[ASTERIXDB-1160] Possible deadlock between LogManager and LogFlusher

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Component/s: TX - Transactions
• Assignee: Murtadha Makki Al Hubail
• Reporter: Murtadha Makki Al Hubail
• Created: 05/Nov/15 04:20
• Updated: 10/Nov/15 01:37
• Resolved: 10/Nov/15 01:37

Description

Currently when LogFlusher thread flushes a full log tail buffer, it calls PrimaryIndexOpeartionTracker.completeOperation. If the completeOperation call results in generating a FLUSH log type and the LogManager has no empty log buffers available to log it, a deadlock happens.

To reproduce the deadlock with high probabilty, set the number of log pages in the asterix configuration to 1 and a small size, then load some data to the system.

The proposed solution is to have a different thread responsible for logging FLUSH logs. This way logging FLUSH logs will have to wait until an empty log buffer becomes available to the LogManager on a different thread than the LogFlusher.