.UTF8String
/**
* Some utility function to convert Spark data source filters to Parquet filters.
*/
private[parquet] class ParquetFilters(
pushDownDate: Boolean,
pushDownTimestamp: Boolean,
pushDownStartWith: Boolean) {
private case class ParquetSchemaType(
originalType: OriginalType,
primitiveTypeName: PrimitiveTypeName,
decimalMetadata: DecimalMetadata)
private val ParquetBooleanType = ParquetSchemaType(null, BOOLEAN, null)
private val ParquetIntegerType = ParquetSchemaType(null, INT32, null)
private val ParquetLongType = ParquetSchemaType(null, INT64, null)
private val ParquetFloatType = ParquetSchemaType(null, FLOAT, null)
private val ParquetDoubleType = ParquetSchemaType(null, DOUBLE, null)
private val ParquetStringType = ParquetSchemaType(UTF8, BINARY, null)
private val ParquetBinaryType = ParquetSchemaType(null, BINARY, null)
private val ParquetDateType = ParquetSchemaType(DATE, INT32, null)
private val ParquetTimestampMicrosType = ParquetSchemaType(TIMESTAMP_MICROS, INT64, null)
private val ParquetTimestampMillisType = ParquetSchemaType(TIMESTAMP_MILLIS, INT64, null)
private def dateToDays(date: Date): SQLDate = {
DateTimeUtils.fromJavaDate(date)
}
private val makeEq: PartialFunction[ParquetSchemaType, (String, Any) => FilterPredicate] = {
case ParquetBooleanType =>
(n: String, v: Any) => FilterApi.eq(booleanColumn(n), v.asInstanceOf[java.lang.Boolean])
case ParquetIntegerType =>
(n: String, v: Any) => FilterApi.eq(intColumn(n), v.asInstanceOf[Integer])
case ParquetLongType =>
(n: String, v: Any) => FilterApi.eq(longColumn(n), v.asInstanceOf[java.lang.Long])
case ParquetFloatType =>
(n: String, v: Any) => FilterApi.eq(floatColumn(n), v.asInstanceOf[java.lang.Float])
case ParquetDoubleType =>
(n: String, v: Any) => FilterApi.eq(doubleColumn(n), v.asInstanceOf[java.lang.Double])
// Binary.fromString and Binary.fromByteArray don't accept null values
case ParquetStringType =>
(n: String, v: Any) => FilterApi.eq(
binaryColumn(n),
Option(v).map(s => Binary.fromString(s.asInstanceOf[String])).orNull)
case ParquetBinaryType =>
(n: String, v: Any) => FilterApi.eq(
binaryColumn(n),
Option(v).map(b => Binary.fromReusedByteArray(v.asInstanceOf[Array[Byte]])).orNull)
case ParquetDateType if pushDownDate =>
(n: String, v: Any) => FilterApi.eq(
intColumn(n),
Option(v).map(date => dateToDays(date.asInstanceOf[Date]).asInstanceOf[Integer]).orNull)
case ParquetTimestampMicrosType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.eq(
longColumn(n),
Option(v).map(t => DateTimeUtils.fromJavaTimestamp(t.asInstanceOf[Timestamp])
.asInstanceOf[JLong]).orNull)
case ParquetTimestampMillisType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.eq(
longColumn(n),
Option(v).map(_.asInstanceOf[Timestamp].getTime.asInstanceOf[JLong]).orNull)
}
private val makeNotEq: PartialFunction[ParquetSchemaType, (String, Any) => FilterPredicate] = {
case ParquetBooleanType =>
(n: String, v: Any) => FilterApi.notEq(booleanColumn(n), v.asInstanceOf[java.lang.Boolean])
case ParquetIntegerType =>
(n: String, v: Any) => FilterApi.notEq(intColumn(n), v.asInstanceOf[Integer])
case ParquetLongType =>
(n: String, v: Any) => FilterApi.notEq(longColumn(n), v.asInstanceOf[java.lang.Long])
case ParquetFloatType =>
(n: String, v: Any) => FilterApi.notEq(floatColumn(n), v.asInstanceOf[java.lang.Float])
case ParquetDoubleType =>
(n: String, v: Any) => FilterApi.notEq(doubleColumn(n), v.asInstanceOf[java.lang.Double])
case ParquetStringType =>
(n: String, v: Any) => FilterApi.notEq(
binaryColumn(n),
Option(v).map(s => Binary.fromString(s.asInstanceOf[String])).orNull)
case ParquetBinaryType =>
(n: String, v: Any) => FilterApi.notEq(
binaryColumn(n),
