Re: pgstat_send_connstats() introduces unnecessary timestamp and UDP overhead

On Wed, Aug 25, 2021 at 01:20:03AM -0700, Andres Freund wrote:
> On 2021-08-25 12:51:58 +0900, Michael Paquier wrote:
> As I said before, this ship has long sailed:
> 
> typedef struct PgStat_MsgTabstat
> {
> 	PgStat_MsgHdr m_hdr;
> 	Oid			m_databaseid;
> 	int			m_nentries;
> 	int			m_xact_commit;
> 	int			m_xact_rollback;
> 	PgStat_Counter m_block_read_time;	/* times in microseconds */
> 	PgStat_Counter m_block_write_time;
> 	PgStat_TableEntry m_entry[PGSTAT_NUM_TABENTRIES];
> } PgStat_MsgTabstat;

Well, I kind of misread what you meant upthread then.
PgStat_MsgTabstat has a name a bit misleading, especially if you
assign connection stats to it.

>> As of the two problems discussed on this thread, aka the increased
>> number of UDP packages and the extra timestamp computations, it seems
>> to me that we had better combine the following ideas for HEAD and 14,
>> for now:
>> - Avoid the extra timestamp computation as proposed by Laurenz in [1]
>> - Throttle the frequency where the connection stat packages are sent,
>> as of [2].
> 
> I think in that case we'd have to do the bigger redesign and move "live"
> connection stats to backend_status.c...

Hmm.  A redesign is not really an option for 14 at this stage.  And I
am not really comfortable with the latest proposal from upthread to
plug in that to pgstat_send_tabstat to report things once per
transaction, either.  It really looks like this needs more thoughts,
and it would mean that a revert may be the most appropriate choice
for the moment.  That's the last-resort option, surely, but we are
post-beta3 so there is no much margin left.
--
Michael