[IMPALA-2940] Parquet DictDecoders accumulate throughout query

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: Impala 2.3.0
• Fix Version/s: Impala 2.5.0, Impala 2.3.4
• Component/s: Backend
• Labels: resource-management
• Target Version: Impala 2.5.0, Impala 2.3.4
• Assignee: Tim Armstrong
• Reporter: Tim Armstrong
• Created: 04/Feb/16 00:19
• Updated: 23/Feb/16 04:57
• Resolved: 05/Feb/16 15:01

Description

Parquet dictionary decoders can accumulate throughout query execution. One is created per-column per-split. The decoder contains an vector of values for the dictionary that is not cleared when the scanner is finished with it.

I've attached a graph of memory usage when running this query on TPC-DS scale factor 100. Before is cdh5-trunk, and after is with a fix that delete the ColumnReader objects after each input split.

use tpcds_100_parquet;                                                           
set num_scanner_threads=8;
select * from store_sales where ss_sold_time_sk = 12512434;

Issue Links

• relates to: IMPALA-2885 Scanners store per-split objects in per-query object pool