Hacker News Reader

Comment by tbrock | Hacker News Reader

Inlining CTEs is a big deal but they also finally have a way to see the progress of an index build without a PhD in postgresology!

Check it out here: https://www.postgresql.org/docs/12/progress-reporting.html

View on HN

I've found CTEs to be one of the few ways to reliably force Postgres evaluation order, which is sometimes quite poor (I have specific queries that run 100x faster with CTE), so retaining old behaviour with materialized is something I didn't know about in 12.

How this should be done was the big roadblock, the actual inlining was a relatively simple patch.

My #1 request for indexes from PostgreSQL or 3rd party tool is concurrently rebuilding indexes.

I have had several cases where query performance dropped significantly (unfortunately in managed AWS Aurora DB, so I have limited inspection abilities).

Index rebuilds is the solution but it locks the entire table.

You can get a concurrent rebuild by creating a new index concurrently, then transactionally drop the old one and rename the new one. It's error prone though.

> You can get a concurrent rebuild by creating a new index concurrently, then transactionally drop the old one and rename the new one. It's error prone though.

actually you can't since you can't create/drop indexes concurrently inside a transaction.

btw. my #1 request would be a better materialized view (which supports indexes) and also makes the performane of some stuff really better. unfortunatly you can't refresh it without a full seq scan. it would be great if postgres would have an auto updating materialized view.

You can certainly drop, or at least rename, inside a transacion. Only the rename part needs to transactional.

> it would be great if postgres would have an auto updating materialized view.

100%. Check out Incremental View Maintainence https://wiki.postgresql.org/wiki/Incremental_View_Maintenanc...

There's been some work done on it, but it's a pretty difficult problem. If/when it comes to PostgreSQL it'll come for only very simple queries.

I wouldn't recommend dropping large indexes within a transaction. Use DROP INDEX CONCURRENTLY instead, Without CONCURRENTLY it has to lock the table which can cause issues in high traffic databases.

But this is mostly a moot point since 12 has REINDEX CONCURRENTLY.

> it has to lock the table which can cause issues in high traffic databases

In my case, I'm okay with a seconds-long lock while an index is dropped. An hours-long lock while an index is created is the real show stopper.

> But this is mostly a moot point since 12 has REINDEX CONCURRENTLY.

Fantastic.

Thanks for the interesting link on IVM - I hadn’t heard about that work. Crikey, it’s a humbling one to solve generally reading through that page.

I have a case of this where I do the incremental updates myself. In my case, figuring out which rows have changed isn’t that tough, I’ve got time stamps to lean on that simplify the process. More complex is that even plucking the data for a single row is slow if you’re not careful. I have a number of large subqueries that Postgres doesn’t optimise well if I just restrict the rows once in the outer query. Instead I need to be very specific with my restrictions in a number of duplicated locations throughout the query to get fast incremental updates. It works well now it’s up and running but it’s a little clumsy to set up.

It gets much easier if you are willing to perform "extra" query work to update the view.

IVM is taking a sophisticated approach that results in a minimum of required work, even in cases of (some) aggregations.

PostgreSQL 12 has REINDEX CONCURRENTLY.

You will need to pick a new #1 request since PostgreSQL 12 finally delivered REINDEX CONCURRENTLY. :)

!!!!!

Wonderful! (Now I only have to wait 12 months for AWS Aurora to upgrade.)

This is worth a lot - thanks for the pointer!

Reply on news.ycombinator.com