[IMPALA-3316] convert_legacy_hive_parquet_utc_timestamps=true makes reading parquet tables 30x slower

• Type: Bug
• Status: Open
• Priority: Minor
• Resolution: Unresolved
• Affects Version/s: impala 2.3
• Component/s: Backend
• Environment: CDH 5.5.2/ Impala 2.3; Parquet table with a timestamp column; Secure cluster; convert_legacy_hive_parquet_utc_timestamps=true; Timestamp column is not being filtered on
• Target Version: Product Backlog
• Assignee: Attila Jeges
• Reporter: Ruslan Dautkhanov
• Created: 07/Apr/16 23:31
• Updated: 29/Mar/18 16:03

Description

Enabling convert_legacy_hive_parquet_utc_timestamps=true
makes simple queries that don't even filter on a timestamp attribute perform really poorly.

Parquet table.
Impala 2.3 / CDH 5.5.2.

convert_legacy_hive_parquet_utc_timestamps=true makes following simple query 30x slower (1.1minutes -> over 30 minutes).

select * from parquet_table_with_a_timestamp_attribute where bigint_attribute=1000771658169

Notice I did not even filter on a timestamp attribute.

Made multiple tests with and without convert_legacy_hive_parquet_utc_timestamps=true impalad present.

Also, from https://issues.cloudera.org/browse/IMPALA-1658

Casey Ching added a comment - 15/Jun/15 5:12 PM
Btw, a perf test showed enabling this flag was 10x slower.

Issue Links

• relates to: IMPALA-1773 Implement TIMESTAMP WITH TIME ZONE data type
• relates to: IMPALA-2125 Improve perf when reading timestamps from parquet files written by hive
• relates to: IMPALA-1658 Add a compatibility option for reading parquet timestamps written by Hive
• links to: Google CCTZ c++ library
• links to: TimestampValue::UtcToLocal