I don't know much about Postgres optimizer, but I imagine that in addition to table/column stats, is also uses structural info as its inputs, like existence of (enabled & valid) constraints for example. If the optimizer knows that some column never has NULLs or is guaranteed to be unique, all kinds of transformation & shortcuts become possible.

(There are plenty of large big-vendor ERP/CRM/etc apps out there that do not use DB constraints for the sake of "portability"... not fun to work with these).

Here's one that surprised me when I found out about it years ago, because I'd never really given it thought: There's a correlation statistic on columns for how well the values in that column match the row order on disk, which can influence a few different things.

In my case a query that retrieved a ton of data with an ORDER BY was using a sort and taking like two hours to run (a data source for an ETL process) - turned out because of a really bad correlation postgres was refusing to use the index, because the random access would be even slower, so it did a table scan then sort. After figuring this out and discovering the CLUSTER command (reorders the data on disk to match an index), it did an index scan and didn't need to sort at the end, was able to start streaming results immediately, and finished the entire query in like ten minutes.

Just a nice example of where the obvious "use query hints to make it use the index" would have been the worst option, instead figuring out why postgres didn't want to use it and fixing that resulted in something much better.

I come to the opposite conclusion. Clustering a table results in an access exclusive lock, and due to MVCC the ordering isn’t permanent.

Here, you as the engineer know you’d like to use a sorted index to stream results out even if the overall query end to end is slower due to the I/O cost. In my opinion there should be a way to express this within the query.

Streaming results doesn't mean anything if the total runtime is that much worse, it just means the overall system will take hours longer to complete. And the good version doesn't need to continuously run CLUSTER, the correlation just has to be high enough for it to be the better choice, so we settled on running it once a week - only takes like 2 minutes to run the CLUSTER.

You should be thinking of it the other way around: my final 10 minute result was the ideal situation I wanted, but when circumstances were bad for it, the postgres query planner was smart enough to tell it wouldn't work and switch to the 2-hour plan instead of blindly following the original plan and taking 5 hours.

This is very true. PostgreSQL does not do any proactive plan caching, so it's important that the planner remains fast. It is possible to adjust the number of stats targets to control the size of the histograms and most common values list. Upping that can be useful for OLAP-type workloads.

> I imagine that in addition to table/column stats, is also uses structural info as its inputs, like existence of (enabled & valid) constraints for example.

Yes. Foreign key constraints are used to assist with join selectivity estimations. PG17 (when released) should be able to make more use of NOT NULL constraints to improve plans.

I would prefer having "externally attached" hints for a query (e.g. identified by it's queryid) like Oracle's stored outlines.

It's also important to consider that with enable_nestloop=0, when Nested Loop must be used (e.g for a CROSS JOIN) that the cost penalty that's added to reduce the chances of Nested Loop being used can dilute the costs so much that the query planner can then go on to make poor subsequent choices later in planning due to the costs for each method of implementing the subsequent operation being so relatively close to each other than the slightly cheaper one might not even be considered. See add_path() and STD_FUZZ_FACTOR. So, running enable_nestloop=0 in production is not without risk.

I wondered whether it would be possible just to add a fixed fuzz to every row estimate, say five rows. It would essentially mean you can never get this issue of a small undercount causing a plan disaster. Overestimating slightly is basically never a big issue as far as I know.

(I should perhaps have considered this when I was actually making a query planner in a previous life, but there were more than enough other things to worry about :-) )

Another avenue is of course trying to avoid the issue to begin with, e.g. through the recent “translation grids” of Müller and Moerkotte for better join selectivities. But I doubt anyone is going to be finding a silver bullet for this anytime soon, so reducing plan risk somehow seems very worthwhile.

Yeah, I think it was probably a mistake to always assume there's zero correlation between columns, but what value is better to use as a default? At least extended statistics allows the correlations of multiple columns to be gathered now. That probably means we'd be less likely to reconsider changing the default assumption of zero correlation when multiplying selectivities.

Multi-selectivities are good, but IIRC they can't be specified across tables and thus across joins, right? Out of curiosity; how do you reconcile multiple selectivities? If you have a multi-column histogram on (a,b) and one on (b,c) and one on (c), and you need to figure out the selectivity of WHERE a=? AND b=? AND c=? from those three? I looked into this at some point, and academia presented me with a nightmare of second-order cone programming and stuff. :-) (I never implemented any of it before leaving the database world.)

Yeah, no extended statistics for join quals yet.

> how do you reconcile multiple selectivities?

Looking at https://doxygen.postgresql.org/extended__stats_8c.html#a3f10... it seems the aim is to find the stats that cover the largest number of clauses tiebreaking on the statistics with the least number of keys. For your example both of those are the same, so it seems that which stats are applied is down to the order the stats appear in the stats list. That list is ordered by OID, which does not seem ideal as a dump and restore could result in the stats getting a different OID. Seems sorting that list by statistics name might be better. That's what we do for triggers, which seems like a good idea as it gives the user some ability to control the trigger fire order.

[1] Similarly, you can sometimes guarantee stuff like P(a AND b) <= 1/n due to a unique multi-column index, which can also inform some of your sub-estimates.