On Fri, Jul 7, 2017 at 4:41 PM, Justin Pryzby wrote:
> The second change averages separate correlation values of small lists of 300
> consecutive TIDs, rather than the course-granularity/aggregate correlation of
> a
> small sample of pages across the entire index. Postgres' existing sampling is
Months ago I reported an issue with very slow index scan of tables with high
"correlation" of its indexed column, due to (we concluded at the time)
duplicate/repeated values of that column causing many lseek()s. References:
https://www.postgresql.org/message-id/flat/20160524173914.GA11880%40telsa
> I don't think the downstream dependencies will let that work without
> rebuilding them as well. The drop fails (without a cascade), and the other
> views and matviews that are built off of this all simply point to x_old.
Wow, ouch. Yeah, I'd neglected to consider dependent objects. Your
only "