Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
3.1.1, 3.2.0, 3.3.0
-
None
Description
Spark on Yarn Cluster
When multiple executors exist on a node, and the same block exists on both executors, with some in memory and some on disk.
Probabilistically, the executor failed to obtain the block,throw Exception:
java.lang.ArrayIndexOutofBoundsException: 0
at org.apache.spark.broadcast.TorrentBroadcast.$anonfun$readBlocks$1(TorrentBroadcast.scala:183)
Next, I will replay the process of the problem occurring:
step 1:
The executor requests the driver to obtain block information(locationsAndStatusOption). The input parameters are BlockId and the host of its own node. Please note that it does not carry port information
line:1092
step 2:
On the driver side, the driver obtains all blockManagers holding the block based on the BlockId. For non remote shuffle scenarios, the driver will retrieve the first one with the blockId and blockManager from the locations
Assuming that there are two BlockManagers holding the BlockId on this node, BM-1 holds the Block and stores it in memory, and BM-2 holds the Block and stores it in disk
Assuming the returned status is of type memory and its disksize is 0
line: 852, 856
step 3:
This method will return a BlockLocationsAndStatus object. If there are BMs using disk, the disk's path information will be stored in localDirs
step 4:
When the executor obtains locationsAndStatusOption, localDirs is not empty, but status.diskSize is 0
line: 1102
step 5:
The readDiskBlockFromSameHostExecutor only determines whether the Block file exists, and then directly uses the incoming blocksize to read the byte array. If the blocksize is 0, it returns an empty byte array
Only checked if the file exists
line: 1234, 1240
Taking values from an empty array, causing an out of bounds problem
