|From:||Смирнов Денис <sd(at)arenadata(dot)io>|
|Subject:||PoC Refactor AM analyse API|
|Views:||Raw Message | Whole Thread | Download mbox | Resend email|
I suggest a refactoring of analyze AM API as it is too much heap specific at the moment. The problem was inspired by Greenplum’s analyze improvement for append-optimized row and column AM with variable size compressed blocks.
Currently we do analyze in two steps.
1. Sample fix size blocks with algorithm S from Knuth (BlockSampler function)
2. Collect tuples into reservoir with algorithm Z from Vitter.
So this doesn’t work for AMs using variable sized physical blocks for example. They need weight random sampling (WRS) algorithms like A-Chao or logical blocks to follow S-Knuth (and have a problem with RelationGetNumberOfBlocks() estimating a physical number of blocks). Another problem with columns - they are not passed to analyze begin scan and can’t benefit from column storage at ANALYZE TABLE (COL).
The suggestion is to replace table_scan_analyze_next_block() and table_scan_analyze_next_tuple() with a single function: table_acquire_sample_rows(). The AM implementation of table_acquire_sample_rows() can use the BlockSampler functions if it wants to, but if the AM is not block-oriented, it could do something else. This suggestion also passes VacAttrStats to table_acquire_sample_rows() for column-oriented AMs and removes PROGRESS_ANALYZE_BLOCKS_TOTAL and PROGRESS_ANALYZE_BLOCKS_DONE definitions as not all AMs can be block-oriented.
Denis Smirnov | Developer
Arenadata | Godovikova 9-17, Moscow 129085 Russia
|Next Message||Bharath Rupireddy||2020-12-07 13:34:24||Re: Parallel Inserts in CREATE TABLE AS|
|Previous Message||Daniel Gustafsson||2020-12-07 13:15:58||Re: Refactor MD5 implementations and switch to EVP for OpenSSL|