Developer Troubleshooting

Out of Memory Issues

Heap Dump on OOM

Set PXF_OOM_DUMP_PATH in ${PXF_CONF}/conf/pxf-env.sh to enable automatic heap dumps:

export PXF_OOM_DUMP_PATH=/tmp/pxf-heapdumps

This adds the following JVM flags:

-XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=$PXF_OOM_DUMP_PATH

注意

Heap dump files can be several GBs. Ensure sufficient disk space at the dump path.

• If the path is a directory, a new dump is created per OOM event: java_pid<PID>.hprof
• If the path is a file, only the first OOM generates a dump (subsequent dumps won't overwrite)

Manual Heap Dump

jmap -dump:live,format=b,file=/tmp/heap_dump "$(pgrep -f tomcat)"

The live option forces a full GC before the dump.

JVM Statistics Collection

Collect GC statistics every 60 seconds:

jstat -gcutil $(pgrep -f tomcat) 60000 > /tmp/jstat_pxf_1min.out &

Reducing Memory Pressure

Set PXF_MAX_THREADS in ${PXF_CONF}/conf/pxf-env.sh to limit concurrent request handling. Excess requests will error out instead of exhausting JVM memory.

CLI Issues

Cluster Commands Fail from Coordinator

Error: pxf cluster commands should only be run from Greenplum master

The hostname must match the entry in the gp_segment_configuration table for the coordinator node. Either update the hostname or the table entry.

Dataproc

Accessing Dataproc from External Networks

Dataproc uses internal IPs for partition locations. To use hostnames instead, set in hdfs-site.xml:

<property>
    <name>dfs.client.use.datanode.hostname</name>
    <value>true</value>
</property>

Kerberos on Dataproc

A kerberized Dataproc cluster may start with permission checking disabled. Enable it in hdfs-site.xml:

<property>
    <name>dfs.permissions.enabled</name>
    <value>true</value>
</property>