produce a stream instead of reading everything into a buffer

[michaelficarra]

It seems like there'd be a number of advantages to producing something more like ReadableStream rather than reading the whole resource into a buffer. For example, you'd be able to consume just part of the stream and cancel it, and you could support potentially infinite resources.

Of course, this would have the prerequisite of adding streams to the language, which would have its own challenges.

[styfle]

Its a good point. Streams would provide more flexibility.

However, I think there is still value in shipping the buffer solution as proposed since it can be statically analyzed and inlined by bundlers as mentioned in the Prior Art section of the readme.

[michaelficarra]

I don't think there's any difference for bundlers. They could inline a stream just as easily as they could inline a buffer.

Also I don't understand the point about static analysis. Do you mean analysis of its use sites? Like you want to do constant propagation and a sync consumption model would be easier than an async one?

[styfle]

You're right, you could inline a stream but it would need to be buffered first and then converted to a stream. Most bundlers use base64 to inline the bytes and then convert that to something a little more usable (either data uri or uin8array). So the idea of consuming part of the stream or canceling is already possible with the buffer, no need for the stream.

However, when we consider the case of browsers fetching the module, then a stream makes more sense because it aligns more with fetch.

That said, why can't we have both? It could be a separate proposal with type stream.

import stream from "./photo.png" with { type: "stream" };

[michaelficarra]

That said, why can't we have both?

Because we usually try to avoid having features that overlap so much in the use cases they cover. I think a stream would cover all of the buffer use cases we care about and more. The hard part is the meta-question of layering since this is yet another thing that we don't have in the language, probably should, but already exists on the web in a different spec.

[bakkot]

I think if we only exposed stream then 90%+ of uses would start with await stream.bytes() (assuming such a thing comes to exist), in which case it would be kind of silly not to provide that functionality directly. It costs relatively little and allows people to directly express what they actually mean.

[styfle]

I’ll update the readme, thanks!