Re: auto-vacuum & Negative "anl" Values

Alvaro Herrera <alvherre(at)commandprompt(dot)com> writes:
> Tom Lane wrote:
>> Apparently somehow last_anl_tuples has managed to get to be bigger than
>> n_live_tuples, which maybe could happen after a delete. Should we be
>> clamping last_anl_tuples to not exceed n_live_tuples somewhere?
>> Alvaro and Matthew, what do you think?

> Hmm ... I'd think that the number of dead tuples plus live tuples should
> never be smaller than the number of tuples seen at last analyze.

The scenario I was imagining was big delete followed by
VACUUM-without-ANALYZE. In this situation, it looks to me like
pgstat_recv_vacuum updates n_live_tuples to the new smaller value
and doesn't do anything to last_anl_tuples. I'm thinking we need

tabentry->n_live_tuples = msg->m_tuples;
tabentry->n_dead_tuples = 0;
if (msg->m_analyze)
{
tabentry->last_anl_tuples = msg->m_tuples;
if (msg->m_autovacuum)
tabentry->autovac_analyze_timestamp = msg->m_vacuumtime;
else
tabentry->analyze_timestamp = msg->m_vacuumtime;
}
+ else
+ {
+ /* last_anl_tuples must never exceed n_live_tuples */
+ tabentry->last_anl_tuples = Min(tabentry->last_anl_tuples,
+ msg->m_tuples);
+ }
}

but perhaps I'm confused.

regards, tom lane