[HDDS-8485] SCM Transaction Buffer Delete block delay Investigate solution

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 1.4.0
• Target Version/s: 1.3.1
• Assignee: Sumit Agrawal
• Reporter: Sumit Agrawal
• Created: 25/Apr/23 02:04
• Updated: 14/Dec/23 13:43
• Resolved: 30/May/23 10:50

Description

Investigate the different solution for handling transaction buffer which is causing delete block delay till 1000 ratis transaction is done in SCM,

• SCM flush the db update after 1000 (default) ratis transaction when ratis takes snapshot.

This creates delay, and investicate different mechanism to handlle

 

3 type of solution as been investigated till now,

1. RocksDB BatchWriteWithIndex (HDDS-7483) - This provides similar performance as BatchWrite currently present, and allow to read the data from batchWrite buffer
2. Manual WAL flush for rocksDB (HDDS-8469) - This is alternative simpler approach, but current performance comparison is slower, 3-9 times for different scenario
3. Timer based solution (HDDS-7111)- Simpler approach flush on time based if not flushed by Ratis (With additional protection of ratis apply transaction parallel - not present currently), Performance will not be hitted

 

erose Nibiruxu Can have discuss and conclude, as per recent discussion, approach "3"seems to be simpler with few additional changes.

Sub-Tasks

• RocksDB BatchWriteWithIndex to handle SCM buffer transaction: Resolved
• SCMHADBTransactionBuffer not flushed in time: Resolved
• SCMHATransactionBuffer flush based on time: Resolved