-
Notifications
You must be signed in to change notification settings - Fork 28k
/
KafkaMicroBatchReader.scala
402 lines (352 loc) · 15.8 KB
/
KafkaMicroBatchReader.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
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
/*
* 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.kafka010
import java.{util => ju}
import java.io._
import java.nio.charset.StandardCharsets
import scala.collection.JavaConverters._
import org.apache.commons.io.IOUtils
import org.apache.spark.SparkEnv
import org.apache.spark.internal.Logging
import org.apache.spark.scheduler.ExecutorCacheTaskLocation
import org.apache.spark.sql.SparkSession
import org.apache.spark.sql.catalyst.InternalRow
import org.apache.spark.sql.catalyst.expressions.UnsafeRow
import org.apache.spark.sql.execution.streaming.{HDFSMetadataLog, SerializedOffset}
import org.apache.spark.sql.kafka010.KafkaSourceProvider.{INSTRUCTION_FOR_FAIL_ON_DATA_LOSS_FALSE, INSTRUCTION_FOR_FAIL_ON_DATA_LOSS_TRUE}
import org.apache.spark.sql.sources.v2.DataSourceOptions
import org.apache.spark.sql.sources.v2.reader.{InputPartition, InputPartitionReader}
import org.apache.spark.sql.sources.v2.reader.streaming.{MicroBatchReader, Offset}
import org.apache.spark.sql.types.StructType
import org.apache.spark.util.UninterruptibleThread
/**
* A [[MicroBatchReader]] that reads data from Kafka.
*
* The [[KafkaSourceOffset]] is the custom [[Offset]] defined for this source that contains
* a map of TopicPartition -> offset. Note that this offset is 1 + (available offset). For
* example if the last record in a Kafka topic "t", partition 2 is offset 5, then
* KafkaSourceOffset will contain TopicPartition("t", 2) -> 6. This is done keep it consistent
* with the semantics of `KafkaConsumer.position()`.
*
* Zero data lost is not guaranteed when topics are deleted. If zero data lost is critical, the user
* must make sure all messages in a topic have been processed when deleting a topic.
*
* There is a known issue caused by KAFKA-1894: the query using Kafka maybe cannot be stopped.
* To avoid this issue, you should make sure stopping the query before stopping the Kafka brokers
* and not use wrong broker addresses.
*/
private[kafka010] class KafkaMicroBatchReader(
kafkaOffsetReader: KafkaOffsetReader,
executorKafkaParams: ju.Map[String, Object],
options: DataSourceOptions,
metadataPath: String,
startingOffsets: KafkaOffsetRangeLimit,
failOnDataLoss: Boolean)
extends MicroBatchReader with Logging {
private var startPartitionOffsets: PartitionOffsetMap = _
private var endPartitionOffsets: PartitionOffsetMap = _
private val pollTimeoutMs = options.getLong(
"kafkaConsumer.pollTimeoutMs",
SparkEnv.get.conf.getTimeAsSeconds("spark.network.timeout", "120s") * 1000L)
private val maxOffsetsPerTrigger =
Option(options.get("maxOffsetsPerTrigger").orElse(null)).map(_.toLong)
private val rangeCalculator = KafkaOffsetRangeCalculator(options)
/**
* Lazily initialize `initialPartitionOffsets` to make sure that `KafkaConsumer.poll` is only
* called in StreamExecutionThread. Otherwise, interrupting a thread while running
* `KafkaConsumer.poll` may hang forever (KAFKA-1894).
