Phoenix LiveView Tutorial: Handling Connection Errors And Push Notifications With JS Hooks
Did we ever say we were going to use Phoenix LiveView to build a Messenger app with no JavaScript?
We lied. This time, we're going to tell you why, how & when you can and should kick in with your JS magic when using Phoenix LiveView.
As of November 2020, the latest Phoenix LiveView version is 0.14.8 - and the series has been updated to match it!
It's been a while since we added Phoenix PubSub and user authentication with Pow to our Phoenix LiveView-based instant messaging app... but there it is, the long-awaited next part of our LiveView tutorial!
Psst! Don't forget to subscribe to our newsletter as there's still more to come.
A brutal wake-up call...
We were keen to point out that Phoenix LiveView can be treated as an alternative to reactive UI frameworks such as React or Vue.
While this is true, let's think about what can go wrong with full reliance on Phoenix LiveView's proposed approach.
Server-side statefulness, which is Phoenix LiveView's major selling point, can also be seen as its major weakness, creating several challenges to resolve.
HTTP is stateless by nature. You pass and mutate cookies back and forth between the browser and the server to create an illusion of state, storing a set of data - a session.
What happens in HTTP when the server's down? HTTP is stateless, so there's not much to worry about. You have your cookies on the browser side of things.
What happens in HTTP when the connection's down? Ditto. Well, technically there is not even such a thing as a connection in HTTP.
What happens in HTTP when an unhandled server exception occurs in data processing? HTTP is stateless, so there's not much to worry about. The server will bail out by throwing a 500, but remains stable and responds to requests, and you still have your state in the cookie.
With LiveView, forget what you're used to in HTTP.
Each instance of a user seeing a LiveView is backed by a stateful Elixir process supervised inside an Erlang VM instance. So whatever you'd like to call a session is part of the process's internal state.
What happens when the server is down? Suppose that the whole Erlang VM has crashed. By default, any state that may have been present is obviously lost.
What happens when the connection is down? Well, Phoenix Channels (backing all LiveViews) are monitored, and when a channel is down, the LiveView process is terminated. Therefore, state is lost.
What happens when an unhandled LiveView process exception occurs? This means the function that processes a state transition has no way to transform the process to a next state, so the process crashes. State is lost.
Let's see how we can overcome some of the challenges we've identified.
Reconnecting to the server
As we've noted, regardless of whether it's caused by the server going entirely down, or by losing connection between the browser at the server, it causes the LiveView process to lose its current state.
We'll do a quick test to learn at how Phoenix LiveView marks "dead" LiveViews in a page.
Phoenix LiveView uses .phx-disconnected
and .phx-error
selectors to visually mark tags that contain malfunctioning LiveView components.
This can be important, because it allows the developer to identify these DOM elements as belonging to a failing LiveView, which is important from a UX standpoint - it needs to be clear that e.g. a form inside such a failing view cannot be interacted with.
To test it, let's add the following selector in assets/css/app.css
:
.phx-error {
opacity: .25;
cursor: not-allowed;
}
.phx-error * {
pointer-events: none;
}
Before we proceed and start the server to test it out, let's temporarily disable watching for file changes and live-reloading the app's pages, because this would interfere with what we want to test here. To do this, jump into /lib/curious_messenger_web/endpoint.ex
and find and comment out the line that declares plug Phoenix.LiveReloader
. Its surroundings should look like this:
if code_reloading? do
socket "/phoenix/live_reload/socket", Phoenix.LiveReloader.Socket
# plug Phoenix.LiveReloader
plug Phoenix.CodeReloader
end
Let's now open up the server by executing mix.phx.server
, navigate to http://localhost:4000 and sign into a registered user.
If you dig into the DOM inspector at this point, you'll notice that there is a <div>
element that has the .phx-connected
class - this is the LiveView's main container.
Since our LiveView's internal state is represented by a changeset corresponding to our "Create Conversation" form, storing IDs of users to be added, let's add them to the to-be-created conversation, but instead of submitting the creation form, terminate the Phoenix server. Here is what you should see:

Note that, as we said, the element now becomes annotated with .phx-disconnected
and .phx-error
, and the CSS we wrote provides a basic way to distinguish a failed component from the rest of the page.
