[HIVE-20247] cleanup issues in LLAP IO after cache OOM

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 4.0.0-alpha-1
• Assignee: Sergey Shelukhin
• Reporter: Prasanth Jayachandran
• Created: 26/Jul/18 22:47
• Updated: 17/Nov/22 08:54
• Resolved: 30/Jul/18 19:56

Description

LLAP IO creates unallocated buffer objects inside the read-related data structures, then allocates them in bulk, then decompresses into them and increfs them.
If allocate or decompress steps fail, it's hard for the higher-level cleanup to tell what the state of the buffers in the read-related structures is - they may be unallocated, allocated but not incref-ed, or incref-ed.
Some cleanup paths only deal with the latter case, resulting in bugs.
Moreover, currently allocator returns partial results on such error. The allocation should be all-or-nothing.

This only happens on one path, others allocate and use buffers in a single place.

Issue Links

• links to: RB