Blocking resource with `SsrMode::PartiallyBlocked` doesn't work consistently

[spencewenski]

Describe the bug
SsrMode::PartiallyBlocked doesn't seem to block rendering as expected in 0.7 (main). Sometimes it works, but occasionally the suspense renders the fallback view instead.

Additionally, I noticed that the behavior seems to be the same as when the default ssr mode is used.

Leptos Dependencies

Reproduced on main with some minor modifications to the hackernews_axum example (in this commit)

To Reproduce
Steps to reproduce the behavior:

1. Disable javascript in your browser
2. Checkout my fork: git clone -b context-resource-startup https://github.com/spencewenski/leptos
3. Run the hackernews_axum example
4. Navigate to the example url (http://127.0.0.1:3000)
5. Refresh the page a few times until only the Loading... fallback is displayed

Expected behavior
The stories content should always be displayed instead of the Loading... fallback.

Screenshots

Additional context
Test on:

• Firefox - 130.0.1
• Arch linux - 6.10.10-arch1-1

[gbj]

Oh at the moment the actual "partially blocked" behavior of replacing the fallback HTML inside the string is just a TODO, it is not yet implemented at all.

Note that in the large majority of cases SsrMode::Async or SsrMode::InOrder is preferable to PartiallyBlocked, and that PartiallyBlocked is only useful if you want a loading placeholder in one part of the page but you definitely don't want a loading placeholder in another part of the page. (For example, a blog post and comments.) In that case with JS disabled you will end up with a perpetual "Loading..." stuck somewhere anyway. In the hackernews app for example, SsrMode::Async is better than Resource::new_blocking() + SsrMode::PartiallyBlocked.

That said yes, it's in the todos.

[gbj]

Oh interesting, in fact I implemented this so long ago that I forgot about it, and in a relatively better way, such that the need for a separate PartiallyBlocked goes away entirely because it will work with regular OutOfOrder too. Fair enough.

I can't reproduce the issue on Actix which suggests to me immediately that it's a multi-threading issue (since Actix pins the whole request/response to a single thread and Axum does not.) Basically something is waking the stream back up to start sending chunks slightly before the chunk is actually ready. I imagine the cases where it sends the placeholder at the cases where Tokio has shifted it between threads, I'll have to take a look.

[metatoaster]

I imagine the cases where it sends the placeholder at the cases where Tokio has shifted it between threads, I'll have to take a look.

I just ran into a similar issue with Resource::new_blocking() even with SsrMode::Async that goes away completely when I switched to the single-threaded mode for tokio using #[tokio::main(flavor = "current_thread")]. The reproduction is a bit complicated but given the design I've already rough implemented when I extended the suspense test case, specifically a variant of the portlet I introduced in this test except rather than using the read signal directly it's fed to a resource as I wanted isomorphic SSR/CSR rendering on my production app as opposed to just having CSR working for the test.

If I find some time I might be able to adjust that example to make a reproduction, but seeing the error being reproduced can be difficult as only about 1 in 250 requests the SSR gets misrendered with my live code.