*/
private lazy val initialPartitionOffsets = getOrCreateInitialPartitionOffsets()
override def setOffsetRange(start: ju.Optional[Offset], end: ju.Optional[Offset]): Unit = {
// Make sure initialPartitionOffsets is initialized
initialPartitionOffsets
startPartitionOffsets = Option(start.orElse(null))
.map(_.asInstanceOf[KafkaSourceOffset].partitionToOffsets)
.getOrElse(initialPartitionOffsets)
endPartitionOffsets = Option(end.orElse(null))
.map(_.asInstanceOf[KafkaSourceOffset].partitionToOffsets)
.getOrElse {
val latestPartitionOffsets =
kafkaOffsetReader.fetchLatestOffsets(Some(startPartitionOffsets))
maxOffsetsPerTrigger.map { maxOffsets =>
rateLimit(maxOffsets, startPartitionOffsets, latestPartitionOffsets)
}.getOrElse {
latestPartitionOffsets
}
}
}
override def planInputPartitions(): ju.List[InputPartition[InternalRow]] = {
// Find the new partitions, and get their earliest offsets
val newPartitions = endPartitionOffsets.keySet.diff(startPartitionOffsets.keySet)
val newPartitionInitialOffsets = kafkaOffsetReader.fetchEarliestOffsets(newPartitions.toSeq)
if (newPartitionInitialOffsets.keySet != newPartitions) {
// We cannot get from offsets for some partitions. It means they got deleted.
val deletedPartitions = newPartitions.diff(newPartitionInitialOffsets.keySet)
reportDataLoss(
s"Cannot find earliest offsets of ${deletedPartitions}. Some data may have been missed")
}
logInfo(s"Partitions added: $newPartitionInitialOffsets")
newPartitionInitialOffsets.filter(_._2 != 0).foreach { case (p, o) =>
reportDataLoss(
s"Added partition $p starts from $o instead of 0. Some data may have been missed")
}
// Find deleted partitions, and report data loss if required
val deletedPartitions = startPartitionOffsets.keySet.diff(endPartitionOffsets.keySet)
if (deletedPartitions.nonEmpty) {
reportDataLoss(s"$deletedPartitions are gone. Some data may have been missed")
}
// Use the end partitions to calculate offset ranges to ignore partitions that have
// been deleted
val topicPartitions = endPartitionOffsets.keySet.filter { tp =>
// Ignore partitions that we don't know the from offsets.
newPartitionInitialOffsets.contains(tp) || startPartitionOffsets.contains(tp)
}.toSeq
logDebug("TopicPartitions: " + topicPartitions.mkString(", "))
val fromOffsets = startPartitionOffsets ++ newPartitionInitialOffsets
val untilOffsets = endPartitionOffsets
untilOffsets.foreach { case (tp, untilOffset) =>
fromOffsets.get(tp).foreach { fromOffset =>
if (untilOffset < fromOffset) {
reportDataLoss(s"Partition $tp's offset was changed from " +
s"$fromOffset to $untilOffset, some data may have been missed")
}
}
}
// Calculate offset ranges
val offsetRanges = rangeCalculator.getRanges(
fromOffsets = fromOffsets,
untilOffsets = untilOffsets,
executorLocations = getSortedExecutorList())
// Reuse Kafka consumers only when all the offset ranges have distinct TopicPartitions,
// that is, concurrent tasks will not read the same TopicPartitions.
val reuseKafkaConsumer = offsetRanges.map(_.topicPartition).toSet.size == offsetRanges.size
// Generate factories based on the offset ranges
offsetRanges.map { range =>
new KafkaMicroBatchInputPartition(
range, executorKafkaParams, pollTimeoutMs, failOnDataLoss, reuseKafkaConsumer
): InputPartition[InternalRow]
}.asJava
}
override def getStartOffset: Offset = {
KafkaSourceOffset(startPartitionOffsets)
}
override def getEndOffset: Offset = {
KafkaSourceOffset(endPartitionOffsets)
}
override def deserializeOffset(json: String): Offset = {
KafkaSourceOffset(JsonUtils.partitionOffsets(json))
}
override def readSchema(): StructType = KafkaOffsetReader.kafkaSchema
override def commit(end: Offset): Unit = {}
override def stop(): Unit = {
kafkaOffsetReader.close()
}
override def toString(): String = s"KafkaV2[$kafkaOffsetReader]"
/**
* Read initial partition offsets from the checkpoint, or decide the offsets and write them to
* the checkpoint.
*/
private def getOrCreateInitialPartitionOffsets(): PartitionOffsetMap = {
// Make sure that `KafkaConsumer.poll` is only called in StreamExecutionThread.
