0

We've got a Postgres database with a table that takes heavy select, update and inserts that needs some tuning (it has too many indexes).

We have a read replica, and I'm wondering if moving all of our SELECT queries over to the replica will be a quick win while we sort out the indexes performance. (Our application can be switched over to a replica fairly easily while sorting the indexes is going to take longer).

My understanding is that pointing read operations at the replica probably won't improve performance in this case, but wanted to check my understanding.

My understanding is that:

  1. INSERT and UPDATE operations will return once they're committed to memory (which is then written to the WAL asynchronously), so they'll be not much faster.
  2. The replica will stream the WAL and will have to update it's own copy of the tables and indexes. So SELECT operations on the replica will have to contend with the same IO and CPU demands that those updates impose on the primary, so will gain no performance.

If our app were simply streaming inserts, then I could see a performance improvement for the inserts, but in practice since the clients are performing SELECT/INSERT together, I can't see a real performance gain here in this use case?

Is this correct, or have I missed something?

2

My understanding is that:

INSERT and UPDATE operations will return once they're committed to memory (which is then written to the WAL asynchronously), so they'll be not much faster.

This sounds like a description of sychronous_commit=off, but that is not the default value for that setting, and you don't describe having changed it.

In any event, modifications can involve calculations and the consultation of any number of rows other than the ones being modified. Having to share CPU and disks with concurrent heavy selects could certainly mutually slow the modifications down. If all your updates are just by pk and setting columns to a literal, and all your inserts are just of a list of literals, and you don't have fk constraints, maybe that is not a concern.

The replica will stream the WAL and will have to update it's own copy of the tables and indexes. So SELECT operations on the replica will have to contend with the same IO and CPU demands that those updates impose on the primary, so will gain no performance.

Replaying the WAL might be less IO than generating it was. The replica doesn't have to verify fk constraints. If the update was HOT (heap only tuple) then the replica doesn't need to consult any indexes, while the original work probably had to consult at least one index in order to identify the tuple to update. And the replica doesn't need to do anything with tuples that the primary considered but then filtered out before modifying.

Accurately predicting how much benefit you might get by redirecting reads is probably at least as hard as just optimizing your indexes will be.

2
  • Thanks, we're using an ORM and not doing anything fancy for these updates. They're literals being written using the PK. So it sounds like the replica is going to be doing as much work
    – ChrisJ
    May 30 at 21:25
  • No, it probably won't, as the answer describes. May 31 at 2:44
1

It could improve performance on your INSERT operations (and thus improve performance where those operations occur in the application) depending on how your read replica is setup and if your INSERT contention is partly due to locking caused by heavy reads against your primary database.

I'm not an expert on PostgreSQL replicas in practice, but the theory is applicable, that if your read replica is asynchronously synchronized from your primary, then any SELECT queries executed on your read replica shouldn't cause locking against your primary database (like they normally would if they were executed on the primary). It sounds like it is synchronized asynchronously if it's streaming the WAL which is being asynchronously written to. If that's the case, and if your database was experiencing heavy read lock contention, it's possible you'll see INSERT performance gains on your primary database by moving your SELECT queries to the read replica.

As far as the performance of the SELECT queries goes, I believe you're likely correct, that you won't necessarily see any performance gains since they'll still be bound by the same CPU, I/O, and write lock contention as the read replica is updated.

If it is easy enough to test, I'd say it's worth trying, if you got nothing to lose.

1
  • 1
    Yeah, locking is a good point. Several of the selects are going to LOCK FOR UPDATE which means the ORM will send them to the primary, so the locking might actually improve performance by splitting across the two servers.
    – ChrisJ
    May 30 at 21:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.