Iterating over More Complex Data with <For/>
This chapter goes into iteration over nested data structures in a bit more depth. It belongs here with the other chapter on iteration, but feel free to skip it and come back if you’d like to stick with simpler subjects for now.
The Problem
I just said that the framework does not rerender any of the items in one of the rows, unless the key has changed. This probably makes sense at first, but it can easily trip you up.
Let’s consider an example in which each of the items in our row is some data structure. Imagine, for example, that the items come from some JSON array of keys and values:
#[derive(Debug, Clone)]
struct DatabaseEntry {
key: String,
value: i32,
}
Let’s define a simple component that will iterate over the rows and display each one:
#[component]
pub fn App() -> impl IntoView {
// start with a set of three rows
let (data, set_data) = create_signal(vec![
DatabaseEntry {
key: "foo".to_string(),
value: 10,
},
DatabaseEntry {
key: "bar".to_string(),
value: 20,
},
DatabaseEntry {
key: "baz".to_string(),
value: 15,
},
]);
view! {
// when we click, update each row,
// doubling its value
<button on:click=move |_| {
set_data.update(|data| {
for row in data {
row.value *= 2;
}
});
// log the new value of the signal
logging::log!("{:?}", data.get());
}>
"Update Values"
</button>
// iterate over the rows and display each value
<For
each=data
key=|state| state.key.clone()
let:child
>
<p>{child.value}</p>
</For>
}
}
Note the
let:child
syntax here. In the previous chapter we introduced<For/>
with achildren
prop. We can actually create this value directly in the children of the<For/>
component, without breaking out of theview
macro: thelet:child
combined with<p>{child.value}</p>
above is the equivalent ofchildren=|child| view! { <p>{child.value}</p> }
When you click the Update Values
button... nothing happens. Or rather:
the signal is updated, the new value is logged, but the {child.value}
for each row doesn’t update.
Let’s see: is that because we forgot to add a closure to make it reactive?
Let’s try {move || child.value}
.
...Nope. Still nothing.
Here’s the problem: as I said, each row is only rerendered when the key changes.
We’ve updated the value for each row, but not the key for any of the rows, so
nothing has rerendered. And if you look at the type of child.value
, it’s a plain
i32
, not a reactive ReadSignal<i32>
or something. This means that even if we
wrap a closure around it, the value in this row will never update.
We have three possible solutions:
- change the
key
so that it always updates when the data structure changes - change the
value
so that it’s reactive - take a reactive slice of the data structure instead of using each row directly
Option 1: Change the Key
Each row is only rerendered when the key changes. Our rows above didn’t rerender, because the key didn’t change. So: why not just force the key to change?
<For
each=data
key=|state| (state.key.clone(), state.value)
let:child
>
<p>{child.value}</p>
</For>
Now we include both the key and the value in the key
. This means that whenever the
value of a row changes, <For/>
will treat it as if it’s an entirely new row, and
replace the previous one.
Pros
This is very easy. We can make it even easier by deriving PartialEq
, Eq
, and Hash
on DatabaseEntry
, in which case we could just key=|state| state.clone()
.
Cons
This is the least efficient of the three options. Every time the value of a row
changes, it throws out the previous <p>
element and replaces it with an entirely new
one. Rather than making a fine-grained update to the text node, in other words, it really
does rerender the entire row on every change, and this is expensive in proportion to how
complex the UI of the row is.
You’ll notice we also end up cloning the whole data structure so that <For/>
can hold
onto a copy of the key. For more complex structures, this can become a bad idea fast!
Option 2: Nested Signals
If we do want that fine-grained reactivity for the value, one option is to wrap the value
of each row in a signal.
#[derive(Debug, Clone)]
struct DatabaseEntry {
key: String,
value: RwSignal<i32>,
}
RwSignal<_>
is a “read-write signal,” which combines the getter and setter in one object.
I’m using it here because it’s a little easier to store in a struct than separate getters
and setters.
#[component]
pub fn App() -> impl IntoView {
// start with a set of three rows
let (data, set_data) = create_signal(vec![
DatabaseEntry {
key: "foo".to_string(),
value: create_rw_signal(10),
},
DatabaseEntry {
key: "bar".to_string(),
value: create_rw_signal(20),
},
DatabaseEntry {
key: "baz".to_string(),
value: create_rw_signal(15),
},
]);
view! {
// when we click, update each row,
// doubling its value
<button on:click=move |_| {
data.with(|data| {
for row in data {
row.value.update(|value| *value *= 2);
}
});
// log the new value of the signal
logging::log!("{:?}", data.get());
}>
"Update Values"
</button>
// iterate over the rows and display each value
<For
each=data
key=|state| state.key.clone()
let:child
>
<p>{child.value}</p>
</For>
}
}
This version works! And if you look in the DOM inspector in your browser, you’ll
see that unlike in the previous version, in this version only the individual text
nodes are updated. Passing the signal directly into {child.value}
works, as
signals do keep their reactivity if you pass them into the view.
Note that I changed the set_data.update()
to a data.with()
. .with()
is the
non-cloning way of accessing a signal’s value. In this case, we are only updating
the internal values, not updating the list of values: because signals maintain their
own state, we don’t actually need to update the data
signal at all, so the immutable
.with()
is fine here.
In fact, this version doesn’t update
data
, so the<For/>
is essentially a static list as in the last chapter, and this could just be a plain iterator. But the<For/>
is useful if we want to add or remove rows in the future.
Pros
This is the most efficient option, and fits directly with the rest of the mental model of the framework: values that change over time are wrapped in signals so the interface can respond to them.
Cons
Nested reactivity can be cumbersome if you’re receiving data from an API or another data source you don’t control, and you don’t want to create a different struct wrapping each field in a signal.
Option 3: Memoized Slices
Leptos provides a primitive called create_memo
,
which creates a derived computation that only triggers a reactive update when its value
has changed.
This allows you to create reactive values for subfields of a larger data structure, without needing to wrap the fields of that structure in signals.
Most of the application can remain the same as the initial (broken) version, but the <For/>
will be updated to this:
<For
each=move || data().into_iter().enumerate()
key=|(_, state)| state.key.clone()
children=move |(index, _)| {
let value = create_memo(move |_| {
data.with(|data| data.get(index).map(|d| d.value).unwrap_or(0))
});
view! {
<p>{value}</p>
}
}
/>
You’ll notice a few differences here:
- we convert the
data
signal into an enumerated iterator - we use the
children
prop explicitly, to make it easier to run some non-view
code - we define a
value
memo and use that in the view. Thisvalue
field doesn’t actually use thechild
being passed into each row. Instead, it uses the index and reaches back into the originaldata
to get the value.
Every time data
changes, now, each memo will be recalculated. If its value has changed,
it will update its text node, without rerendering the whole row.
Pros
We get the same fine-grained reactivity of the signal-wrapped version, without needing to wrap the data in signals.
Cons
It’s a bit more complex to set up this memo-per-row inside the <For/>
loop rather than
using nested signals. For example, you’ll notice that we have to guard against the possibility
that the data[index]
would panic by using data.get(index)
, because this memo may be
triggered to re-run once just after the row is removed. (This is because the memo for each row
and the whole <For/>
both depend on the same data
signal, and the order of execution for
multiple reactive values that depend on the same signal isn’t guaranteed.)
Note also that while memos memoize their reactive changes, the same calculation does need to re-run to check the value every time, so nested reactive signals will still be more efficient for pinpoint updates here.