// Otherwise, interrupting a thread while running `KafkaConsumer.poll` may hang forever
// (KAFKA-1894).
assert(Thread.currentThread().isInstanceOf[UninterruptibleThread])
// SparkSession is required for getting Hadoop configuration for writing to checkpoints
assert(SparkSession.getActiveSession.nonEmpty)
val metadataLog =
new KafkaSourceInitialOffsetWriter(SparkSession.getActiveSession.get, metadataPath)
metadataLog.get(0).getOrElse {
val offsets = startingOffsets match {
case EarliestOffsetRangeLimit =>
KafkaSourceOffset(kafkaOffsetReader.fetchEarliestOffsets())
case LatestOffsetRangeLimit =>
KafkaSourceOffset(kafkaOffsetReader.fetchLatestOffsets(None))
case SpecificOffsetRangeLimit(p) =>
kafkaOffsetReader.fetchSpecificOffsets(p, reportDataLoss)
}
metadataLog.add(0, offsets)
logInfo(s"Initial offsets: $offsets")
offsets
}.partitionToOffsets
}
/** Proportionally distribute limit number of offsets among topicpartitions */
private def rateLimit(
limit: Long,
from: PartitionOffsetMap,
until: PartitionOffsetMap): PartitionOffsetMap = {
val fromNew = kafkaOffsetReader.fetchEarliestOffsets(until.keySet.diff(from.keySet).toSeq)
val sizes = until.flatMap {
case (tp, end) =>
// If begin isn't defined, something's wrong, but let alert logic in getBatch handle it
from.get(tp).orElse(fromNew.get(tp)).flatMap { begin =>
val size = end - begin
logDebug(s"rateLimit $tp size is $size")
if (size > 0) Some(tp -> size) else None
}
}
val total = sizes.values.sum.toDouble
if (total < 1) {
until
} else {
until.map {
case (tp, end) =>
tp -> sizes.get(tp).map { size =>
val begin = from.get(tp).getOrElse(fromNew(tp))
val prorate = limit * (size / total)
// Don't completely starve small topicpartitions
val prorateLong = (if (prorate < 1) Math.ceil(prorate) else Math.floor(prorate)).toLong
// need to be careful of integer overflow
// therefore added canary checks where to see if off variable could be overflowed
// refer to [https://issues.apache.org/jira/browse/SPARK-26718]
val off = if (prorateLong > Long.MaxValue - begin) {
Long.MaxValue
} else {
begin + prorateLong
}
// Paranoia, make sure not to return an offset that's past end
Math.min(end, off)
}.getOrElse(end)
}
}
}
private def getSortedExecutorList(): Array[String] = {
def compare(a: ExecutorCacheTaskLocation, b: ExecutorCacheTaskLocation): Boolean = {
if (a.host == b.host) {
a.executorId > b.executorId
} else {
a.host > b.host
}
}
val bm = SparkEnv.get.blockManager
bm.master.getPeers(bm.blockManagerId).toArray
.map(x => ExecutorCacheTaskLocation(x.host, x.executorId))
.sortWith(compare)
.map(_.toString)
}
/**
* If `failOnDataLoss` is true, this method will throw an `IllegalStateException`.
* Otherwise, just log a warning.
*/
private def reportDataLoss(message: String): Unit = {
if (failOnDataLoss) {
throw new IllegalStateException(message + s". $INSTRUCTION_FOR_FAIL_ON_DATA_LOSS_TRUE")
} else {
logWarning(message + s". $INSTRUCTION_FOR_FAIL_ON_DATA_LOSS_FALSE")
}
}
/** A version of [[HDFSMetadataLog]] specialized for saving the initial offsets. */
class KafkaSourceInitialOffsetWriter(sparkSession: SparkSession, metadataPath: String)
extends HDFSMetadataLog[KafkaSourceOffset](sparkSession, metadataPath) {
val VERSION = 1
override def serialize(metadata: KafkaSourceOffset, out: OutputStream): Unit = {
out.write(0) // A zero byte is written to support Spark 2.1.0 (SPARK-19517)
val writer = new BufferedWriter(new OutputStreamWriter(out, StandardCharsets.UTF_8))
writer.write("v" + VERSION + "\n")
writer.write(metadata.json)
writer.flush
}
override def deserialize(in: InputStream): KafkaSourceOffset = {
in.read() // A zero byte is read to support Spark 2.1.0 (SPARK-19517)
val content = IOUtils.toString(new InputStreamReader(in, StandardCharsets.UTF_8))
// HDFSMetadataLog guarantees that it never creates a partial file.
