Comment by bshipp | Hacker News Reader

Comment by bshipp | Hacker News Reader

I administer a medium-sized postgres db, around (now) 110GB of data, and recently had the opportunity to revisit the old .conf file when I redeployed that database to a different server. What I quickly realized was that years of tweaks I had implemented as I learned postgres from various (now outdated) online sources were making it difficult to optimize my new instance.

I decided to start again from scratch. I discovered a handy little utility called pghero and set up it's internal monitoring:

https://github.com/ankane/pghero

I backed up my old .conf replaced it with a default and blew away all my old indices, and fired up the database. Of course, nothing worked well. But step by step, working with postgres' internal tools and pghero, I took a brand new look at every previous assumption and was much more strategic about where and when I created an index and which tuning parameter I tweaked.

Long story short, I was able to knock 30 GB off of the install disk size (based, in large part, on unnecessary duplicate indices), accelerate the most common queries, and everything worked a lot smoother. Plus my postgres installation was no longer fighting old postgres .conf tuning tweaks which were designed for much older versions of postgres.

Just adding my vote for revising your old tuning parameters after your database (and your programmer/administrator) has matured a bit.

This is a common operation for my team as well. In my old project, so many indexes that were added were just assumptions about the usage pattern of the data. Not only did they blow up the size of the table and INSERT time, they became the opposite of a red-herring where developers would stop and look elsewhere because the table "already has indexes".

Only adding indexes with good use of EXPLAIN/EXPLAIN ANALYZE on expected usage patterns yielded good results. Often we'd just blow the indexes away and fiddle until we added the correct one to enable an index scan / index only scan.

Other big boons for us have been using ENUM types where necessary (small known dataset for a column) -- now your column takes 4 bytes instead of N bytes for a string. I find them a bit easier to work in than foreign keys for this optimization because of their direct string mapping.

We've had really good mileage with Datadog's APM tools, RDS performance insights and pghero as you mentioned.

The ENUM type sounds like a great little tweak; I've got quite a few small tables that describe a fixed number of rows so I'll definitely investigate that.

I'm my own worst enemy for indexes, because I'm not only the database administrator but also the analyst. So when I'm trying to solve an ad-hoc problem and encounter a long-running query there's a 99% probability that I may have "accidentally" generated a new index to accelerate that query and then forgotten to remove it when I was done.

Lesson learned, and also a good argument for disaggregating the administrative and analytical user permissions.

More recently I’ve got into the habit of creating a new schema when I’m analysing / patching data so I can pull out the data I need into new tables and index accordingly. Because I do it from vim and have a plugin to execute the queries, I just create a record as I go (and I set the path to hit my new schema first so any new tables go there).

Ooooh, I like that idea a lot, especially since I can assign it a tablespace on a different drive.

They were useful. Unfortunately enums can't be changed (e.g. add a value) in a transaction without some extra work

> Other big boons for us have been using ENUM types where necessary (small known dataset for a column) -- now your column takes 4 bytes instead of N bytes for a string. I find them a bit easier to work in than foreign keys for this optimization because of their direct string mapping.

Can you give an example?

It sounds like normalising the schema would have similar (perhaps more) benefits.

Postgres keeps ENUM types in memory at all times. A join table is probably a slower option, all other things being equal. There are cases where it's worthwhile (dynamically expanding the values that would be in that enum, for example), but it's a tradeoff.

Normalizing has similar size benefits. For our use case (big aggregate reporting tables), incoming external data is not normalized, meaning we'd have to have another process iterate over it with foreign table in memory and map it to the foreign key. Enum's can be transparently ingested without this normalization requirement, while taking equivalent space. There's also now one level of indirection between the column and its value. Both are fine.

PGHero is great, and as someone who's reported (minor) issues and made feature requests on the issue tracker, the author is quite open to improvements too.

just wondering, did you have to plan a maintenance downtime for doing all that or could it be done without interrupting the production DB?

Not OP, but I imagine you could just duplicate the DB into a test environment to mess around with it.

That's exactly what it was. I had a bit of time after buying the server before I wanted to deploy it in production so I copied the last backup over and just started messing around. Because this database primarily ingests data from outside sources and doesn't drive user experiences, I was able to easily duplicate the data flow to replicate the production environment.

I recognize this would be a much different beast when it comes to a database that supports a hundred thousand users or something. However, Postgresql does have the ability to do hot standby and streaming replication so it's likely not as difficult as one might imagine.

https://www.postgresql.org/docs/current/warm-standby.html

Reply on news.ycombinator.com