Re: Performance degradation of REFRESH MATERIALIZED VIEW

Hi,

On 2021-03-11 17:44:37 +0900, Masahiko Sawada wrote:
> The execution time of REFRESH MATERIALIZED VIEW are:
>
> w/ HEAP_INSERT_FROZEN flag : 42 sec
> w/o HEAP_INSERT_FROZEN flag : 33 sec
>
> After investigation, I found that such performance degradation happens
> on only HEAD code. It seems to me that commit 39b66a91b (and
> 7db0cd2145) is relevant that has heap_insert() set VM bits and
> PD_ALL_VISIBLE if HEAP_INSERT_FROZEN is specified (so CCing Tomas
> Vondra and authors). Since heap_insert() sets PD_ALL_VISIBLE to the
> page when inserting a tuple for the first time on the page (around
> L2133 in heapam.c), every subsequent heap_insert() on the page reads
> and pins a VM buffer (see RelationGetBufferForTuple()). Reading and
> pinning a VM buffer for every insertion is a very high cost. This
> doesn't happen in heap_multi_insert() since it sets VM buffer after
> filling the heap page with tuples. Therefore, there is no such
> performance degradation between COPY and COPY FREEZE if they use
> heap_multi_insert() (i.g., CIM_MULTI). Paul also reported it in that
> thread.

Probably worth adding as an open item for 14.

Greetings,

Andres Freund