I think if we wanted to do that we would also need to change the name of the method.

We want Go 1 and Go 2 APIs to be able to coexist in the same program, and per https://research.swtch.com/vgo-import, that requires that “when you change a function’s behavior, you also change its name.”

It would be useful to know whether there are any existing implementations of io.Reader that can return zero with no error. If there are any, why do they exist?

See #5309 and #5310. In particular #5309 suggests that (*tls.Conn).Read() could return (0, nil). Is that still the case?

It would be useful to know whether there are any existing implementations of io.Reader that can return zero with no error. If there are any, why do they exist?

I think the net.Conns returned by net.Pipe can. I believe the reason is so that Write can unblock a pending Read without closing the pipe, but I'm not really sure.

(I didn't find a compelling reason the last time I checked: #24205 (comment).)

Though undocumented, the flate.Reader can return (0, nil) when it encounters successive Z_SYNC_FLUSH markers. There are tests to preserve this behavior and some network protocols depend on this.

EDIT: I just checked. flate.Reader actually only returns on SYNC_FLUSH if there at least some number of bytes read (example). That said, I got an email recently complaining that the reader did not return on all SYNC_FLUSH, which would imply that it may return (0, nil).

surely the contract, put as a conversation, is...

reader.Read == 'i want some bytes in this buffer, let me know how many you have put in it and if there is a problem that effects me repeating this action'.

the reader Reads without knowledge of how many are available, isn't this isolation a fundamental point of a reader.

It is interesting that https://golang.org/pkg/io/#Reader discourages this but allows it.

Implementations of Read are discouraged from returning a zero byte count with a nil error, except when len(p) == 0.

It appears this was considered when designing the io package. Perhaps to allow readers where Read can be used to poll for data to be read instead of blocking? It's unidiomatic for sure but that's the only reason I can think of. And anyone who definitely wants to block uses io.ReadFull anyway.

@nhooyr The guidelines for io.Reader have evolved over time. We've come to understand more clearly how a Read method should behave, and we've tried to capture that in the io.Reader documentation, but at the same time we don't want to break existing types that already have Read methods.

This proposal has been added to the active column of the proposals project
and will now be reviewed at the weekly proposal review meetings.
— rsc for the proposal review group

I don't see a way to do this compatibly. There exist implementations that can return zero-length successful reads. Any reader that is reading and stopping on packet boundaries might do that, for example. This is another example of something that we can't easily change without invalidating all the implementations in the wild, which means we can't really change it.

Based on the discussion above, this proposal seems like a likely decline.
— rsc for the proposal review group

No change in consensus, so declined.
— rsc for the proposal review group

If returning 0, nil is a valid Reader implementation, shouldn't bufio.maxConsecutiveEmptyReads be removed because it doesn't combine well with such readers?

A reader that returns 0, nil is valid, but a reader that returns it 100 times in a row probably has something wrong with it.

A reader that returns 0, nil is valid, but a reader that returns it 100 times in a row probably has something wrong with it.

@ianlancetaylor what about when working with CGO and compression libraries. I was under the impression you want to avoid blocking the OS thread and I'd see this as a valid use-case where it could return 0, nil 100+ times and it remain valid.

For example this library https://github.com/DataDog/zstd

Edit: Wanted to add if there's a better approach of some sort, I'm more than open to implementing it. Just from my understanding the C library itself does not have a way to safely block and hook that with Go.

@ethan-gallant You possibly should create an issue in DataDog/zstd (ideally, with a repro). I agree with @ianlancetaylor, returning 0-bytes 100 times in a row is suspicious. Normally, compressing library reads uncompressed data from input stream, does its compression magic and writes result to the output. Such logic is not supposed to produce many 0-bytes results.

Some compression formats have the concept of a "sync flush" where a special block is a signal to the decompressor that it should report all uncompressed data to the application. If a decompression stream contains a series of back-to-back sync flush blocks without any data in-between, I would expect it to still read (0, nil) a number of times repeatedly.

Some compression formats have the concept of a "sync flush" where a special block is a signal to the decompressor that it should report all uncompressed data to the application. If a decompression stream contains a series of back-to-back sync flush blocks without any data in-between, I would expect it to still read (0, nil) a number of times repeatedly.

Agreed, additionally 100 is a very ambiguous catch-all case. What makes 100 so significant here, why not 1000? Who's to say how frequently a consumer may call read or the frequency at which they do it.

My opinion is that we should leave bufio as is unless and until somebody finds a real world problem. The current code was worked out in the discussion on https://go.dev/cl/76400048.

@ianlancetaylor the real-world problem is that using bufio.Reader with implementations that return 0, nil results in an unrecoverable error.

While the documentation says "this is discouraged" it does not state it is unsupported.

@ethan-gallant What I mean by a real-world problem is a real existing Go program that breaks because of how bufio currently behaves, and would be fixed if bufio's behavior changes. I would also be willing to accept a real existing Go program that has extra code to work around the current bufio behavior.

A Reader that returns 0, nil upon occasion will not lead to an unrecoverable error when using bufio. A problem will only arise with a Reader that returns 0, nil 100 times in a row. Do any such implementations of Reader actually exist in real Go programs outside of tests? If not, it's not worth worrying about.

@ethan-gallant What I mean by a real-world problem is a real existing Go program that breaks because of how bufio currently behaves, and would be fixed if bufio's behavior changes. I would also be willing to accept a real existing Go program that has extra code to work around the current bufio behavior.

A Reader that returns 0, nil upon occasion will not lead to an unrecoverable error when using bufio. A problem will only arise with a Reader that returns 0, nil 100 times in a row. Do any such implementations of Reader actually exist in real Go programs outside of tests? If not, it's not worth worrying about.

This library when compressing large images and reading them with bufio.Reader produces the error. From what I gather it's related to how the library interacts with CGO and the ZStd library.

Not sure if there's an alternate way, but the zstd library does have a similar implementation that doesn't translate well into Go bindings for this reason.

Link: https://github.com/DataDog/zstd

When in actual practice are those packages going to return 0, nil 100 times in a row?

When in actual practice are those packages going to return 0, nil 100 times in a row?

Whenever you're compressing a large object, like a database backup or container image that exceeds 600MB on a slower CPU.

I'm sorry, I don't understand how that could happen. Can you show an example? Thanks.

An example is a reader that returns 0, nil without blocking. I'm a bit short on time and making a short example is a little tedious (writer, buffer, reader, chunk size, cgo) but I can guarantee it is an issue.

Maybe it's isolated to https://github.com/DataDog/zstd but regardless, it's documented that returning 0, nil is a valid reader implementation, it is simply "discouraged" so I'd assume internals packages like this would work alongside that assumption.

As I understand, @ianlancetaylor wants a real-world usecase, not an artificially crafted reader that would infinitely return 0, nil.

If you say that https://github.com/DataDog/zstd is capable of doing that, do you have a repro code?