From 7e63bb49c526c3f872619ae14e4b5273f4c535e9 Mon Sep 17 00:00:00 2001 From: Josh Rosen Date: Sun, 19 Oct 2014 00:31:06 -0700 Subject: [SPARK-2546] Clone JobConf for each task (branch-1.0 / 1.1 backport) This patch attempts to fix SPARK-2546 in `branch-1.0` and `branch-1.1`. The underlying problem is that thread-safety issues in Hadoop Configuration objects may cause Spark tasks to get stuck in infinite loops. The approach taken here is to clone a new copy of the JobConf for each task rather than sharing a single copy between tasks. Note that there are still Configuration thread-safety issues that may affect the driver, but these seem much less likely to occur in practice and will be more complex to fix (see discussion on the SPARK-2546 ticket). This cloning is guarded by a new configuration option (`spark.hadoop.cloneConf`) and is disabled by default in order to avoid unexpected performance regressions for workloads that are unaffected by the Configuration thread-safety issues. Author: Josh Rosen Closes #2684 from JoshRosen/jobconf-fix-backport and squashes the following commits: f14f259 [Josh Rosen] Add configuration option to control cloning of Hadoop JobConf. b562451 [Josh Rosen] Remove unused jobConfCacheKey field. dd25697 [Josh Rosen] [SPARK-2546] [1.0 / 1.1 backport] Clone JobConf for each task. (cherry picked from commit 2cd40db2b3ab5ddcb323fd05c171dbd9025f9e71) Signed-off-by: Josh Rosen Conflicts: core/src/main/scala/org/apache/spark/rdd/HadoopRDD.scala --- .../scala/org/apache/spark/rdd/HadoopRDD.scala | 53 ++++++++++++++++------ docs/configuration.md | 9 ++++ 2 files changed, 47 insertions(+), 15 deletions(-) diff --git a/core/src/main/scala/org/apache/spark/rdd/HadoopRDD.scala b/core/src/main/scala/org/apache/spark/rdd/HadoopRDD.scala index 8010dd9008..775141775e 100644 --- a/core/src/main/scala/org/apache/spark/rdd/HadoopRDD.scala +++ b/core/src/main/scala/org/apache/spark/rdd/HadoopRDD.scala @@ -132,27 +132,47 @@ class HadoopRDD[K, V]( // used to build JobTracker ID private val createTime = new Date() + private val shouldCloneJobConf = sc.conf.get("spark.hadoop.cloneConf", "false").toBoolean + // Returns a JobConf that will be used on slaves to obtain input splits for Hadoop reads. protected def getJobConf(): JobConf = { val conf: Configuration = broadcastedConf.value.value - if (conf.isInstanceOf[JobConf]) { - // A user-broadcasted JobConf was provided to the HadoopRDD, so always use it. - conf.asInstanceOf[JobConf] - } else if (HadoopRDD.containsCachedMetadata(jobConfCacheKey)) { - // getJobConf() has been called previously, so there is already a local cache of the JobConf - // needed by this RDD. - HadoopRDD.getCachedMetadata(jobConfCacheKey).asInstanceOf[JobConf] - } else { - // Create a JobConf that will be cached and used across this RDD's getJobConf() calls in the - // local process. The local cache is accessed through HadoopRDD.putCachedMetadata(). - // The caching helps minimize GC, since a JobConf can contain ~10KB of temporary objects. - // Synchronize to prevent ConcurrentModificationException (Spark-1097, Hadoop-10456). + if (shouldCloneJobConf) { + // Hadoop Configuration objects are not thread-safe, which may lead to various problems if + // one job modifies a configuration while another reads it (SPARK-2546). This problem occurs + // somewhat rarely because most jobs treat the configuration as though it's immutable. One + // solution, implemented here, is to clone the Configuration object. Unfortunately, this + // clone can be very expensive. To avoid unexpected performance regressions for workloads and + // Hadoop versions that do not suffer from these thread-safety issues, this cloning is + // disabled by default. HadoopRDD.CONFIGURATION_INSTANTIATION_LOCK.synchronized { + logDebug("Cloning Hadoop Configuration") val newJobConf = new JobConf(conf) - initLocalJobConfFuncOpt.map(f => f(newJobConf)) - HadoopRDD.putCachedMetadata(jobConfCacheKey, newJobConf) + if (!conf.isInstanceOf[JobConf]) { + initLocalJobConfFuncOpt.map(f => f(newJobConf)) + } newJobConf } + } else { + if (conf.isInstanceOf[JobConf]) { + logDebug("Re-using user-broadcasted JobConf") + conf.asInstanceOf[JobConf] + } else if (HadoopRDD.containsCachedMetadata(jobConfCacheKey)) { + logDebug("Re-using cached JobConf") + HadoopRDD.getCachedMetadata(jobConfCacheKey).asInstanceOf[JobConf] + } else { + // Create a JobConf that will be cached and used across this RDD's getJobConf() calls in the + // local process. The local cache is accessed through HadoopRDD.putCachedMetadata(). + // The caching helps minimize GC, since a JobConf can contain ~10KB of temporary objects. + // Synchronize to prevent ConcurrentModificationException (SPARK-1097, HADOOP-10456). + HadoopRDD.CONFIGURATION_INSTANTIATION_LOCK.synchronized { + logDebug("Creating new JobConf and caching it for later re-use") + val newJobConf = new JobConf(conf) + initLocalJobConfFuncOpt.map(f => f(newJobConf)) + HadoopRDD.putCachedMetadata(jobConfCacheKey, newJobConf) + newJobConf + } + } } } @@ -276,7 +296,10 @@ class HadoopRDD[K, V]( } private[spark] object HadoopRDD extends Logging { - /** Constructing Configuration objects is not threadsafe, use this lock to serialize. */ + /** + * Configuration's constructor is not threadsafe (see SPARK-1097 and HADOOP-10456). + * Therefore, we synchronize on this lock before calling new JobConf() or new Configuration(). + */ val CONFIGURATION_INSTANTIATION_LOCK = new Object() /** diff --git a/docs/configuration.md b/docs/configuration.md index f0204c640b..96fa1377ec 100644 --- a/docs/configuration.md +++ b/docs/configuration.md @@ -619,6 +619,15 @@ Apart from these, the following properties are also available, and may be useful output directories. We recommend that users do not disable this except if trying to achieve compatibility with previous versions of Spark. Simply use Hadoop's FileSystem API to delete output directories by hand. + + spark.hadoop.cloneConf + false + If set to true, clones a new Hadoop Configuration object for each task. This + option should be enabled to work around Configuration thread-safety issues (see + SPARK-2546 for more details). + This is disabled by default in order to avoid unexpected performance regressions for jobs that + are not affected by these issues. + spark.executor.heartbeatInterval 10000 -- cgit v1.2.3