[KUDU-1931] Servers using the FBM may crash after a Flush

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.2.0
• Fix Version/s: 1.4.0
• Component/s: fs
• Code Review: http://gerrit.cloudera.org:8080/6359
• Assignee: Adar Dembo
• Reporter: Adar Dembo
• Created: 12/Mar/17 20:35
• Updated: 13/Mar/17 23:33
• Resolved: 13/Mar/17 23:33

Description

The file block manager interacts with RandomAccessFiles via the file cache, and the file cache RandomAccessFile descriptor's implementation of memory_footprint() can lead to a crash, due to its use of std::make_shared() to allocate the underlying memory.

Just how likely the crash is to occur is unclear: it happens with near certainty using glibc's malloc, but I haven't seen it happen with tcmalloc. Because of that, I'm inclined to restrict the fix to 1.4 rather than get it into 1.3 (or even 1.2.1).

Issue Links

• is related to: KUDU-1932 Run at least one tablet-level test against all block managers