On 9/22/25 15:57, Andrei Lepikhov wrote:
On 22/9/2025 15:37, Frédéric Yhuel wrote:
I wonder if this is an argument in favour of decoupling the sample
size and the precision of the statistics. Here, we basically want the
sample size to be as big as the table in order to include the few
(NULL, WARNING) values.
I also have seen how repeating ANALYZE on the same database drastically
changes query plans ;(.
It seems to me that with massive samples, many of the ANALYZE algorithms
should be rewritten. In principle, statistical hooks exist. So, it is
possible to invent an independent table analyser which will scan the
whole table to get precise statistics.
Interesting! I wonder how difficult it would be.
However, in this specific case, I realised that it wouldn't solve the
issue of ANALYZE being triggered when there are zero rows with (ackid,
crit) = (NULL, WARNING).
Partitioning would still work in this case, though, because ackid's
null_frac would be zero for the partition containing the 'WARNING' value.
I wonder if we could devise another kind of extended statistic that
would provide these "partitioned statistics" without actually partitioning.