Now let's start the Phoenix server again. The browser will reconnect to the server promptly, but clearly the form will now be in its initial state, with nobody but the current user added to the conversation.
To test the scenario of a user's connection going down while the server is still up, connect to the app using your local IP address (as opposed to localhost
which uses a loopback interface) and disconnect from your network - and then reconnect - you'll again notice that the form is now back in the initial state, which means that the process that held the pre-disconnect state has been terminated.
This is actually good, because having to deal with dangling processes of lost LiveView connections would be disastrous. But it obviously also means that we need to make some conscious decisions about how we deal with reconnecting, if the LiveView's state contains some non-persistent data.
We'll look into how to tackle this in a moment. Now let's create a conversation and get into its view, and we'll see if we need to do anything with ConversationLive
with regard to behavior on reconnecting.
One could imagine that the most annoying thing that could happen is to lose a message you've been typing when connection issues occur. Will it happen? Let's find out: type in a message and close the server instead of submitting.

Reopen the server. You surely have noticed that the message is still there. Isn't that inconsistent with what we've seen with the behaviour of DashboardLive
? Let's do one more experiment to find out what's going on.
Go into /lib/curious_messenger_web/templates/conversation/show.html.leex
and find the line that declares the message content field:
<%= text_input f, :content %>
...and modify it to have a programmatically assigned value - it needn't be anything meaningful, let's just use a function from Erlang's :rand
module:
<%= text_input f, :content, value: :rand.uniform() %>
Refresh the page. You'll see a number entered into the "Content" field. Now enter whatever text you want in that field, close the server and reopen it. You'll notice that the text you had entered is lost, and there is a different random number entered in the field now.
Long story short, Phoenix LiveView is very good at splitting templates into static and dynamic parts, as well as identifying those elements in the DOM that needn't be replaced. While the <%= text_input f, :content %>
line is dynamic by nature regardless of whether there's a random number used there, it is rendered identically every single time - so it doesn't need to be replaced, because changing the field's value doesn't create any DOM change.
Controlling reconnection, restoring state
Coming back to the DashboardLive
module - we elected to store the form's state as an Ecto.Changeset
because it's pretty idiomatic to anyone familiar with Ecto and very easy to persist once we're done with selecting contacts to chat with. It's easy to reason about changesets when there are forms that correspond to a specific Ecto schema defining a changeset - in this case, we've got a Conversation
that accepts nested attributes for associated ConversationMember
records.
The changeset is a part of the component's state which can't be easily retrieved from elsewhere, so let's think of how we can make it at least a bit fail-safe.
If we wanted to make the form state remain intact between page visits, we could think of server-side solutions such as continously storing data related to the current user's form state in the relational database (PostgreSQL), Mnesia or the Erlang Term Storage (the latter being non-persistent, though). Stored data would then be retrieved at the moment the LiveView is mounted.
One could be tempted to think about cookies. But we don't want to use classic HTTP, and Phoenix's secure session cookie cannot be manipulated without getting into the HTTP request-response cycle - which we don't want to do.
With client-side (or mixed) approaches, it gets a bit more complicated. Depending on the nature of processed data, it might be acceptable to use browser's mechanisms such as localStorage
to store the form's state. And, finally, there is the need to actually write some JS code to feed the LiveView with an initial, post-reconnect state.
Since in the case of the "Create Conversation" form we don't necessarily need it to be persistent between page reloads, we'll only make it fail-safe in reconnection scenarios described earlier.
JavaScript Hooks
Phoenix LiveView allows us to write JS functions reacting to a LiveView instance's lifecycle events. According to the documentation, we can react to the following events:
mounted
- the element has been added to the DOM and its server LiveView has finished mountingupdated
- the element has been updated in the DOM by the serverdestroyed
- the element has been removed from the page, either by a parent update, or the parent being removed entirelydisconnected
- the element's parent LiveView has disconnected from the serverreconnected
- the element's parent LiveView has reconnected to the server
Create a /assets/js/create_conversation_form_hooks.js
file:
const CreateConversationFormHooks = {
disconnected() {
console.log("Disconnected", this)
},
reconnected() {
console.log("Reconnected", this)
},
mounted() {
console.log("Mounted", this)
},
destroyed() {
console.log("Destroyed", this)
},
disconnected() {
console.log("Disconnected", this)
},
updated() {
console.log("Updated", this)
}
}
export default CreateConversationFormHooks
In /assets/js/app.js
, ensure that the hooks are loaded and passed to the constructor of LiveSocket
. Let's have the file contain exactly the following code (excluding comments):
import css from "../css/app.css"
import "phoenix_html"
import { Socket } from "phoenix"
import LiveSocket from "phoenix_live_view"
import CreateConversationFormHooks from "./create_conversation_form_hooks";
let Hooks = { CreateConversationFormHooks };
let liveSocket = new LiveSocket("/live", Socket, { hooks: Hooks })
liveSocket.connect()
Fine, but how does LiveView know what view instance to react on? For this purpose, we annotate an element within a template with the phx-hook
attribute. In this case, we'll do this with the form we use for creating a conversation.
In /lib/curious_messenger_web/templates/dashboard/show.html.leex
, let's have the form declared like this:
<%= form_for @conversation_changeset, "", [id: "create-conversation-form", phx_submit: :create_conversation, phx_hook: "CreateConversationFormHooks"], fn f -> %>
...
<% end %>
If you open up your browser's developer tools and refresh the page at this point, you'll be able to observe what events are initially triggered - clearly these are mounted
and updated
. When you do anything that makes the LiveView update its state (e.g. add a conversation member), the updated
event is fired again. The destroyed
event would be triggered when e.g. a LiveView (possibly nested) completely disappears from the page - experience tells that it is not a reliable event on e.g. a page exit.
Now, try closing and restarting the Phoenix server, and here's what you'll probably end up with:

Most importantly, disconnected
and reconnected
happened. Conceptually, what we'd like to do is serialize the form to a changeset-friendly representation (e.g. URL-encoded form data) and send it over the LiveView's WebSocket connection when the server is up again (reconnected
).
Inside these hook functions, this
is a special object encapsulating the client-side LiveView representation and allowing us to communicate with the server, most importantly to send events that we'll handle in Elixir using handle_event
callbacks.
Let's now change the reconnected
function definition to the following:
reconnected() {
console.log("Reconnected", this)
let formData = new FormData(this.el)
let queryString = new URLSearchParams(formData)
this.pushEvent("restore_state", { form_data: queryString.toString() })
}
This will serialize the form into a form that we'll then convert into a changeset at the Elixir side. Recall that, at this point, the LiveView has been mounted and its process is up and running, so let's make it respond to the restore_state
event we send from JS code. In /lib/curious_messenger_web/live/dashboard_live.ex
, add a new clause for handle_event/3
:
def handle_event("restore_state", %{"form_data" => form_data}, socket) do
# Decode form data sent from the pre-disconnect form
decoded_form_data = Plug.Conn.Query.decode(form_data)
# Since the new LiveView has already run the mount function, we have the changeset assigned
%{assigns: %{conversation_changeset: changeset}} = socket
# Now apply decoded form data to that changeset
restored_changeset =
changeset
|> Conversation.changeset(decoded_form_data["conversation"])
# Reassign the changeset, which will then trigger a re-render
{:noreply, assign(socket, :conversation_changeset, restored_changeset)}
end
Psst! If you follow this pattern in multiple LiveViews, do yourself a favor and extract this into a reusable module.
Now, when you try disconnecting and reconnecting again, your "Create Conversation" form will be restored to the pre-disconnect state.
Reacting to new conversations
As an exercise at the end of our previous episode, we suggested that you can (and should) use Phoenix PubSub to notify a user's DashboardLive view when they've got a new conversation they participate in.
