math/rand: Read should return consistent results, regardless of buffer size

[pin]

Read returns byte sequence that depends on the buffer length. That's because Rand uses Int63 value to produce 7 random bytes and throws away unused bits from Int63 when done with the particular call. Next call always starts with fresh Int63 leaving "gap" from previous Read if it was not finished on 7-bytes boundary.

Therefore, you can't fix seed and read pseudo-random bytes deterministically without fixing buffer size.

Proposed fix

My proposal is to keep unused Int63 reminder in the Rand struct. It let us fix the issue and still have the same bytes from the first Read call as we have now. The performance hit is minimal also. The downside is an extra state in Rand struct.

One of the other options is to use only single byte from each Int63 call. No state in Rand but the 2x-4x performance hit.

1. What version of Go are you using (go version)?
   go version devel +f3d5478 Fri Jun 17 19:15:29 2016 +0000 linux/amd64
2. What operating system and processor architecture are you using (go env)?

GOARCH="amd64"
GOOS="linux"

1. What did you do?
   This is play-adopted unit-test using testing/iotest that fails: https://play.golang.org/p/dbz-VlFPCa
2. What did you expect to see?
   The same random bytes stream no matter read buffer size used.

[adg]

Seems like something that's easily fixed.

The one concern I see is compatibility. If someone relies on this broken behavior, their program may cease to function correctly. However, I don't think we have made any promises about the stability of math/rand as a pseudo-random source. And this is certainly a bug. So the compatibility concern is probably not relevant.

[randall77]

I don't see the point behind this proposal. If you're giving math/rand a seed (and the same seed multiple times in different runs), it is because you're trying to make some nondeterministic code deterministic. But surely you're also issuing the same sized reads in that case?

Therefore, you can't fix seed and read pseudo-random bytes deterministically without fixing buffer size.

In other words, why would you not fix the buffer size if you were trying to make a computation deterministic?

[gopherbot]

CL https://golang.org/cl/24251 mentions this issue.

[dsnet]

Issue #8013 has some discussion about whether we could change the output of the rand package. The Read method was introduced in Go1.6, so the number of users dependent on it may be low.

That said, if we do change it, should we consider this change for Go1.7 to reduce the amount of time the faulty behavior stays present? I'm almost certain that this change will break some code. The longer people are using and relying on the current behavior, the worse it will be later.

[josharian]

The same random bytes stream no matter read buffer size used.

I'm with @randall77; it's not clear to me why this needs to be the case. Is there a concrete context where you need this?

[pin]

@randall77 I would like to have deterministic sequence with known distribution but in the same time luxury of dealing with io.Reader which allows me to take buffer size out of the equation and combine/nest/wrap it without worrying about buffer size.

Say io.CopyN one instance to hash.Hash implementation, then send stream produced with other instance (made with the same seed) over the network and write it to the same hash.Hash implementation on the other end. Buffer size will differ for both cases, but I would expect to get the same hash sum.

This is the natural way of thinking for me:

• PRNG generates a sequence of random values that is determined by a seed
• io.Reader is a way to read/receive a sequence of bytes, a buffer is just the way to get the data
• io.Reader implemented by rand.Rand is a way to generate sequence of bytes determined by a seed

@josharian The reasonable expectation I believe is that, for instance, io.LimitReader(rand.New(rand.NewSource(seed), length) is the way to produce pseudo-random byte stream of given length and that's content is determined by the seed only.

So that one could send that stream over the network, store on disk, etc. Then afterwards read it back (i.e. get io.Reader with the same stream but now coming from the network, disk, etc), create the second stream using io.LimitReader/rand.Reader with the same seed/length and validate data (read both streams in turn and compare bytes).

It is convenient because one can use arbitrary-sized streams that could exceed RAM size, have many of them and just keep seed and length in order to be able to validate content. It is sound way to do some kinds of testing of network device, storage service, etc. And doing such testing one can't always fix buffer size because it could be determined by say network buffers (or say RPC mechanism and validator implementation could have different buffers).

One of the real-world examples is TFTP library unit-tests, see https://github.com/pin/tftp/blob/master/tftp_test.go#L126 for instance. There is roll-your-own io.Reader wrapper around rand.Source exactly because of the this issue: when I implement validator I don't want to take TFTP library buffer size into the equation.

In general, I would expect that provided the same seed and calling no other methods on Rand except Read (and wrapping it into something like io.LimitReader actually enforces that) I would get the same stream of bytes no matter how io.Reader is read.

[adg]

I agree with @dsnet that this might be worth getting in for 1.7.

[randall77]

I see. It does sound useful to do io.LimitReader(rand.New(rand.NewSource(seed), length) to generate a deterministic random-looking stream.

One way to get the determinism you want would be to wrap rand.NewSource with a iotest.OneByteReader. And then probably a bufio.Reader for efficiency. But you've convinced me, I'm ok with doing this directly in rand.

[josharian]

I am a bit concerned about the backwards compatibility still. We've declined to do reasonable things with rand in the past. I wonder whether it would be preferable to do this in 1.8 and fix those other bugs as well, all in one release. If this were introduced in 1.7 the case for changing it now would be clear.

[randall77]

Another case where we decided not to change math/rand because of compatibility issues:

#13215