[IMPALA-2400] Unpredictable locality behavior for reading Parquet files

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Cannot Reproduce
• Affects Version/s: Impala 2.3.0
• Component/s: Perf Investigation
• Labels: ramp-up
• Target Version: Product Backlog
• Assignee: Unassigned
• Reporter: Mostafa Mokhtar
• Created: 23/Sep/15 20:47
• Updated: 13/Jun/20 00:04
• Resolved: 13/Jun/20 00:04

Description

When running the query below I noticed exceptionally high variance even after running "invalidate metadata".

select * from tpch_bin_flat_parquet_30.lineitem limit 10;

• Fetched 10 row(s) in 1.08s
  WARNINGS: Read 139.48 MB of data across network that was expected to be local. Block locality metadata for table 'tpch_bin_flat_parquet_30.lineitem' may be stale. Consider running "INVALIDATE METADATA `tpch_bin_flat_parquet_30`.`lineitem`".
• Fetched 10 row(s) in 1.32s
• Fetched 10 row(s) in 0.09s
• Fetched 10 row(s) in 1.08s
• "invalidate metadata"
• Fetched 10 row(s) in 0.89s
• Fetched 10 row(s) in 0.07s
  WARNINGS: Read 76.15 MB of data across network that was expected to be local. Block locality metadata for table 'tpch_bin_flat_parquet_30.lineitem' may be stale. Consider running "INVALIDATE METADATA `tpch_bin_flat_parquet_30`.`lineitem`".
• Fetched 10 row(s) in 1.11s
• Fetched 10 row(s) in 0.73s
• Fetched 10 row(s) in 0.09s

The behavior above is tied to Parquet tables and doesn't repro against text data.

Profile files attached.