[SPARK-27966] input_file_name empty when listing files in parallel

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Incomplete
• Affects Version/s: 2.4.0
• Component/s: Input/Output
• Labels: bulk-closed

Environment

Databricks 5.3 (includes Apache Spark 2.4.0, Scala 2.11)

 
Worker Type: 14.0 GB Memory, 4 Cores, 0.75 DBU Standard_DS3_v2
Workers: 3
Driver Type: 14.0 GB Memory, 4 Cores, 0.75 DBU Standard_DS3_v2

• Assignee: Unassigned
• Reporter: Christian Homberg
• Created: 06/Jun/19 12:35
• Updated: 12/Dec/22 18:11
• Resolved: 25/May/21 01:43

Description

I ran into an issue similar and probably related to SPARK-26128. The org.apache.spark.sql.functions.input_file_name is sometimes empty.

 

df.select(input_file_name()).show(5,false)

 

+-----------------+
|input_file_name()|
+-----------------+
| |
| |
| |
| |
| |
+-----------------+

My environment is databricks and debugging the Log4j output showed me that the issue occurred when the files are being listed in parallel, e.g. when 

19/06/06 11:50:47 INFO InMemoryFileIndex: Start listing leaf files and directories. Size of Paths: 127; threshold: 32
19/06/06 11:50:47 INFO InMemoryFileIndex: Listing leaf files and directories in parallel under:

 

Everything's fine as long as

19/06/06 11:54:43 INFO InMemoryFileIndex: Start listing leaf files and directories. Size of Paths: 6; threshold: 32
19/06/06 11:54:43 INFO InMemoryFileIndex: Start listing leaf files and directories. Size of Paths: 0; threshold: 32
19/06/06 11:54:43 INFO InMemoryFileIndex: Start listing leaf files and directories. Size of Paths: 0; threshold: 32
19/06/06 11:54:43 INFO InMemoryFileIndex: Start listing leaf files and directories. Size of Paths: 0; threshold: 32
19/06/06 11:54:43 INFO InMemoryFileIndex: Start listing leaf files and directories. Size of Paths: 0; threshold: 32
19/06/06 11:54:43 INFO InMemoryFileIndex: Start listing leaf files and directories. Size of Paths: 0; threshold: 32
19/06/06 11:54:43 INFO InMemoryFileIndex: Start listing leaf files and directories. Size of Paths: 0; threshold: 32

 

Setting spark.sql.sources.parallelPartitionDiscovery.threshold to 9999 resolves the issue for me.

 

edit: the problem is not exclusively linked to listing files in parallel. I've setup a larger cluster for which after parallel file listing the input_file_name did return the correct filename. After inspecting the log4j again, I assume that it's linked to some kind of MetaStore being full. I've attached a section of the log4j output that I think should indicate why it's failing. If you need more, please let me know.

 ** 

 

 

Issue Links

• is related to: SPARK-28153 Use AtomicReference at InputFileBlockHolder (to support input_file_name with Python UDF)