Feature: async

[mfbx9da4]

While mmkv is super fast, and that's awesome 😎 it still blocks the main JS thread and for situations where you have to get a bunch of keys in a loop of some kind, that can add up. It would be great if mmkv could provide a threaded async version where

• Stored values are loaded from mmkv in a thread
• JSI strings and numbers are constructed in a thread
• Promise / callback is resolved back to main JS thread with jsCallInvoker

Do you think that would be possible?

[mrousavy]

Yep, that's definitely possible!

I'm not sure if I'd be happy adding an async method for every sync method though, because the API will look a bit messy then. Right now it's very simple and minimal. 😅 Maybe we can find a way around that

Either way, it's a little bit of work, not sure how much exactly but we need to set up a Promise resolver that runs on a separate Thread, custom batched methods, etc etc

[mfbx9da4]

API wise maybe the counterpart to all the standard methods eg. mmkv.getString() and mmkv.getNumber() could be mirrored under the property async ie mmkv.async.getString(), mmkv.async.getNumber(), mmkv.async.getMany()?

Do you like that? @mrousavy

Another hacky thing you could do is have mmkv.getString.async(), I've seen Zustand do that but I kind of hate it and it won't help with batch methods.

Personally I think the API mmkv.getStringAsync() and mmkv.getNumberAsync() are the most natural as they will be the easiest to find for typescript users and I think I've seen some NodeJS APIs use that pattern.

[mfbx9da4]

Just one question around thread safety, say for example we implement the method mmkv.setManyAsync(items: Array<[string, string | number | boolean]). Will it actually be thread safe to create a thread and then inside the thread do convertJSIStringToNSString(argument.getNumber()), argument.getBool(), argument.getNumber() or will that have to be done outside the thread?

[mrousavy]

Arg parsing will need to happen outside, but that's fast anyways.

[mrousavy]

I'm thinking of a potential solution for setting.

I can implement one additional method on the MMKV instance called batch.

const storage = new MMKV()
await storage.batch((s) => {
  s.set('hello', 'world')
  s.set('test', 'another')
  // ...
})

Inside you will use the s instance (same MMKV instance, but a Proxy), which will schedule the methods instead of actually running them.

The only problem I see with this now is that you cannot use getters (s.getString, s.getNumber, s.getBoolean) because those would have to be synchronous.

A "solution" for that would be to make those funcs async, but that's a bit complicated again.

[mfbx9da4]

🤔

Perhaps batch would have to return an array of items, does that solve it?

const storage = new MMKV()
const [foo] = await storage.batch((s) => [
  s.getString('foo'),
  s.set('hello', 'world'),
  s.set('test', 'another')
])

[mfbx9da4]

Also could you not make storage to be context aware, then you could have this API:

const storage = new MMKV()
const [foo] = await storage.batch(() => [
  storage.getString('foo'),
  storage.set('hello', 'world'),
  storage.set('test', 'another')
])

Provided you enforce that batch is always a synchronous function.

BTW, I love the batch() idea

[mrousavy]

Perhaps batch would have to return an array of items

That's a good idea, but then you'd be very limited in terms of doing stuff contextually (e.g. only set if the value in storage is ....), but yea that would probably work.

Provided you enforce that batch is always a synchronous function.

Wdym? The whole idea is to have batch asynchronous and awaitable

[mfbx9da4]

Wdym? The whole idea is to have batch asynchronous and awaitable

Sorry I mean the batch implementation would enforce that the batch argument callback is synchronous, this is so that the batchFn doesn't need to be injected with a proxy.

class Storage {
  // ...
  batch(fn) {
    ensureSynchronous(fn)
    this._batchingEnabled = true
    try {
      const items = fn().filter(x => x)
      items.every(x => ensureIsBatchItem(x))
      return await this._executeBatch(items)
    } finally {
      this._batchingEnabled = false
    }
  }
}

[mfbx9da4]

but then you'd be very limited in terms of doing stuff contextually (e.g. only set if the value in storage is ....)

What use case might be limited?

[mrousavy]

eg. this (but on a larger scale with tons of items like arrays):

const storage = new MMKV()
await storage.batch((s) => {
  const x = s.getBoolean('bla')
  if (x === true) {
    s.set('hello', 'world') 
  } else {
    s.set('test', 'another')
  }
  // ...
})

[mfbx9da4]

Hmm yeah it's true that wouldn't work and having a callback does give the impression that it should be possible. One saving grace is that typescript users will get the following error:

const storage = new MMKV()
await storage.batch((s) => {
  const x = s.getBoolean('bla')
  // ❌ This condition will always return 'false' since the types 'MMKVBatchItem' and 'boolean' have no overlap
  if (x === true) {
    s.set('hello', 'world') 
  } else {
    s.set('test', 'another')
  }
  // ...
})

I don't see how you could support the above API, unless... you implemented a similar strategy to react-native-multithreading and all batchFns were marked automatically with a "worklet" like tag? While it's probably overkill for MMKV, it's fun to dream.

[mfbx9da4]

BTW one thing which is way better about .batch vs storage.async.{{method}} is that there will be less back and forth between Microtasks Queue, C++, JSI and JS and you will only pay the cost once per batch. Whereas, with my original API you would expect to pay the cost three times here:

await Promise.all([
  storage.async.getString('foo'), 
  storage.async.getString('bar'), 
  storage.async.getString('baz')
])

Although, thinking about it more, even with the .async approach, in theory you could batch any calls within the same microtask 🤔. Still though, explicit batching is probably preferable.

[mfbx9da4]

Also maybe just twitter poll the options?

[elliotsayes]

+1 For async, as it will make larger buffer storage play nicely :)

In the meantime, is there any way to wrap the existing methods so they don't block the main thread ?