[cuishuang]

Description

Fix the client-side use_linked_state retry loop caused by initial LinkError::Internal results not being cached.

Initial fetches now cache all completed results, including infrastructure errors. This allows the suspension to resolve and exposes the error through LinkedStateHandle::as_result() instead of creating another identical request.

Refresh behavior remains unchanged: LinkError::Internal results from background refreshes are not cached, so an existing stale value is preserved.

Fixes #4218

Checklist

• I have reviewed my own code
• I have added tests

[github-actions]

Visit the preview URL for this PR (updated for commit 1a9fd92):

https://yew-rs-api--pr4219-master-im3mvlbo.web.app

_{(expires Sat, 14 Nov 2026 11:51:35 GMT)}

_{🔥 via Firebase Hosting GitHub Action 🌎}

[github-actions]

Size Comparison

Details

examples	master (KB)	pull request (KB)	diff (KB)	diff (%)
async_clock	100.169	100.169	0	0.000%
boids	164.983	164.983	0	0.000%
communication_child_to_parent	93.702	93.702	0	0.000%
communication_grandchild_with_grandparent	105.983	105.983	0	0.000%
communication_grandparent_to_grandchild	102.312	102.312	0	0.000%
communication_parent_to_child	91.099	91.099	0	0.000%
contexts	106.298	106.298	0	0.000%
counter	86.243	86.243	0	0.000%
counter_functional	88.285	88.285	0	0.000%
dyn_create_destroy_apps	90.153	90.153	0	0.000%
file_upload	98.562	98.562	0	0.000%
function_delayed_input	94.622	94.622	0	0.000%
function_memory_game	168.692	168.692	0	0.000%
function_router	401.547	401.547	0	0.000%
function_todomvc	164.991	164.991	0	0.000%
futures	236.843	236.843	0	0.000%
game_of_life	100.612	100.612	0	0.000%
immutable	259.449	259.449	0	0.000%
inner_html	80.784	80.784	0	0.000%
js_callback	109.766	109.766	0	0.000%
keyed_list	176.314	176.314	0	0.000%
mount_point	84.267	84.267	0	0.000%
nested_list	112.704	112.704	0	0.000%
node_refs	91.772	91.772	0	0.000%
password_strength	1767.894	1767.894	0	0.000%
portals	93.578	93.578	0	0.000%
router	367.774	367.774	0	0.000%
suspense	113.953	113.953	0	0.000%
timer	88.515	88.515	0	0.000%
timer_functional	99.403	99.403	0	0.000%
todomvc	142.016	142.016	0	0.000%
two_apps	86.185	86.185	0	0.000%
web_worker_fib	136.346	136.346	0	0.000%
web_worker_prime	186.646	186.646	0	0.000%
webgl	82.914	82.914	0	0.000%

✅ None of the examples has changed their size significantly.

[cuishuang]

@WorldSEnder Modified. Please review it again.

[WorldSEnder]

Took a bit, since I am myself not familiar with this part of the library and was hoping Mado would chime in. So bear with and feel free to correct me too.

completed being an (unbounded) hashmap is I believe a memory leak (of all failed requests that never succeed). Keeping the old property of the LRU cache seems worthwhile.

I would phrase the property we want as follows: Every use of the use_linked_state hook has associated with it an input, call this the query. So when the input changes, the query does but it's the same hook. Every query should lead to at most one request to the backend (this is what I understand the bug is about). Multiple simultaneous queries from different uses of the hook for the same input should share the request (only one at best) - implemented by in_flight and refreshing. A query we already answered in SSR should lead to no request (subtly a different hook might have the same query without SSR, this can lead to a query request since cache is unreliable). And finally, memory should be proportional to the hooks used. After unmounting all, only a constant should be left, that's why LRU cache is used but HashMap is used only for things in flight.

The last property I think implies that two different hooks querying for the same input might not always share the result (which could have dropped out of the LRU already) and the hook needs to have a local cache, too.

EDIT: To come back to my earlier comment: each component should remember if a request failed and not retry again. The shared cache should be used only for successful requests.

Sorry this took so long, caching is hard.

[WorldSEnder]

Note: prior art for request coalescing in rs-singleflight could be helpful.

[cuishuang]

@WorldSEnder Thanks for spelling out the query lifetime and for the rs-singleflight pointer. I removed the shared completed map. Each hook now keeps the result for its current input, including failures, and the bounded LRU stores only successful values.

Initial fetches and explicit refreshes share one in-flight request per key. Hooks that join a request can still read its result after it leaves the in-flight map, even if the LRU entry has been evicted. SSR-prepared results are used directly, and an internal refresh error keeps the previous value visible.

I added a browser regression test for request coalescing, explicit retries, and LRU eviction. The updated PR is squashed into one commit and ready for another look.