Conversation

I just wonder what happens if multiple keys with the same value are deduplicated and then all the others except one key are deleted. If each entry is not marked as LP_DEAD, unncessary lookup for the table can happen?

Quote Tweet

Fujii Masao

@fujii_masao

Oct 22, 2020

Regarding B-tree deduplication, LP_DEAD flag is set in the posting list only when all the entries in the list are known as dead? Or each entry can be marked as LP_DEAD? #PostgreSQL

Replying to

Yes, it can, but it's not too bad in practice. Remember that deduplication is lazy, and only happens when the alternative is a page split. We must have not have had any LP_DEAD bits set when any dedup pass happens (if we did there would be no dedup pass at all).

Replying to

and

Relying on opportunistic LP_DEAD setting is risky. New pg 14 work on "delete deduplication" doesn't care about LP_DEAD bits being by somebody else already. It works really well for non-hot updates that cause duplicate version churn: postgr.es/m/CAH2-WzmUAf=

5:00 PM · Oct 22, 2020Twitter Web App

Retweet