Skip to content

Reading manifest Avro hangs when S3A inputstream returns less bytes than requested #1654

@mickjermsurawong-stripe

Description

Observation: Spark job intermittently hangs with driver thread stuck in reading avro file with S3A inputstream, according to Thread dump here.

java.io.DataInputStream.read(DataInputStream.java:149)
org.apache.iceberg.hadoop.HadoopStreams$HadoopSeekableInputStream.read(HadoopStreams.java:112)
org.apache.iceberg.avro.AvroIO$AvroInputStreamAdapter.read(AvroIO.java:106)
org.apache.iceberg.shaded.org.apache.avro.file.DataFileReader.openReader(DataFileReader.java:61)
org.apache.iceberg.avro.AvroIterable.newFileReader(AvroIterable.java:94)
org.apache.iceberg.avro.AvroIterable.iterator(AvroIterable.java:77)
org.apache.iceberg.avro.AvroIterable.iterator(AvroIterable.java:37)
org.apache.iceberg.relocated.com.google.common.collect.Iterables.addAll(Iterables.java:320)
org.apache.iceberg.relocated.com.google.common.collect.Lists.newLinkedList(Lists.java:237)
org.apache.iceberg.ManifestLists.read(ManifestLists.java:46)
org.apache.iceberg.BaseSnapshot.cacheManifests(BaseSnapshot.java:127)
org.apache.iceberg.BaseSnapshot.dataManifests(BaseSnapshot.java:149)
org.apache.iceberg.DataTableScan.planFiles(DataTableScan.java:74)
org.apache.iceberg.BaseTableScan.planFiles(BaseTableScan.java:212)
org.apache.iceberg.DataTableScan.planFiles(DataTableScan.java:28)
org.apache.iceberg.BaseTableScan.planTasks(BaseTableScan.java:245)
org.apache.iceberg.DataTableScan.planTasks(DataTableScan.java:28)
org.apache.iceberg.spark.source.Reader.tasks(Reader.java:365)

Affected code: DataFileReader.openReader

private DataFileReader<D> newFileReader() {
try {
return (DataFileReader<D>) DataFileReader.openReader(
AvroIO.stream(file.newStream(), file.getLength()), reader);
} catch (IOException e) {
throw new RuntimeIOException(e, "Failed to open file: %s", file);
}
}

Root cause: Avro reader checking header has incorrect logic and will get stuck in infinite loop when S3A input stream reads fewer bytes than requested. This has been raised with Avro https://issues.apache.org/jira/browse/AVRO-2944

Current: https://github.com/apache/avro/blob/328c539afc77da347ec52be1e112a6a7c371143b/lang/java/avro/src/main/java/org/apache/avro/file/DataFileReader.java#L61-L62

for (int c = 0; c < magic.length; c = in.read(magic, c, magic.length - c)) {}

Fix in Avro:

for (int c = 0; c < magic.length; c += in.read(magic, c, magic.length - c)) {}

Iceberg change needed: There is PR for Avro apache/avro#969 and I'm wondering if Iceberg can take this Avro patch in, once merged.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions