Re: more detailed description of tup_returned and tup_fetched

On 2021/05/20 9:46, Masahiro Ikeda wrote:
> 
> 
> On 2021/05/18 20:10, Fujii Masao wrote:
>>
>>
>> On 2021/05/18 18:23, Masahiro Ikeda wrote:
>>>
>>>
>>> On 2021/05/18 16:01, Fujii Masao wrote:
>>>> On 2021/05/18 13:20, Masahiro Ikeda wrote:
>>>>> 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.
>>>
>>> The following comment says that this behavior is expected. But, I agree it's
>>> odd and it's natural both seq_tup_read and seq_scan are incremented at the
>>> same time or not...
>>>
>>> /*
>>>    * Currently, we only have a stats counter for sequential heap scans (but
>>>    * e.g for bitmap scans the underlying bitmap index scans will be counted,
>>>    * and for sample scans we update stats for tuple fetches).
>>>    */
>>> if (scan->rs_base.rs_flags & SO_TYPE_SEQSCAN)
>>>      pgstat_count_heap_scan(scan->rs_base.rs_rd);
>>>
>>>
>>>>> 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,
>>>
>>> Yes, my understanding is same now.
>>>
>>>
>>>> pg_stat_database.tup_returned:> Number of live rows fetched by sequential
>>>> and index scans in this database
>>>
>>> I wonder "live rows fetched by index scans" may mislead. I think "live" means
>>> it's not dead tuple and "rows" mean the tuple user want to get.
>>>
>>> But, pg_stat_all_indexes.idx_tup_read says that "index entires returned by
>>> scans on this index". There is no meaning of "live" and "rows", so I thought
>>> it's better to distinguish them.
>>>
>>> So, why don't you change to "Number of live rows fetched by sequential scans
>>> and index entries returned by index scans in this database"?
>>
>> Yes, LGTM.
>>
>>
>>>> pg_stat_database.tup_fetched:
>>>> Number of index entries returned by scans on indexes in this database
>>> Is this the sum of pg_stat_all_indexes.idx_tup_read? This is accounted to
>>> pg_stat_database.tup_returned.
>>
>> I was thinking that pg_stat_database.tup_fetched is the same as
>> the sum of pg_stat_all_tables.idx_tup_fetch. Because they both
>> are incremented by bitmap index scans, but pg_stat_all_indexes.idx_tup_read
>> is not.
> 
> Yes. So, "Number of index entries returned by scans on indexes in this
> database" is incorrect, and "Number of live rows fetched by index scans in
> this database" is correct?

Yes, I think so!

Regards,

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