Again, to recap, it's just about making conversation creation broadcast a notification to all interested members - let's go to /lib/curious_messenger/chat.ex
and rework the create_conversation
function:
def create_conversation(attrs \\ %{}) do
result =
%Conversation{}
|> Conversation.changeset(attrs)
|> Repo.insert()
case result do
{:ok, conversation} ->
conversation.conversation_members
|> Enum.each(
&CuriousMessengerWeb.Endpoint.broadcast!(
"user_conversations_#{&1.user_id}",
"new_conversation",
conversation
)
)
result
_ ->
result
end
end
This needs to be reacted to in DashboardLive
, so we need to make changes in /lib/curious_messenger_web/live/dashboard_live.ex
.
In the mount/2
callback, insert the following code at the beginning:
def mount(%{current_user: current_user}, socket) do
CuriousMessengerWeb.Endpoint.subscribe("user_conversations_#{current_user.id}")
# prior implementation
end
Since the information about a new conversation being created can now come from elsewhere as well as from self, we'll now handle this information in handle_info
instead of handle_event
- the latter's clause for create_conversation
can now be simplified:
def handle_event("create_conversation", %{"conversation" => conversation_form},
%{assigns: %{conversation_changeset: changeset, contacts: contacts}} = socket) do
title = if conversation_form["title"] == "" do
build_title(changeset, contacts)
else
conversation_form["title"]
end
conversation_form = Map.put(conversation_form, "title", title)
case Chat.create_conversation(conversation_form) do
{:ok, _} ->
{:noreply, socket}
{:error, err} ->
Logger.error(inspect(err))
{:noreply, socket}
end
end
...and a new handle_info
clause for new_conversation
can be added:
def handle_info(%{event: "new_conversation", payload: new_conversation}, socket) do
user = socket.assigns[:current_user]
user = %{user | conversations: user.conversations ++ [new_conversation]}
{:noreply, assign(socket, :current_user, user)}
end
This way, users creating conversations for each other will get live updates of their conversation lists.
How to intercept this event so that, for example, a push notification mechanism can be created?
Phoenix LiveView itself does not expose an API to push messages from the server to the browser; everything that's pushed in this direction is DOM diffs.
You could think of leveraging Phoenix Channels, also built on top of Phoenix PubSub, to handle this kind of notifications; or, to keep your technological stack thin, you can just rely on JS updated
hooks interpreting data present in the new DOM.
Let's enhance the newly added handle_info
clause and piggyback a notify
flag on the newly created record. Remember how we once explained that a struct is just a map?
def handle_info(%{event: "new_conversation", payload: new_conversation}, socket) do
user = socket.assigns[:current_user]
annotated_conversation = new_conversation |> Map.put(:notify, true)
user = %{user | conversations: (user.conversations |> Enum.map(&(Map.delete(&1, :notify)))) ++ [annotated_conversation]}
{:noreply, assign(socket, :current_user, user)}
end
Consume this in the /lib/curious_messenger_web/templates/dashboard/show.html.leex
view the following way - plugging in the new ConversationListHooks
that we'll create in a moment into the conversation list.
<article id="conversation-list" class="column" phx-hook="ConversationListHooks">
<h2>Ongoing Conversations</h2>
<%= for conversation <- @current_user.conversations do %>
<div>
<%= link conversation.title,
to: Routes.conversation_path(@socket, CuriousMessengerWeb.ConversationLive, conversation.id, @current_user.id),
data: if Map.get(conversation, :notify), do: [notify: true], else: [] %>
</div>
<% end %>
</article>
This way, only one conversation link will be annotated with the data-notify
attribute that we'll then look for in JS code so that we can notify the user about it.
Add the following import to /assets/js/app.js
and modify the Hooks
declaration. Also ask the user for permission to display notifications:
import ConversationListHooks from "./conversation_list_hooks";
let Hooks = { CreateConversationFormHooks, ConversationListHooks };
Notification.requestPermission() // returns a Promise that we're not much interested in for now
...and create a new file in /assets/js/conversation_list_hooks.js
. Notice that, in this case, this.el
is the element to which the hook was attached in the HTML template:
const ConversationListHooks = {
updated() {
let newConversationLink = this.el.querySelector('[data-notify]')
if (!newConversationLink) return
let notification = new Notification(newConversationLink.innerText)
notification.onclick = () => window.open(newConversationLink.href)
}
}
export default ConversationListHooks
This way, everyone involved with the conversation will now receive a push notification when it is created! It is slightly imperfect, because the notification will also be displayed to the user who created the conversation. Nonetheless, it is a good starting point to expand the push functionality to messages as well.

