Example: I wrote a recursive query that provided a huge reduction in queries issued from our app for a traversal. Worked beautifully in local, and halved p95 in our dogfood environment. Yippee! In prod? The query always timed out after 60 seconds, even though it had the same query plan as dogfood env and staging. Sad trombone noises.
For query semantics regression testing, we just write tests in our normal test framework and run them in CI like any other test. Test data setup works like any other test, and we can call some of our millions of lines of app code to help out with repetitive tasks. To prevent cross-test leakage we wrap each test in a BEGIN..ROLLBACK, and transform inner use of transactions to save points in our db client layer. I’d like to add libeatmydata to speed things up further but haven’t done so yet.
But from my experience, this is low-hanging fruit for catching regressions that could break your application without you noticing. The performance work you're describing is the real nitty-gritty database management - the deep optimization work that truly makes or breaks an app.
For that there's no silver bullet.
I think one interesting angle would be to integrate your assertions directly into Postgres as a plugin. That way, you can spy on every query issued in any test framework, dump the text automatically, and prove additional assertions about the query plan index use perf etc, regardless of the user’s primary language. That would also get you out of needing to build 1000 different integrations for all the programming language / ORM combos in the universe. If you don’t want to do a Postgres plugin you could also do this as a Postgres wire protocol proxy.
I'm wondering how I would be able to regression-test functions in my project (pgflow [0]) - it tracks a graph of state machines modeled in few tables. State is mutated only by calling few exposed SQL functions (task queue worker does it).
Given I can't enforce everything I need with check constraints and I try to avoid triggers if possible, I opted for only using exposed SQL API [1] for setting up state in my pgTAP tests.
It is imperative and harder to maintain, like scripts you described in the article, but really my only option, as I want to have maximum confidence level.
Does RegreSQL support some kind of init scripts or I would need to wire it myself and just run RegreSQL after the proper state is set? Would lose the "run once and get report on everything" benefit then :-(
[0] https://pgflow.dev/ [1] https://github.com/pgflow-dev/pgflow/blob/main/pkgs/core/sup...
Plus I have whole set of other requirements where RegreSQL suddenly seems to be a good solution.
And without sounding cliche - Thank you for the comment! This is exactly why I forced myself to go public and get this level of feedback.
I was postponing proper, dedicated performance testing for some time and would really love to up my game in that regard.
I'm very happy with pgTAP approach of running stuff in transaction and rolling them back after the test - how this works in RegreSQL?
Would love to provide feedback and test the hooks when you will be working on them. I'm mostly interested in performance testing and my use case would be to run them on CI and compare to previous metrics stored somewhere in order to fail CI when performance regressions are introduced.
Happy to connect, got contact info in my profile.
You can instruct postgres to raise exceptions using the same error code that constraints use: that way your clients do not need to know the difference.
That said, there ARE other places that would benefit from triggers (aggregate counts). I've avoided them because they're hot paths and I was worried about perf impact - relyi ng on pgTAP coverage instead.
Your defense-in-depth argument is solid though. I should revisit this and benchmark whether the safety is worth the perf cost. Something like RegreSQL would come in handy
So it would be nice if RegreSQL would support fixture hooks for those who like this route.
Would also be fun if you could support PGLite [0], that's what I've been using to write "unit" tests connected to a "real" database.
[0] https://github.com/eradman/ephemeralpg [1] https://github.com/theory/pgtap
I deal with a lot of complex querying logic with mostly Elasticsearch. My appproach is to either unit test or integration test and just skip everything in between. With queries what I care about is that they work under realistic scenarios against the same version of Elasticsearch that we use in production. Creating test fixtures is expensive. So tests for read only querying shares the same test fixtures. That speeds things up. I don't care about separately testing simple crud operations because most of my scenarios trigger plenty of those. A unit test for that has no value to me. Unit testing whether my API can send queries to some fake Elasticsearch, empty Elasticsearch, etc. has limited value to me. There's some limited value in knowing the queries I'm sending are syntactically correct. But you get that for free with a proper integration test not failing. The unit test is redundant if you have a proper integration test. And a unit test without a proper integration test provided little practical value.
What I actually do care about is all the complicated dashboard and other queries that make lot of assumptions about how data is structured, what fields are there, how they are indexed, whether they can be null, blank, or have invalid values, etc. work as intended. That kind of calls for an integration test. Anything trivial enough that a unit test would be good enough doesn't tend to need a lot of testing. Any scenario that touches enough of that stuff, kind of covers most of that.
I put a lot of effort in ensuring that my integration tests can run quickly, concurrently, and don't interact with each other (data randomization). That allows me to get away with not deleting a lot of data between tests and gets me a lot of realism for free because real users don't have an empty system completely to themselves. So having a lot of tests running against a busy system is integration testing gold. I have close to 300 full API integration tests running in around 30 seconds on my laptop. Close enough to unit testing performance that I run them many times per day.
The same approach applies to database testing. Probably more so because all the interesting bugs usually relate to constraints, transactionality, database locks, etc. If you have flaky tests because of that, it might actually be because your database layer has some issues with the notion of users not being polite enough to queue up one by one.
This is not for everyone, I realize. Do what works for you. I've butted heads with people over this more than a few times. But in my company (of which I'm the CTO), we unit test functions, small classes, regular expressions, parsing logic, etc. We integration test systems and APIs. Testing individual queries without the rest of the system is hard and pointless. Test the API that triggers the query. That 30 second performance for test runs is something I spent a lot of time on getting. It means we can do major changes without fear. If tests pass, our users should be fine.
One question, since you seem to have more experience with this approach.
> Test the API that triggers the query.
When you test the API, do you also check the state of the database?
Let's say there's an endpoint which creates an entity, and that operation involves setting a bunch of columns that aren't returned in the API.
When testing, do you check only the expected API response, or also the database state to be sure the operation was fully correct?
I currently do the latter, but this feels a bit iffy as the database internals are being leaked in an otherwise "blackbox" test.
If you learn towards the first option, how then do you verify all the unreturned database fields are correctly set?
Yes. This becomes especially obvious when you rewrite ORM garbage for something complicated, and are told that they can’t accept it, because they’re not sure how to test it.
Looking for performance issues on a machine with different baseline IO and CPU load, buffer state, query plans, cardinality, etc. is just theater and will lead to a false sense of security. RegreSQL is approaching a stateful problem as if it were stateless and deterministic. A linter like https://squawkhq.com is a good partial solution but only addresses DDL problems.
RegreSQL would be better served by focusing only on the aspects of correctness that tools like SQLx and sqlc fundamentally cannot address. This is a real need that too few tools try to address.
With current feature set it's something I have already considered but still undecided.
