-
Type: Improvement
-
Resolution: Unresolved
-
Priority: Major - P3
-
None
-
Affects Version/s: None
-
Component/s: None
-
StorEng - Defined Pipeline
We have seen some recent evidence that WiredTiger is inefficiently laying data out on leaf pages. That is managed by the reconciliation split logic - we should review that logic looking for places where it might make poor decisions.
Note: It's possible that the poor page layout is due to access patterns not reconciliation inefficiency. If data in a table is populated then a majority of data is selectively removed it could lead to a similar pattern of small leaf pages.
The case in hand has pages with an average of 600 bytes of data, when leaf_page_max is configured to be 16KB. Durable history is enabled. It is on an index in MongoDB, which probably has small key/value pairs being randomly updated across the data, along with 5 minutes of pinned history.
- has to be done after
-
WT-10831 Improve cache tracking for small uncompressed pages
- Open
-
WT-10845 Add statistics that give insight to cached disk image size
- Open
-
WT-10846 Create a tool that can anonymously dump table/database shape
- Closed
- has to be done before
-
WT-10833 Implement a mechanism to combine small on-disk pages together
- Open
- is related to
-
WT-9182 Explore what should be the correct way to calculate upd_memsize in the durable history era
- Open