Massive perf drop in probably 5.41.1

[gyzerok]

Describe the bug

After recent upgrade from 5.19.12 to 5.41.1 I found our app is freezing to multiple seconds (got as much as 5s in some cases) when I click through UI.

Reproduction

I am not yet sure how to make a reproduction for it outside our code, but I think the problem is important enough to start discussion about it and iterate on the findings.

What I know so far:

1. We are not using async
2. It seems like the regression appears in 5.41.1. I can't say for sure, because I don't quite know a specific set of steps that lead to it. My guess it's branch refactoring PR
3. The best way to trigger it seems to be fast clicking through some tabs in our UI. So probably something isn't cleaned up correctly and get interleaved?

Bottoms up

Source with timings

Logs

System Info

I can repro it both in Chrome and Firefox, so seems like env is irrelevant

Severity

blocking an upgrade to

[dummdidumm]

There's two layers to this it seems:

• fix: don't rerun async effects unnecessarily #16944 added logic to check if a block or async effect should rerun. That depends_on seems to be very slow in your case
• chore: centralise branch management #16977 somehow makes it so that you get multiple branches despite not using the experimental async feature, and then running into the depends_on when you should not at all

[gyzerok]

UPDATE

Seems like even though we are not using async there is a possibility to arrive at 2 simultaneous batches which result in some infinite loop of depends_on calls.

How to arrive at this is still unclear though.

[timc13]

I believe I am also experiencing the same thing - symptoms are navigation and reactivity freezing up. I tested each minor upgrade and found this regression was introduced 5.41.0 -> 5.41.1

[gyzerok]

I believe I am also experiencing the same thing - symptoms are navigation and reactivity freezing up. I tested each minor upgrade and found this regression was introduced 5.41.0 -> 5.41.1

Any chance your code is public or you can extract a reproduction from it?

[MathiasWP]

Are there no performance tests that capture these types of performance impacts? Considering Svelte really cares about performance. This has a substantial impact on our performance in production - it froze the application many cases :/

Here's our flame graph:

[gyzerok]

Are there no performance tests that capture these types of performance impacts? Considering Svelte really cares about performance. This has a substantial impact on our performance in production - it froze the application many cases :/

The problem is not the lack of perf tests but some race condition in which sync svelte starts behaving like async one (at least in our case). Unless there were a test specifically designed for that race condition you wouldn't catch it. And you can't write such performance test for a bug you don't know is going to exist. But arguably if there were such a test it'd be a unit test, not a perf test.

Are you able to reproduce it reliably? Do you think you can share a reproduction? That'd help a lot!