To put it bluntly after managing both for years, Postgres doesn't scare the shit out of me like MySQL does. I've had quite a few moments with MySQL doing stupid things that aren't intuitive or right, particularly in the backup/restore space. For example backup one schema and restore onto later versions doesn't always work.
Postgres has the same feel as *BSD i.e. deterministic. It is well documented and does exactly what the manual says and is devoid of surprises. It feels engineered and I can provide reliably engineered solutions because of this.
Basically I can sleep at night.
I really give less of a crap about scalability. I'd throw a bigger box at the problem or use heavy caching. I've built much larger ecommerce solutions with orders of magnitude more hits/orders than sparkfun on much smaller kit
I have been woken up in the night multiple times and found that the problem is that Postgres has arbitrarily decided to stop using an index and changed to do a full table scan and the database has ground to a halt under the load.
Doing an ANALYZE sometimes fixes it. Increasing the statistics target for the table in addition sometimes fixes it. One time we had to add an extra column to an index to make it decide to use it even though it shouldn't have needed it (and didn't a few hours before!).
The developers seem dead set against allowing you to override the query planner to add determinism.
I'm not saying MySQL is any better, as I have not used that in production before.
That's crazy. How would you rate a shell that just said nothing and copied nothing when asked to copy an important backup file somewhere? For me it is the hell of software, not even telling there is something wrong.
The vessel is in danger when the captain don't know her position, but in much greater danger when he is not even knowing he doesn't know where she is.
That said, it's unfair to compare with MySQL's planner, which, although it is there, is so infantile that by comparison you can say MySQL has no planner. Example: Invariant sub selects run on every row of the master query, instead of being stored in a temporary table. Think SELECT a FROM foo WHERE foo.b IN (SELECT b FROM bar). Or another basic one SELECT a FROM foo ORDER BY b LIMIT 5 where there exists an index on b actually sorts the entire table instead of using the index and jumping out when it hits the LIMIT.
Query planner changing its mind can be a problem for most DBs, Postgres is not special in this regard.
I use MSSQL a lot in my day job and sometimes the query planner there surprises us by being fairly bright about this and other optimisations. Of course it sometimes gets things very wrong and you have to use index hints to push it in the right direction, but generally speaking if your queries are complex enough to hit that problem you might need to consider refactoring your design for other reasons too.
With Postgres, while it usually seems to do a better job, when it refuses to use your indexes it can be a massive pain to coax it into doing what you want it to, even when you know it's making a bad choice.
Postgres has a number of issues like that, where ideological purity has a negative impact on usability. It still mainly use Postgres these days, but I regularly swear at it over things I know would be far simple than MySQL for those reasons (replication used to be a huge one, though that is now finally getting there, 8 years after I set up my first "large" replicated MySQL setup)
The query planner evaluates multiple query plans, and chooses the one with the lowest cost based. If that results in the wrong plan, you should attempt to correct the query planner input (that is, the statistics and the server configuration, and possibly the tables and their indices) so that it can better evaluate the possible query plans.
EXPLAIN is a great tool to debug query plans.
You mean hacks like these?
http://docs.python.org/2/library/abc.html
I agree.
From my experience with PostgreSQL and MySQL, it is a bit like the difference between python and php, one is a great work of craftmanship that is reliable, predictible, coherent, enjoyable to work with and minimise the wtf/mn rate (which is the best quality measure in software), while the other is a bunch of hacks knit together to make it work asap and its wtf/mn is skyscraping.