dragos commented on a change in pull request #7648: [SPARK-8979] Add a PID based rate estimator URL: https://github.com/apache/spark/pull/7648#discussion_r386411575
########## File path: streaming/src/main/scala/org/apache/spark/streaming/scheduler/rate/PIDRateEstimator.scala ########## @@ -0,0 +1,100 @@ +/* + * 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.streaming.scheduler.rate + +/** + * Implements a proportional-integral-derivative (PID) controller which acts on + * the speed of ingestion of elements into Spark Streaming. A PID controller works + * by calculating an '''error''' between a measured output and a desired value. In the + * case of Spark Streaming the error is the difference between the measured processing + * rate (number of elements/processing delay) and the previous rate. + * + * @see https://en.wikipedia.org/wiki/PID_controller + * + * @param batchDurationMillis the batch duration, in milliseconds + * @param proportional how much the correction should depend on the current + * error. This term usually provides the bulk of correction. A value too large would + * make the controller overshoot the setpoint, while a small value would make the + * controller too insensitive. The default value is -1. + * @param integral how much the correction should depend on the accumulation + * of past errors. This term accelerates the movement towards the setpoint, but a large + * value may lead to overshooting. The default value is -0.2. + * @param derivative how much the correction should depend on a prediction + * of future errors, based on current rate of change. This term is not used very often, + * as it impacts stability of the system. The default value is 0. + */ +private[streaming] class PIDRateEstimator( + batchIntervalMillis: Long, + proportional: Double = -1D, + integral: Double = -.2D, + derivative: Double = 0D) + extends RateEstimator { + + private var firstRun: Boolean = true + private var latestTime: Long = -1L + private var latestRate: Double = -1D + private var latestError: Double = -1L + + require( + batchIntervalMillis > 0, + s"Specified batch interval $batchIntervalMillis in PIDRateEstimator is invalid.") + + def compute(time: Long, // in milliseconds + elements: Long, + processingDelay: Long, // in milliseconds + schedulingDelay: Long // in milliseconds + ): Option[Double] = { + + this.synchronized { + if (time > latestTime && processingDelay > 0 && batchIntervalMillis > 0) { + + // in seconds, should be close to batchDuration + val delaySinceUpdate = (time - latestTime).toDouble / 1000 + + // in elements/second + val processingRate = elements.toDouble / processingDelay * 1000 + + // in elements/second + val error = latestRate - processingRate + + // in elements/second + val sumError = schedulingDelay.toDouble * processingRate / batchIntervalMillis Review comment: @YikSanChan sorry for the delay, it's been a while since I didn't visit this code. I admit I am not sure I understand where you're coming from, and I don't want to simply repeat what I wrote above. So I'll ask a question or two myself first: What prompted you to look into this code? Do you see bad behavior on certain loads? > The explanation that if # overflowing elements = 0, then historical error = 0 does not explain the 1 / batch interval coefficient. If the overflowing elements are 0, doesn't this imply the `schedulingDelay` is zero, so this whole term is 0 as well (instead of `1/batchInterval`)? >I feel the hard part is: It is easy to reason about the accumulated error if error is an absolute number, for example, the number of overflowing elements, but what is an accumulated error of a rate? If we agree that the accumulated error is what it is computed above as number of elements, then "all we need to do" is to transform it into a rate (elements/second). We need to do that since that's the quantity that this PID is controlling. By dividing it by the batch interval we get how many `elements/second` are overflowing, in other words, by how much is the existing rate "off", if we wanted to have no overflowing elements. Does this make sense to you? ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