assert(content.length != 0)
if (content(0) == 'v') {
val indexOfNewLine = content.indexOf("\n")
if (indexOfNewLine > 0) {
val version = parseVersion(content.substring(0, indexOfNewLine), VERSION)
KafkaSourceOffset(SerializedOffset(content.substring(indexOfNewLine + 1)))
} else {
throw new IllegalStateException(
s"Log file was malformed: failed to detect the log file version line.")
}
} else {
// The log was generated by Spark 2.1.0
KafkaSourceOffset(SerializedOffset(content))
}
}
}
}
/** A [[InputPartition]] for reading Kafka data in a micro-batch streaming query. */
private[kafka010] case class KafkaMicroBatchInputPartition(
offsetRange: KafkaOffsetRange,
executorKafkaParams: ju.Map[String, Object],
pollTimeoutMs: Long,
failOnDataLoss: Boolean,
reuseKafkaConsumer: Boolean) extends InputPartition[InternalRow] {
override def preferredLocations(): Array[String] = offsetRange.preferredLoc.toArray
override def createPartitionReader(): InputPartitionReader[InternalRow] =
new KafkaMicroBatchInputPartitionReader(offsetRange, executorKafkaParams, pollTimeoutMs,
failOnDataLoss, reuseKafkaConsumer)
}
/** A [[InputPartitionReader]] for reading Kafka data in a micro-batch streaming query. */
private[kafka010] case class KafkaMicroBatchInputPartitionReader(
offsetRange: KafkaOffsetRange,
executorKafkaParams: ju.Map[String, Object],
pollTimeoutMs: Long,
failOnDataLoss: Boolean,
reuseKafkaConsumer: Boolean) extends InputPartitionReader[InternalRow] with Logging {
private val consumer = KafkaDataConsumer.acquire(
offsetRange.topicPartition, executorKafkaParams, reuseKafkaConsumer)
private val rangeToRead = resolveRange(offsetRange)
private val converter = new KafkaRecordToUnsafeRowConverter
private var nextOffset = rangeToRead.fromOffset
private var nextRow: UnsafeRow = _
override def next(): Boolean = {
if (nextOffset < rangeToRead.untilOffset) {
val record = consumer.get(nextOffset, rangeToRead.untilOffset, pollTimeoutMs, failOnDataLoss)
if (record != null) {
nextRow = converter.toUnsafeRow(record)
nextOffset = record.offset + 1
true
} else {
false
}
} else {
false
}
}
override def get(): UnsafeRow = {
assert(nextRow != null)
nextRow
}
override def close(): Unit = {
consumer.release()
}
private def resolveRange(range: KafkaOffsetRange): KafkaOffsetRange = {
if (range.fromOffset < 0 || range.untilOffset < 0) {
// Late bind the offset range
val availableOffsetRange = consumer.getAvailableOffsetRange()
val fromOffset = if (range.fromOffset < 0) {
assert(range.fromOffset == KafkaOffsetRangeLimit.EARLIEST,
s"earliest offset ${range.fromOffset} does not equal ${KafkaOffsetRangeLimit.EARLIEST}")
availableOffsetRange.earliest
} else {
range.fromOffset
}
val untilOffset = if (range.untilOffset < 0) {
assert(range.untilOffset == KafkaOffsetRangeLimit.LATEST,
s"latest offset ${range.untilOffset} does not equal ${KafkaOffsetRangeLimit.LATEST}")
availableOffsetRange.latest
} else {
range.untilOffset
}
KafkaOffsetRange(range.topicPartition, fromOffset, untilOffset, None)
} else {
range
}
}
}