Re: more detailed description of tup_returned and tup_fetched

On 2021/05/18 13:20, Masahiro Ikeda wrote:
> 
> 
> On 2021/05/17 20:46, Fujii Masao wrote:
>>
>>
>> On 2021/05/17 18:58, Masahiro Ikeda wrote:
>>>
>>>
>>> On 2021/05/17 15:32, Fujii Masao wrote:
>>>>
>>>>
>>>> On 2021/05/14 17:00, Masahiro Ikeda wrote:
>>>>> Hi,
>>>>>
>>>>> I worried the difference between "tup_returned" and "tup_fetched" in
>>>>> pg_stat_database. I assumed that "tup_returned" means the number of tuples
>>>>> that returned to clients. Of course, this is wrong.
>>>>
>>>> -       Number of rows returned by queries in this database
>>>> +       Number of live rows returned by sequential scans of queries in this
>>>> database
>>>>
>>>> -       Number of rows fetched by queries in this database
>>>> +       Number of live rows fetched by index scan of queries in this database
>>>>
>>>> I found the following comments in pgstat.h. So maybe even these
>>>> new descriptions are incorrect?
>>>>
>>>>    * Note: for a table, tuples_returned is the number of tuples successfully
>>>>    * fetched by heap_getnext, while tuples_fetched is the number of tuples
>>>>    * successfully fetched by heap_fetch under the control of bitmap indexscans.
>>>>    * For an index, tuples_returned is the number of index entries returned by
>>>>    * the index AM, while tuples_fetched is the number of tuples successfully
>>>>    * fetched by heap_fetch under the control of simple indexscans for this
>>>> index.
>>>
>>> Oh, Thanks!
>>>
>>> I updated the sentences using the descriptions of
>>> "pg_stat_all_tables.seq_tup_read", "pg_stat_all_tables.idx_tup_fetch", and
>>> "pg_stat_all_index.idx_tup_read".
>>>
>>> -       Number of rows returned by queries in this database
>>> +       Number of rows returned by queries in this database. The rows
>>> correspond to the live rows fetched by sequential scans and index entries
>>> returned by scans on indexes
>>
>> This is still not correct because this counter is incremented even when
>> other scan like TidScan happens?
> 
> Sorry, I couldn't find the way to increment tup_returned by TidScan.
> Do you mean that Tid Range Scan increments the counter?

Yes, what I tried to mean is Tid Range Scan.

> 
> Tid Range Scan increments the tup_returned, and
> pg_stat_all_tables.seq_tup_read is also incremented. I thought it's ok because
> Tid Range Scan is like sequential scan.

Yes, you're right. One interesting thing I found is;
when Tid Range Scan happens, seq_tup_read is incremented
but seq_scan is not. I'm not sure if this is expected behavior or not.

> That's the reason why the document of
> pg_stat_all_tables.seq_tup_read says "Number of live rows fetched by
> sequential scans"

Regarding the original issue, as far as I understand correctly,

* pg_stat_database.tup_returned = sum(pg_stat_all_tables.seq_tup_read) + sum(pg_stat_all_indexes.idx_tup_read)
* pg_stat_database.tup_fetched = sum(pg_stat_all_tables.idx_tup_fetch)

But the counters for some system catalogs like pg_database shared
across all databases of a cluster are excluded from that calculation.
Is this my understanding right? If right, probably we can reuse
the existing descriptions for those counters to document
pg_stat_database counters. For example,

pg_stat_database.tup_returned:
Number of live rows fetched by sequential and index scans in this database

pg_stat_database.tup_fetched:
Number of index entries returned by scans on indexes in this database

Regards,

-- 
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION