[WIP] using db-pool library to create a pool of databases

[momentary-lapse]

Addresses: #4979

[momentary-lapse]

I created this WIP PR to share the progress and the issue I'm stuck with currently. The crate I use operates with its own structure wrapping connection pools: code
And we have our own ActualDbPool. They are kinda same, but it's not obvious for me how to correctly convert one to another.
I had an idea to make ActualDbPool a enum with two possible values: RegularPool and ReusablePool, but stuck on trying to adapt stuff like LemmyContext, which also requires pool struct to be clone-able (and ReusablePool is not). And it seems a lot of changes to the main codebase for purely test changes.
Do you folks have any ideas how to manage that? Or should I stick to the initial plan without using this library?

[dessalines]

Our ActualDbPool is just a type alias for deadpool Pool<AsyncPgConnection>.

Their crate should be able to work with deadpool pools, but I'm not familiar with how to plug that into their crate... you'll have to ask them.

[momentary-lapse]

Yeah, I see. I returned to this issue today after a week of a break. I'm in contact with the db-pool author and they're helping to understand a lot of moments and really willing to collaborate, so i think we'll make this work.

I'd like to clarify one moment: do we want build_db_pool_for_tests to return still ActualDbPool? db-pool has its own wrapper ReusableConnectionPool which works like a deadpool Pool, but a bit different and needs adaptation. And it might be easier to adapt tests for working with ReusableConnectionPool than converting ReusableConnectionPool to ActualDbPool

[dullbananas]

The return type of build_db_pool_for_tests may be changed. Also, a DbPool variant may be added if needed.

[momentary-lapse]

This is the moment which currently blocks me, and i think it's better to consult with you again. LemmyContext structure must be cloneable, therefore all the fields, therefore the pool. But unfortunately, reusable pool from db-pool crate is not, and i don't have access to its fields to implement the trait here.
But before asking db-pool developer, i'd like to be sure we really need this pool cloning stuff, especially for the tests. Cloning the pool seems a bit strange to me, but i may miss something. I'm looking at the code now, but maybe you folks already have some insights on this

[dullbananas]

Wrap it in Arc for now.

[momentary-lapse]

Update: I'm working on the topic; cannot devote much time for it, but it slowly going forward, and i keep the code in the branch up-to-date. I connected db-pool crate to our tests, and reworked most of them. Currently have a runtime error, planning to look at it and fix this week.
After this, what is left is to change a few tests which are using build_db_pool function.

[momentary-lapse]

Hey! I've a question about the replaceable schema — found out it's the reason tests are failing in the branch: the tests are being run in restricted privilege mode, and have no access to the r schema. I can do some workaround, but would like to ask about motivation for making that replaceable schema — for me it seems a bit unsafe, since there in theory might be migrations which require changes in these files. Also, am I right that the triggers and utils are separated into another schema because it's just easier to clean them up at once?

[dessalines]

Yes I believe the main reason is so that they're easier to clean up.

for me it seems a bit unsafe, since there in theory might be migrations which require changes in these files.

Everything in the r schema should only contain triggers and functions used at runtime, not by any migrations. They're only run once, at the end of all the migrations. The main reason to do it like this, is because it was extremely cumbersome to tweak every trigger and function within the migrations, copying and pasting every time, rather than doing tweaks to a simple sql file.

the tests are being run in restricted privilege mode

Is this a limitation of the db-pool library? Because regular cargo tests can already run the build_db_pool_for_tests function which runs the replaceable schema.

[momentary-lapse]

Is this a limitation of the db-pool library?

Yep, it's designed the way you have privileged connection pool to prepare dbs and do migrations, and restricted connection pool to run tests. And the latest can only access public schema of each db. So, the library would work out of the box if we had these runtime functions in the same schema

[dessalines]

Maybe its just a matter of running GRANT for the current lemmy user on that schema, as part of the setup?

https://stackoverflow.com/questions/17338621/what-does-grant-usage-on-schema-do-exactly

cc @dullbananas

[momentary-lapse]

that's an option i'm thinking about. but unfortunately it has certain caveats.
what does the db-pool library do, it prepares multiple databases with names like:

• db-pool-uuid1
• db-pool-uuid2
  etc.
  and it creates a separate role with restricted rights for each db named exactly as this db. so to grant permissions after the migrations are done here, i need to know the current db name, which is afaik impossible to retrieve from the connection only.

however, i can change it in a fork, so it will use the same role for all the restricted connections (e.g. _db_pool_test_role). then we can grant access to schema r for that user after migrations are done

[dullbananas]

Maybe its just a matter of running GRANT for the current lemmy user on that schema, as part of the setup?

Maybe. The problem is almost certainly caused by something being different between the "public" schema and the "r" schema. Otherwise migrations would have failed before replaceable_schema runs.

[momentary-lapse]

There's logic in the db-pool library which grants permission to restricted user fro public schema assuming it's the only one tests will deal with.
Anyway, i reported that, and we agreed he'll take time to think how to change that. So, we're in touch. And if he decides to keep the logic, i have some ideas for a workaround for our case