On 25/03/2019 15:21, Heikki Linnakangas wrote:
> I had another quick look.
>
> I still think using the "generic xlog AM" for this is a wrong level of
> abstraction, and we should use the XLOG_FPI records for this directly.
> We can extend XLOG_FPI so that it can store multiple pages in a single
> record, if it doesn't already handle it.
>
> Another counter-point to using the generic xlog record is that you're
> currently doing unnecessary two memcpy's of all pages in the index, in
> GenericXLogRegisterBuffer() and GenericXLogFinish(). That's not free.
>
> I guess the generic_log_relation() function can stay where it is, but it
> should use XLogRegisterBuffer() and XLogInsert() directly.
Patch set v.3 uses XLOG_FPI records directly.
As a benchmark I use the script (test.sql in attachment) which show WAL
size increment during index build. In the table below you can see the
influence of the patch on WAL growth.
Results
=======
AM | master | patch |
GIN | 347 MB | 66 MB |
GiST | 157 MB | 43 MB |
SP-GiST | 119 MB | 38 MB |
--
Andrey Lepikhov
Postgres Professional
https://postgrespro.com
The Russian Postgres Company