Let's get to messages now. Add an import and change the hooks declaration again in /assets/js/app.js
:
import ConversationHooks from "./conversation_hooks"
let Hooks = { CreateConversationFormHooks, ConversationListHooks, ConversationHooks }
Create a new file in /assets/js/conversation_hooks.js
. We'll rely on incoming messages - that we want to have notifications on - to have a data-incoming
attribute.
const ConversationHooks = {
updated() {
if (!this.notifiedMessages)
this.notifiedMessages = []
this.el.querySelectorAll('[data-incoming]').forEach(element => {
if (!this.notifiedMessages.includes(element)) {
this.notifiedMessages.push(element)
let notification = new Notification(element.innerText)
notification.onclick = () => window.focus()
}
})
}
}
export default ConversationHooks
Go to /lib/curious_messenger_web/templates/conversation/show.html.leex
. Find the for
loop that renders message markup and replace it like this, so that if a message is annotated with an additional :incoming
key, the div
tag should be displayed with an additional data attribute:
<div id="conversation" phx-hook="ConversationHooks">
<%= for message <- @messages do %>
<%= if Map.get(message, :incoming) do %>
<div data-incoming="true">
<% else %>
<div>
<% end %>
<b><%= message.user.nickname %></b>: <%= message.content %>
</div>
<% end %>
</div>
Then, go to /lib/curious_messenger_web/live/conversation_live.ex
and modify the definition for handle_info
dealing with new messages:
def handle_info(%{event: "new_message", payload: new_message}, socket) do
annotated_message =
if new_message.user.id != socket.assigns[:user].id do
new_message |> Map.put(:incoming, true)
else
new_message
end
updated_messages = socket.assigns[:messages] ++ [annotated_message]
{:noreply, socket |> assign(:messages, updated_messages)}
end
This time, the message's notification will not appear for the user that sent it.
You can now test this out - notifications will appear in your conversations! One caveat is that you'll only see them while the conversation tab is running. **If you'd like to use notifications that don't require the tab to be open, read up on the Push API and Service Workers - and, at the Elixir side of things, get familiar with the web_push_encryption
library.
What else can go wrong with Phoenix LiveView?
It has to be kept in mind that, in LiveView, with the whole logic behind events and UI updates being on the server side of things.
Server-side event handling makes LiveView far from ideal for aplications that require offline access**.
If an application is intended for offline use, it might be a better idea to go for solutions that don't use LiveView for driving the logic behind form controls, and perhaps restrict it to displaying some live-updated data - which, with Phoenix PubSub available, is still very convenient.
You could go more ambitious and use the disconnected
hook to create a plug custom logic into your form when it has lost connection. If you've got an UI whose concept is simple and linear - for instance, let's say you've got a button which you count clicks on - you could try to trace any activity that occurs in the relevant controls (or a form) during the time the connection is down - and push it as events to the server using pushEvent
when the connection is back. It's not perfect, though, because the UI will still not appear to be reactive during the connection outage.
If your application requires full UI reactivity in offline mode, look for other solutions than Phoenix LiveView for the most critical components.
Wrapping up
Here's the repository of our app, current as of the end of this episode.
There is still some polishing left to do in our app - we'll review the elements of our schema that we haven't covered in the next episode.
Without, though, we've managed to explain a few of the issues you can run into in a production LiveView app, and resolve a few of them - learning how to use JS hooks to react to Phoenix LiveView lifecycle events.
Have you liked the new episode? Give us your feedback via contact form or our social media (don't forget to share using links below). If you don't mind, please also do us a favour and subscribe to our newsletter.
Check out other articles in this series: