[IGNITE-7998] SQL: Improve MVCC vacuum performance by iterating over data pages instead of cache tree.

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Component/s: mvcc
• Assignee: Unassigned
• Reporter: Roman Kondakov
• Created: 20/Mar/18 13:14
• Updated: 27/Feb/19 15:08
• Resolved: 27/Feb/19 15:08

Description

At the moment vacuum process uses cache trees to find outdated (dead) entries and cache and index trees to cleanup them. It is not efficient due to several reasons. For example, we should lock a datapage for each cache tree entry to find out if entry is dead.

We can consider a direct iteration over datapages as a possible improvement of  the vacuum process. Data page iteration prototype demonstrated 5-10 times time improvement over the tree iteration.

At first stage we need to implement direct datapages iteration only for collecting dead entries links.

At the second stage we need to consider removing links to dead entries from index pages directly. In other words, we need to efficiently remove batches of dead links from indexes without traversing cache and index tree one dead link by one.

Issue Links

• duplicates: IGNITE-10729 MVCC TX: Improve VAC using visibility maps
• is related to: IGNITE-5936 Cleanup of not needed and committed versions