Detect multiple jars on the classpath when init plugin [databricks]

sql-plugin/src/main/scala/com/nvidia/spark/rapids/Plugin.scala

@@ -17,6 +17,7 @@
 package com.nvidia.spark.rapids
 
 import java.lang.reflect.InvocationTargetException
+import java.net.URL
 import java.time.ZoneId
 import java.util.Properties
 
@@ -25,6 +26,7 @@ import scala.sys.process._
 import scala.util.Try
 
 import ai.rapids.cudf.{Cuda, CudaException, CudaFatalException, CudfException, MemoryCleaner}
+import com.nvidia.spark.rapids.RapidsConf.AllowMultipleJars
 import com.nvidia.spark.rapids.filecache.{FileCache, FileCacheLocalityManager, FileCacheLocalityMsg}
 import com.nvidia.spark.rapids.python.PythonWorkerSemaphore
 import org.apache.commons.lang3.exception.ExceptionUtils
@@ -111,6 +113,52 @@ object RapidsPluginUtils extends Logging {
     }
   }
 
+  private def detectMultipleJar(propName: String, jarName: String, conf: RapidsConf): Unit = {
+    val classloader = ShimLoader.getShimClassLoader()
+    val possibleRapidsJarURLs = classloader.getResources(propName).asScala.toSet.toSeq.filterNot {
+      url => url.toString.contains("test")
+    }
+    val revisionMap: Map[String, Seq[URL]] = possibleRapidsJarURLs.map { url =>
+      val versionInfo = scala.io.Source.fromURL(url).getLines().toSeq
+      lazy val revision = versionInfo
+        .filter(_.startsWith("revision="))
+        .map(_.split("=").last)
+        .headOption
+        .getOrElse("UNKNOWN")
+      (revision, url)
+    }.groupBy(_._1).mapValues(_.map(_._2)).toMap
+    lazy val rapidsJarsVersMsg = revisionMap.map {
+      case (revision, urls) => {
+        s"revison: $revision" + urls.map {
+          url => "\n\tjar URL: " + url.toString.split("!").head + "\n\t" + 
+              scala.io.Source.fromURL(url).getLines().toSeq.mkString("\n\t")
+        }.mkString + "\n"
+      }
+    }.mkString
+    // scalastyle:off line.size.limit
+    lazy val msg = s"""Multiple $jarName jars found in the classpath:
+        |$rapidsJarsVersMsg
+        |Please make sure there is only one $jarName jar in the classpath.
+        |If it is impossible to fix the classpath you can suppress the error by setting ${RapidsConf.ALLOW_MULTIPLE_JARS.key} to SAME_REVISION or ALWAYS.
+        """.stripMargin
+    // scalastyle:on line.size.limit
+
+    conf.allowMultipleJars match {
+      case AllowMultipleJars.ALWAYS =>
+        logWarning(msg)
+      case AllowMultipleJars.SAME_REVISION =>
+        require(revisionMap.size == 1, msg)
+      case AllowMultipleJars.NEVER =>
+        require(revisionMap.size == 1 && revisionMap.values.forall(_.size == 1), msg)
+    }
+  }
+
+  def detectMultipleJars(conf: RapidsConf): Unit = {
+    detectMultipleJar(PLUGIN_PROPS_FILENAME, "rapids-4-spark", conf)
+    detectMultipleJar(JNI_PROPS_FILENAME, "spark-rapids-jni", conf)
+    detectMultipleJar(CUDF_PROPS_FILENAME, "cudf", conf)
+  }
+
   // This assumes Apache Spark logic, if CSPs are setting defaults differently, we may need
   // to handle.
   def estimateCoresOnExec(conf: SparkConf): Int = {
@@ -309,6 +357,7 @@ class RapidsDriverPlugin extends DriverPlugin with Logging {
     val sparkConf = pluginContext.conf
     RapidsPluginUtils.fixupConfigsOnDriver(sparkConf)
     val conf = new RapidsConf(sparkConf)
+    RapidsPluginUtils.detectMultipleJars(conf)
     RapidsPluginUtils.logPluginMode(conf)
     GpuCoreDumpHandler.driverInit(sc, conf)
 
@@ -363,6 +412,9 @@ class RapidsExecutorPlugin extends ExecutorPlugin with Logging {
       val numCores = RapidsPluginUtils.estimateCoresOnExec(sparkConf)
       val conf = new RapidsConf(extraConf.asScala.toMap)
 
+      // Fail if there are multiple plugin jars in the classpath.
+      RapidsPluginUtils.detectMultipleJars(conf)
+
       // Compare if the cudf version mentioned in the classpath is equal to the version which
       // plugin expects. If there is a version mismatch, throw error. This check can be disabled
       // by setting this config spark.rapids.cudfVersionOverride=true

sql-plugin/src/main/scala/com/nvidia/spark/rapids/RapidsConf.scala

@@ -1839,6 +1839,22 @@ object RapidsConf {
     .booleanConf
     .createWithDefault(false)
 
+  object AllowMultipleJars extends Enumeration {
+    val ALWAYS, SAME_REVISION, NEVER = Value
+  }
+
+  val ALLOW_MULTIPLE_JARS = conf("spark.rapids.sql.allowMultipleJars")
+    .internal()
+    .startupOnly()
+    .doc("Allow multiple rapids-4-spark, spark-rapids-jni, and cudf jars on the classpath. " +
+      "Spark will take the first one it finds, so the version may not be expected. Possisble " +
+      "values are ALWAYS: allow all jars, SAME_REVISION: only allow jars with the same " +
+      "revision, NEVER: do not allow multiple jars at all.")
+    .stringConf
+    .transform(_.toUpperCase(java.util.Locale.ROOT))
+    .checkValues(AllowMultipleJars.values.map(_.toString))
+    .createWithDefault(AllowMultipleJars.SAME_REVISION.toString)
+
   val ALLOW_DISABLE_ENTIRE_PLAN = conf("spark.rapids.allowDisableEntirePlan")
     .internal()
     .doc("The plugin has the ability to detect possibe incompatibility with some specific " +
@@ -2634,6 +2650,17 @@ class RapidsConf(conf: Map[String, String]) extends Logging {
 
   lazy val cudfVersionOverride: Boolean = get(CUDF_VERSION_OVERRIDE)
 
+  lazy val allowMultipleJars: AllowMultipleJars.Value = {
+    get(ALLOW_MULTIPLE_JARS) match {
+      case "ALWAYS" => AllowMultipleJars.ALWAYS
+      case "NEVER" => AllowMultipleJars.NEVER
+      case "SAME_REVISION" => AllowMultipleJars.SAME_REVISION
+      case other =>
+        throw new IllegalArgumentException(s"Internal Error $other is not supported for " +
+            s"${ALLOW_MULTIPLE_JARS.key}")
+    }
+  }
+
   lazy val allowDisableEntirePlan: Boolean = get(ALLOW_DISABLE_ENTIRE_PLAN)
 
   lazy val useArrowCopyOptimization: Boolean = get(USE_ARROW_OPT)