[IMPALA-8243] ConcurrentModificationException in Catalog stress tests

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: Impala 3.1.0
• Fix Version/s: Impala 3.2.0
• Component/s: Catalog
• Target Version: Impala 3.2.0
• Epic Color: ghx-label-6
• Assignee: Bharath Vissapragada
• Reporter: Bharath Vissapragada
• Created: 25/Feb/19 05:34
• Updated: 14/Mar/19 15:34
• Resolved: 14/Mar/19 15:34

Description

Following is the full stack from the Catalog server logs.

14:09:29.474424 14829 jni-util.cc:256] java.util.ConcurrentModificationException
java.util.ArrayList$Itr.checkForComodification(ArrayList.java:901)
java.util.ArrayList$Itr.next(ArrayList.java:851)
org.apache.hadoop.hive.metastore.api.StorageDescriptor$StorageDescriptorStandardScheme.write(StorageDescriptor.java:1449)
org.apache.hadoop.hive.metastore.api.StorageDescriptor$StorageDescriptorStandardScheme.write(StorageDescriptor.java:1278)
org.apache.hadoop.hive.metastore.api.StorageDescriptor.write(StorageDescriptor.java:1144)
org.apache.hadoop.hive.metastore.api.Partition$PartitionStandardScheme.write(Partition.java:1062)
org.apache.hadoop.hive.metastore.api.Partition$PartitionStandardScheme.write(Partition.java:919)
org.apache.hadoop.hive.metastore.api.Partition.write(Partition.java:815)
org.apache.impala.thrift.TPartialPartitionInfo$TPartialPartitionInfoStandardScheme.write(TPartialPartitionInfo.java:862)
org.apache.impala.thrift.TPartialPartitionInfo$TPartialPartitionInfoStandardScheme.write(TPartialPartitionInfo.java:759)
org.apache.impala.thrift.TPartialPartitionInfo.write(TPartialPartitionInfo.java:665)
org.apache.impala.thrift.TPartialTableInfo$TPartialTableInfoStandardScheme.write(TPartialTableInfo.java:731)
org.apache.impala.thrift.TPartialTableInfo$TPartialTableInfoStandardScheme.write(TPartialTableInfo.java:624)
org.apache.impala.thrift.TPartialTableInfo.write(TPartialTableInfo.java:543)
org.apache.impala.thrift.TGetPartialCatalogObjectResponse$TGetPartialCatalogObjectResponseStandardScheme.write(TGetPartialCatalogObjectResponse.java:977)
org.apache.impala.thrift.TGetPartialCatalogObjectResponse$TGetPartialCatalogObjectResponseStandardScheme.write(TGetPartialCatalogObjectResponse.java:857)
org.apache.impala.thrift.TGetPartialCatalogObjectResponse.write(TGetPartialCatalogObjectResponse.java:739)
org.apache.thrift.TSerializer.serialize(TSerializer.java:79)
org.apache.impala.service.JniCatalog.getPartialCatalogObject(JniCatalog.java:233)

It looks like the bug is in the following piece of code.

/**
   * Returns a Hive-compatible partition object that may be used in calls to the
   * metastore.
   */
  public org.apache.hadoop.hive.metastore.api.Partition toHmsPartition() {
    if (cachedMsPartitionDescriptor_ == null) return null;
    Preconditions.checkNotNull(table_.getNonPartitionFieldSchemas());
    // Update the serde library class based on the currently used file format.
    org.apache.hadoop.hive.metastore.api.StorageDescriptor storageDescriptor =
        new org.apache.hadoop.hive.metastore.api.StorageDescriptor(
            table_.getNonPartitionFieldSchemas(),  <===== Reference to the actual field schema list.
            getLocation(),
            cachedMsPartitionDescriptor_.sdInputFormat,
            cachedMsPartitionDescriptor_.sdOutputFormat,
            cachedMsPartitionDescriptor_.sdCompressed,

It appears we are leaking a reference to nonPartFieldSchemas_ in to the thrift object and once the thread leaves the lock scope, some other thread (load() for ex: ) can potentially change the source list and the serialization code could throw ConcurrentModificationException

While the stack above is Catalog-v2 only, it is possible that some other threads can race in a similar fashion.