[IGNITE-6057] SQL: Full scan should be performed through data pages bypassing primary index

• Type: Task
• Status: Closed
• Priority: Major
• Resolution: Won't Fix
• Affects Version/s: 2.1
• Component/s: persistence, sql
• Labels: performance
• Assignee: Unassigned
• Reporter: Vladimir Ozerov
• Created: 14/Aug/17 14:20
• Updated: 05/Jan/19 12:12
• Resolved: 27/Nov/17 20:06

Description

Currently both SQL full scan and CREATE INDEX commands iterate through primary index to get all existing values. Consider that we have 10 entries per data page on average. In this case we will have to read the same data page 10 times when reaching relevant keys in different parts of index tree. This could be very inefficient on certain workloads.

We should iterate over data pages directly instead. This way a page with 10 entries will be accessed only once. However, we should take cache groups in count - if there are too many entries from other logical caches, this approach could make situation even worse, unless we have a mechanism to skip unnecessary entries (or the whole pages!) efficiently.

Probably we should develop a cost-based model, which will take in count the following statistics:
1) Average entry size. The longer the entry, the lesser the benefit. Especially if overflow pages are used frequently.
2) Cache groups. Ideally, we should estimate number of entries from all logical caches. The more entries from other caches, the lesser the benefit.

Issue Links

• causes: IGNITE-10798 Data page scan for ScanQuery, SqlQuery and SqlFieldsQuery