Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Duplicate
-
0.13.1
-
None
-
Loading into orc
Description
When loading into an un-paritioned ORC table WriterImpl$StructTreeWriter.write method is synchronized.
When hive.optimize.sort.dynamic.partition is enabled the current thread will be the only writer and the synchronization is not needed.
Also checking for memory per row is an over kill , this can be done per 1K rows or such
public void addRow(Object row) throws IOException { synchronized (this) { treeWriter.write(row); rowsInStripe += 1; if (buildIndex) { rowsInIndex += 1; if (rowsInIndex >= rowIndexStride) { createRowIndexEntry(); } } } memoryManager.addedRow(); }
This can improve ORC load performance by 7%
Stack Trace Sample Count Percentage(%) WriterImpl.addRow(Object) 5,852 65.782 WriterImpl$StructTreeWriter.write(Object) 5,163 58.037 MemoryManager.addedRow() 666 7.487 MemoryManager.notifyWriters() 648 7.284 WriterImpl.checkMemory(double) 645 7.25 WriterImpl.flushStripe() 643 7.228 WriterImpl$StructTreeWriter.writeStripe(OrcProto$StripeFooter$Builder, int) 584 6.565
Attachments
Issue Links
- is related to
-
HIVE-10191 ORC: Cleanup writer per-row synchronization
- Closed