Representing ViewModel events with StateFlow vs. SharedFlow vs. Channel
One of the most heated debates in the Android community is the use of StateFlow
, SharedFlow
, and Channel
in the ViewModel. In this article, we will discuss the differences between these three, and suggest which one is the best for representing events.
StateFlow
for events
Let's start from saying, that for most elements of the UI, StateFlow
is undoubtedly the best choice. We can assume it is best for all the elements of our "state".
However, it can be tricky when we use it for representing events, like showing a toast, navigation, or performing an action. Here we have a problem, as StateFlow
is not designed for this purpose.
Let's consider using MutableStateFlow
with a nullable value. It is null
at start, and after an event is handled, so a non-null value can be seen as an event to handle.
There are some problems though:
- We need to remember to clear the state after handling the event, which makes this pattern a bit more complicated.
- If two coroutines send two different events in a short period of time, one of those events might be lost. Especially if they have the same value, because
MutableStateFlow
ignores updates with the same value, but even different values can be lost if the second event is sent before the first one is handled, becauseStateFlow
is conflated. So essentially, there are a few ways in which events can be lost.
Let's consider some other options now.
SharedFlow
for events
In general, the best practice to represent events is to use SharedFlow
. It is a much simpler abstraction that always emits value to all its current collectors.
However, SharedFlow
has a problem with view models, where UI can change, and in the meantime there is no observer. We deal with such a situation, for instance, when a user rotates the screen. If an event is sent during the rotation, it will be lost. I saw some people using a workaround for this problem: awaiting for the first observer to appear, and then sending the event. I like this solution, but I cannot guarantee that it is bulletproof, so I wouldn't use it for important events.
The advantage of SharedFlow
is that it is a simple, and it is the only solution that can be freely used by more than one observer.
Channel
for events
The natural way of dealing with the aforementioned problem is using Channel
with an unlimited capacity. It can be seen as a queue of events that are received by collectors. We can transform it into a Flow
using receiveAsFlow
function. If an event is sent when there is no observer, it will be received by the next observer.
The problem of using a Channel
is that there is a possibility that this event is lost, in an unlikely event of cancellation after sending the event, but before invoking its action (see this article and this issue).
The most important argument against using Channel
is that it does not guarantee event delivery (that should not be a problem if we both send and receive events in Dispatchers.Main.immediate
, what is quite often the case, as it is used by both viewModelScope
and lifecycleScope
, but always ensuring that is a fragile solution). That is why we should avoid this solution for events that are important for the user experience, like a transaction result. Such an event is better represented with StateFlow
, that guarantees that the event will be delivered to the observer.
Turning events into state and using StateFlow
Time for the option, that I most often hear being recommended by Googlers. It is turning events into state, and using StateFlow
for them. That is the only option I am sure is bulletproof, but my problem is that it is not the simplest one.
Let's consider an example of showing a snackbar. To represent it as a state, we would need to use a list of messages, show only the first one, and remove it after showing.
It is probably possible that the same message will be shown twice, but not that it will be lost, which is preferable for important events.
A good argument behind this solution is that we do not need to learn intricacies of SharedFlow
or Channel
, we can just use StateFlow
for everything, and develop patterns of using it for different purposes.
However, on the other hand, there is a good argument that events should be represented as events, and not as state (event happen, state is), and for that SharedFlow
or Channel
are better solutions (see this article).
Summary
I will not have a clear answer to the question of which one is the best, as from my point of view, for most cases it does not matter that much which one you choose. However, I do agree that for important events, it is safest to use StateFlow
. Nevertheless, for most cases, I still see the advantage of using SharedFlow
, for its possibility of having more than one observer, or Channel
, for its simplicity. On the other hand, I wouldn't oppose if Google developers decided to standardize on using StateFlow
for everything, as it might be easier and safer for those who are not familiar with SharedFlow
or Channel
.