Option(v).map(b => Binary.fromReusedByteArray(v.asInstanceOf[Array[Byte]])).orNull)
case ParquetDateType if pushDownDate =>
(n: String, v: Any) => FilterApi.notEq(
intColumn(n),
Option(v).map(date => dateToDays(date.asInstanceOf[Date]).asInstanceOf[Integer]).orNull)
case ParquetTimestampMicrosType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.notEq(
longColumn(n),
Option(v).map(t => DateTimeUtils.fromJavaTimestamp(t.asInstanceOf[Timestamp])
.asInstanceOf[JLong]).orNull)
case ParquetTimestampMillisType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.notEq(
longColumn(n),
Option(v).map(_.asInstanceOf[Timestamp].getTime.asInstanceOf[JLong]).orNull)
}
private val makeLt: PartialFunction[ParquetSchemaType, (String, Any) => FilterPredicate] = {
case ParquetIntegerType =>
(n: String, v: Any) => FilterApi.lt(intColumn(n), v.asInstanceOf[Integer])
case ParquetLongType =>
(n: String, v: Any) => FilterApi.lt(longColumn(n), v.asInstanceOf[java.lang.Long])
case ParquetFloatType =>
(n: String, v: Any) => FilterApi.lt(floatColumn(n), v.asInstanceOf[java.lang.Float])
case ParquetDoubleType =>
(n: String, v: Any) => FilterApi.lt(doubleColumn(n), v.asInstanceOf[java.lang.Double])
case ParquetStringType =>
(n: String, v: Any) =>
FilterApi.lt(binaryColumn(n), Binary.fromString(v.asInstanceOf[String]))
case ParquetBinaryType =>
(n: String, v: Any) =>
FilterApi.lt(binaryColumn(n), Binary.fromReusedByteArray(v.asInstanceOf[Array[Byte]]))
case ParquetDateType if pushDownDate =>
(n: String, v: Any) =>
FilterApi.lt(intColumn(n), dateToDays(v.asInstanceOf[Date]).asInstanceOf[Integer])
case ParquetTimestampMicrosType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.lt(
longColumn(n),
DateTimeUtils.fromJavaTimestamp(v.asInstanceOf[Timestamp]).asInstanceOf[JLong])
case ParquetTimestampMillisType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.lt(
longColumn(n),
v.asInstanceOf[Timestamp].getTime.asInstanceOf[JLong])
}
private val makeLtEq: PartialFunction[ParquetSchemaType, (String, Any) => FilterPredicate] = {
case ParquetIntegerType =>
(n: String, v: Any) => FilterApi.ltEq(intColumn(n), v.asInstanceOf[Integer])
case ParquetLongType =>
(n: String, v: Any) => FilterApi.ltEq(longColumn(n), v.asInstanceOf[java.lang.Long])
case ParquetFloatType =>
(n: String, v: Any) => FilterApi.ltEq(floatColumn(n), v.asInstanceOf[java.lang.Float])
case ParquetDoubleType =>
(n: String, v: Any) => FilterApi.ltEq(doubleColumn(n), v.asInstanceOf[java.lang.Double])
case ParquetStringType =>
(n: String, v: Any) =>
FilterApi.ltEq(binaryColumn(n), Binary.fromString(v.asInstanceOf[String]))
case ParquetBinaryType =>
(n: String, v: Any) =>
FilterApi.ltEq(binaryColumn(n), Binary.fromReusedByteArray(v.asInstanceOf[Array[Byte]]))
case ParquetDateType if pushDownDate =>
(n: String, v: Any) =>
FilterApi.ltEq(intColumn(n), dateToDays(v.asInstanceOf[Date]).asInstanceOf[Integer])
case ParquetTimestampMicrosType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.ltEq(
longColumn(n),
DateTimeUtils.fromJavaTimestamp(v.asInstanceOf[Timestamp]).asInstanceOf[JLong])
case ParquetTimestampMillisType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.ltEq(
longColumn(n),
v.asInstanceOf[Timestamp].getTime.asInstanceOf[JLong])
}
private val makeGt: PartialFunction[ParquetSchemaType, (String, Any) => FilterPredicate] = {
case ParquetIntegerType =>
(n: String, v: Any) => FilterApi.gt(intColumn(n), v.asInstanceOf[Integer])
case ParquetLongType =>
(n: String, v: Any) => FilterApi.gt(longColumn(n), v.asInstanceOf[java.lang.Long])
case ParquetFloatType =>
(n: String, v: Any) => FilterApi.gt(floatColumn(n), v.asInstanceOf[java.lang.Float])
case ParquetDoubleType =>
(n: String, v: Any) => FilterApi.gt(doubleColumn(n), v.asInstanceOf[java.lang.Double])
case ParquetStringType =>
(n: String, v: Any) =>
FilterApi.gt(binaryColumn(n), Binary.fromString(v.asInstanceOf[String]))
case ParquetBinaryType =>
(n: String, v: Any) =>
FilterApi.gt(binaryColumn(n), Binary.fromReusedByteArray(v.asInstanceOf[Array[Byte]]))
case ParquetDateType if pushDownDate =>
(n: String, v: Any) =>
FilterApi.gt(intColumn(n), dateToDays(v.asInstanceOf[Date]).asInstanceOf[Integer])
case ParquetTimestampMicrosType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.gt(
longColumn(n),
DateTimeUtils.fromJavaTimestamp(v.asInstanceOf[Timestamp]).asInstanceOf[JLong])
case ParquetTimestampMillisType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.gt(
longColumn(n),
v.asInstanceOf[Timestamp].getTime.asInstanceOf[JLong])
}
private val makeGtEq: PartialFunction[ParquetSchemaType, (String, Any) => FilterPredicate] = {
case ParquetIntegerType =>
(n: String, v: Any) => FilterApi.gtEq(intColumn(n), v.asInstanceOf[Integer])
case ParquetLongType =>
(n: String, v: Any) => FilterApi.gtEq(longColumn(n), v.asInstanceOf[java.lang.Long])
case ParquetFloatType =>
(n: String, v: Any) => FilterApi.gtEq(floatColumn(n), v.asInstanceOf[java.lang.Float])
case ParquetDoubleType =>
(n: String, v: Any) => FilterApi.gtEq(doubleColumn(n), v.asInstanceOf[java.lang.Double])
case ParquetStringType =>
(n: String, v: Any) =>
FilterApi.gtEq(binaryColumn(n), Binary.fromString(v.asInstanceOf[String]))
case ParquetBinaryType =>
(n: String, v: Any) =>
FilterApi.gtEq(binaryColumn(n), Binary.fromReusedByteArray(v.asInstanceOf[Array[Byte]]))
case ParquetDateType if pushDownDate =>
(n: String, v: Any) =>
FilterApi.gtEq(intColumn(n), dateToDays(v.asInstanceOf[Date]).asInstanceOf[Integer])
case ParquetTimestampMicrosType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.gtEq(
longColumn(n),
DateTimeUtils.fromJavaTimestamp(v.asInstanceOf[Timestamp]).asInstanceOf[JLong])
case ParquetTimestampMillisType if pushDownTimestamp =>
(n: String, v: Any) => FilterApi.gtEq(
longColumn(n),
v.asInstanceOf[Timestamp].getTime.asInstanceOf[JLong])
}
/**
* Returns a map from name of the column to the data type, if predicate push down applies.
*/
private def getFieldMap(dataType: MessageType): Map[String, ParquetSchemaType] = dataType match {
case m: MessageType =>
// Here we don't flatten the fields in the nested schema but just look up through
// root fields. Currently, accessing to nested fields does not push down filters
// and it does not support to create filters for them.
m.getFields.asScala.filter(_.isPrimitive).map(_.asPrimitiveType()).map { f =>
f.getName -> ParquetSchemaType(
f.getOriginalType, f.getPrimitiveTypeName, f.getDecimalMetadata)
}.toMap
case _ => Map.empty[String, ParquetSchemaType]
}
/**
* Converts data sources filters to Parquet filter predicates.
*/
def createFilter(schema: MessageType, predicate: sources.Filter): Option[FilterPredicate] = {
val nameToType = getFieldMap(schema)
// Parquet does not allow dots in the column name because dots are used as a column path
// delimiter. Since Parquet 1.8.2 (PARQUET-389), Parquet accepts the filter predicates
// with missing columns. The incorrect results could be got from Parquet when we push down
// filters for the column having dots in the names. Thus, we do not push down such filters.
// See SPARK-20364.
def canMakeFilterOn(name: String): Boolean = nameToType.contains(name) && !name.contains(".")
// NOTE:
//
// For any comparison operator `cmp`, both `a cmp NULL` and `NULL cmp a` evaluate to `NULL`,
// which can be casted to `false` implicitly. Please refer to the `eval` method of these
// operators and the `PruneFilters` rule for details.
