[HIVE-22446] Make IO decoding quantiles counters less contended resource.

• Type: Improvement
• Status: Patch Available
• Priority: Major
• Resolution: Unresolved
• Component/s: llap
• Assignee: Slim Bouguerra
• Reporter: Slim Bouguerra
• Created: 01/Nov/19 21:25
• Updated: 21/Oct/22 07:31

Description

Currently LLAP IO relies on Hadoop Lock-based quantiles data structure and updates the IO decoding sample on a per batch based using.

 
org.apache.hadoop.hive.llap.metrics.LlapDaemonIOMetrics#addDecodeBatchTime

via

 
org.apache.hadoop.hive.llap.io.decode.EncodedDataConsumer#consumeData

This can be a source of thread contention.
Goal of this ticket is to reduce the frequency of updates to avoid major bottleneck.