viirya commented on code in PR #55420: URL: https://github.com/apache/spark/pull/55420#discussion_r3255192422
########## connector/kafka-0-10-sql/src/test/scala/org/apache/spark/sql/kafka010/benchmark/RTMKafkaKafkaBenchmark.scala: ########## @@ -0,0 +1,353 @@ +/* + * 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.benchmark + +import java.nio.file.Files +import java.util.{Properties, Timer, TimerTask} +import java.util.concurrent.{CountDownLatch, TimeUnit} +import java.util.concurrent.atomic.{AtomicInteger, AtomicLong} + +import scala.concurrent.duration._ + +import org.apache.kafka.clients.producer.{Callback, KafkaProducer, Producer, ProducerRecord, RecordMetadata} + +import org.apache.spark.benchmark.{Benchmark, BenchmarkBase} +import org.apache.spark.internal.Logging +import org.apache.spark.sql.{Column, SparkSession} +import org.apache.spark.sql.execution.streaming.RealTimeTrigger +import org.apache.spark.sql.functions._ +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.sql.kafka010.KafkaTestUtils +import org.apache.spark.sql.streaming.StreamingQueryListener + +/** + * Stateless Kafka-to-Kafka RTM benchmark. Reads from an input Kafka topic, applies a + * stateless transformation, and writes results to an output Kafka topic using + * [[RealTimeTrigger]]. After the run it reports e2e latency percentiles. + * + * The benchmark spins up a real local-cluster Spark context and a live embedded Kafka + * broker, so a single run takes several minutes. + * + * To run this benchmark: + * {{{ + * 1. without sbt: + * bin/spark-submit --class <this class> + * --jars <spark core test jar>,<spark sql test jar> <spark sql kafka 0-10 test jar> + * 2. build/sbt "sql-kafka-0-10/Test/runMain <this class>" + * 3. generate result: + * SPARK_GENERATE_BENCHMARK_FILES=1 build/sbt "sql-kafka-0-10/Test/runMain <this class>" + * Results will be written to: + * "connector/kafka-0-10-sql/benchmarks/RTMKafkaKafkaBenchmark-results.txt". + * }}} + * + * See `benchmarks/RTMKafkaKafkaBenchmark-results.txt` for a recorded run. + */ +object RTMKafkaKafkaBenchmark extends BenchmarkBase with Logging { + + private val topicId = new AtomicInteger(0) + private var spark: SparkSession = _ + private var testUtils: KafkaTestUtils = _ + + override def runBenchmarkSuite(mainArgs: Array[String]): Unit = { + // BenchmarkBase.main does not wrap this call in try/finally, so we must own + // teardown ourselves: partial setup, a timeout, or a getLatencies failure + // would otherwise leak the embedded Kafka broker and local-cluster workers. + testUtils = new KafkaTestUtils(Map.empty) + try { + testUtils.setup() + spark = SparkSession.builder() + .master("local-cluster[3, 5, 1024]") + .appName(this.getClass.getCanonicalName) + .getOrCreate() + runBenchmark("RTM stateless kafka-to-kafka") { + benchmark(60.seconds.toMillis, 4) + } + } finally { + cleanup() + } + } + + /** + * Idempotent cleanup of the Spark session and embedded Kafka broker. Safe to call + * after any combination of partial setup, normal completion, or exception. + */ + private def cleanup(): Unit = { + if (spark != null) { + try { + spark.stop() + } catch { + case t: Throwable => logWarning("Failed to stop SparkSession during cleanup", t) + } + spark = null + } + if (testUtils != null) { + try { + testUtils.teardown() + } catch { + case t: Throwable => logWarning("Failed to teardown KafkaTestUtils during cleanup", t) + } + testUtils = null + } + } + + private def newTopic(): String = s"topic-${topicId.getAndIncrement()}" + + def benchmark(longRunningBatchDurationMs: Long, numBatches: Long): Unit = { + val inputTopic = newTopic() + testUtils.createTopic(inputTopic, partitions = 5) + + val outputTopic = newTopic() + testUtils.createTopic(outputTopic, partitions = 5) + + spark.conf.set(SQLConf.STREAMING_POLLING_DELAY.key, 10) + + val kafkaStream = spark.readStream + .format("kafka") + .option("kafka.bootstrap.servers", testUtils.brokerAddress) + .option("subscribe", inputTopic) + .option("kafka.fetch.max.wait.ms", "10") + .option("kafka.max.partition.fetch.bytes", "10485760") // 10MB + .load() + + val currentTimestampUDF = udf(() => System.currentTimeMillis()) Review Comment: Spark's built-in current_timestamp() in a streaming context is evaluated once per batch for determinism — which is the exact opposite of what this benchmark wants (per-row wall-clock timestamp). This is a subtle correctness point: anyone seeing a UDF wrapping System.currentTimeMillis() will be tempted to "clean it up" to the built-in and silently change the semantics. Please add an inline comment, e.g.: ``` // UDF instead of current_timestamp(): the built-in is evaluated once per batch // for streaming determinism, but we want per-row wall-clock to measure per-record latency. val currentTimestampUDF = udf(() => System.currentTimeMillis()) ``` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