// Hyukjin:
// I added [[EqualNullSafe]] with [[org.apache.parquet.filter2.predicate.Operators.Eq]].
// So, it performs equality comparison identically when given [[sources.Filter]] is [[EqualTo]].
// The reason why I did this is, that the actual Parquet filter checks null-safe equality
// comparison.
// So I added this and maybe [[EqualTo]] should be changed. It still seems fine though, because
// physical planning does not set `NULL` to [[EqualTo]] but changes it to [[IsNull]] and etc.
// Probably I missed something and obviously this should be changed.
predicate match {
case sources.IsNull(name) if canMakeFilterOn(name) =>
makeEq.lift(nameToType(name)).map(_(name, null))
case sources.IsNotNull(name) if canMakeFilterOn(name) =>
makeNotEq.lift(nameToType(name)).map(_(name, null))
case sources.EqualTo(name, value) if canMakeFilterOn(name) =>
makeEq.lift(nameToType(name)).map(_(name, value))
case sources.Not(sources.EqualTo(name, value)) if canMakeFilterOn(name) =>
makeNotEq.lift(nameToType(name)).map(_(name, value))
case sources.EqualNullSafe(name, value) if canMakeFilterOn(name) =>
makeEq.lift(nameToType(name)).map(_(name, value))
case sources.Not(sources.EqualNullSafe(name, value)) if canMakeFilterOn(name) =>
makeNotEq.lift(nameToType(name)).map(_(name, value))
case sources.LessThan(name, value) if canMakeFilterOn(name) =>
makeLt.lift(nameToType(name)).map(_(name, value))
case sources.LessThanOrEqual(name, value) if canMakeFilterOn(name) =>
makeLtEq.lift(nameToType(name)).map(_(name, value))
case sources.GreaterThan(name, value) if canMakeFilterOn(name) =>
makeGt.lift(nameToType(name)).map(_(name, value))
case sources.GreaterThanOrEqual(name, value) if canMakeFilterOn(name) =>
makeGtEq.lift(nameToType(name)).map(_(name, value))
case sources.And(lhs, rhs) =>
// At here, it is not safe to just convert one side if we do not understand the
// other side. Here is an example used to explain the reason.
// Let's say we have NOT(a = 2 AND b in ('1')) and we do not understand how to
// convert b in ('1'). If we only convert a = 2, we will end up with a filter
// NOT(a = 2), which will generate wrong results.
// Pushing one side of AND down is only safe to do at the top level.
// You can see ParquetRelation's initializeLocalJobFunc method as an example.
for {
lhsFilter <- createFilter(schema, lhs)
rhsFilter <- createFilter(schema, rhs)
} yield FilterApi.and(lhsFilter, rhsFilter)
case sources.Or(lhs, rhs) =>
for {
lhsFilter <- createFilter(schema, lhs)
rhsFilter <- createFilter(schema, rhs)
} yield FilterApi.or(lhsFilter, rhsFilter)
case sources.Not(pred) =>
createFilter(schema, pred).map(FilterApi.not)
case sources.StringStartsWith(name, prefix) if pushDownStartWith && canMakeFilterOn(name) =>
Option(prefix).map { v =>
FilterApi.userDefined(binaryColumn(name),
new UserDefinedPredicate[Binary] with Serializable {
private val strToBinary = Binary.fromReusedByteArray(v.getBytes)
private val size = strToBinary.length
override def canDrop(statistics: Statistics[Binary]): Boolean = {
val comparator = PrimitiveComparator.UNSIGNED_LEXICOGRAPHICAL_BINARY_COMPARATOR
val max = statistics.getMax
val min = statistics.getMin
comparator.compare(max.slice(0, math.min(size, max.length)), strToBinary) < 0 ||
comparator.compare(min.slice(0, math.min(size, min.length)), strToBinary) > 0
}
override def inverseCanDrop(statistics: Statistics[Binary]): Boolean = {
val comparator = PrimitiveComparator.UNSIGNED_LEXICOGRAPHICAL_BINARY_COMPARATOR
val max = statistics.getMax
val min = statistics.getMin
comparator.compare(max.slice(0, math.min(size, max.length)), strToBinary) == 0 &&
comparator.compare(min.slice(0, math.min(size, min.length)), strToBinary) == 0
}
override def keep(value: Binary): Boolean = {
UTF8String.fromBytes(value.getBytes).startsWith(
UTF8String.fromBytes(strToBinary.getBytes))
}
}
)
}
case _ => None
}
}
}