core.async - handling nils

[Mikera]

Hi all,

I'm experimenting with core.async. Most of it is exceptionally good, but bit I'm finding it *very* inconvenient that nil can't be sent over channels. In particular, you can't pipe arbitrary Clojure sequences through channels (since sequences can contain nils). 

I see this as a pretty big design flaw given the ubiquity of sequences in Clojure code - it appears to imply that you can't easily compose channels with generic sequence-handling code without some pretty ugly special-case handling.

Am I missing something? Is this a real problem for others too? 

If it is a design flaw, can it be fixed before the API gets locked down?

[Brandon Bloom]

I ran into the other half of this problem: If you expect nils to signify closed channels, then you can't leverage the logically false nature of nil without excluding explicit boolean false values. Given the pleasant syntax of if-let / <! pairs, I reworked my early experiments to use if-recv which is defined as follows:

(defmacro if-recv
  "Reads from port, binding to name. Evaluates the then block if the
  read was successful. Evaluates the else block if the port was closed."
  ([[name port :as binding] then]
   `(if-recv ~binding ~then nil))
  ([[name port] then else]
   `(let [~name (<! ~port)]
      (if (nil? ~name)
        ~else
        ~then))))

I've considered some alternative core.async designs, such as an additional "done" sentinel value, or a pair of quote/unquote operators (see "reduced"), but nothing seems as simple as just avoiding booleans and nils, as annoying as that is. I'd be curious to here what Rich & team considered and how they're thinking about it. However, my expectation is that the nil approach won't change, since it's pretty much good enough.

[Ben Wolfson]

On Fri, Aug 16, 2013 at 4:50 PM, Brandon Bloom <brandon...@gmail.com> wrote:

I ran into the other half of this problem: If you expect nils to signify closed channels, then you can't leverage the logically false nature of nil without excluding explicit boolean false values. Given the pleasant syntax of if-let / <! pairs, I reworked my early experiments to use if-recv which is defined as follows:

(defmacro if-recv
  "Reads from port, binding to name. Evaluates the then block if the
  read was successful. Evaluates the else block if the port was closed."
  ([[name port :as binding] then]
   `(if-recv ~binding ~then nil))
  ([[name port] then else]
   `(let [~name (<! ~port)]
      (if (nil? ~name)
        ~else
        ~then))))

I've considered some alternative core.async designs, such as an additional "done" sentinel value, or a pair of quote/unquote operators (see "reduced"), but nothing seems as simple as just avoiding booleans and nils, as annoying as that is. I'd be curious to here what Rich & team considered and how they're thinking about it. However, my expectation is that the nil approach won't change, since it's pretty much good enough.

One possibility is to retain nil as "closed" and have every other value come wrapped in a Just or Some-like constructor, which an if-recv macro would unwrap transparently, something like:

(defmacro if-recv [[name port] then else]

   `(let [~name (<! ~port)]
      (if (nil? ~name)
          ~else

        (let [~name (unwrap ~name)]
          ~then))))

That would enable sending of nils, since on the far side they'd be wrapped.

--
Ben Wolfson
"Human kind has used its intelligence to vary the flavour of drinks, which may be sweet, aromatic, fermented or spirit-based. ... Family and social life also offer numerous other occasions to consume drinks for pleasure." [Larousse, "Drink" entry]

[Brandon Bloom]

> have every other value come wrapped in a Just or Some-like constructor

That's what I meant by "a pair of quote/unquote operators"

[Ben Wolfson]

On Fri, Aug 16, 2013 at 5:07 PM, Brandon Bloom <brandon...@gmail.com> wrote:

> have every other value come wrapped in a Just or Some-like constructor

That's what I meant by "a pair of quote/unquote operators"

 ah, gotcha

[Mikera]

My overall sense is that the convenience of using if-let directly in a few use cases doesn't justify making channels fall short of being able to send arbitrary values (nil specifically, and clearly boolean false can cause some problems too). 

I think it would be a much better design to have a sentinel value and a couple of specialised functions or macros that can detect  / interact with it appropriately. With a sentinel value the key part of your if-recv code could just be something like:

`(let [~name (<! ~port)]

      (if (end-of-stream? ~name)
        ~else
        ~then))))

I can see that wrappers for nil values could also work, but that seems to be a more complex solution (and also potentially with more overhead) than a sentinel value....

[Brandon Bloom]

That's precisely the design followed by Magpie, described here: http://journal.stuffwithstuff.com/2013/04/17/well-done/

Parts 1 & 2 of that series are worth reading too.

--
--
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to clo...@googlegroups.com
Note that posts from new members are moderated - please be patient with your first post.
To unsubscribe from this group, send email to
clojure+u...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
---
You received this message because you are subscribed to a topic in the Google Groups "Clojure" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/clojure/pF9FEP7b77U/unsubscribe.
To unsubscribe from this group and all its topics, send an email to clojure+u...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[Ben Wolfson]

A sentinel value also prevents channels from being able to send/receive arbitrary values, without further wrapping.

Sent from my iPhone

--

--
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to clo...@googlegroups.com
Note that posts from new members are moderated - please be patient with your first post.
To unsubscribe from this group, send email to
clojure+u...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
---

You received this message because you are subscribed to the Google Groups "Clojure" group.
To unsubscribe from this group and stop receiving emails from it, send an email to clojure+u...@googlegroups.com.

[Mikera]

If you use a singleton sentinel value that is generated privately within the core.async implementation, then the sentinel isn't really a regular "value" in the sense that it can be created by regular user code.

nil, on the other hand, gets used very frequently as a value in regular Clojure code.

That's a key reason IMHO why it's better to use a sentinel as a closed channel indicator than nil.

[Max Penet]

It's a real problem for me too, I also wonder what was the intention behind this. I guess there could be a very good reason for this special treatement of nils, but I haven't seen it yet. 

I would love to hear about this from people involved in core.async development.

[Timothy Baldridge]

The reason for not allowing nils isn't a complex one, and basically boils down to the following:

a) to avoid race conditions, we need a single value to signal "the channel is closed". As mentioned, nil is the obvious choice for this as it matches lazy seqs and fits well with the rest of clojure:

(when-let [v (<! c)]
  (process v))

If we chose a different value, this becomes much more ugly:

(let [v (<! c)]
  (when-not (= v :async/closed)
    (process v)))

b) I question if there are any valid uses for putting nil in a channel. With all due respect to all who have written here, thus far, every complaint about nils and channels boils down to a conversion from seqs to channels. This is the wrong way to look at the problem. Channels are co-ordination primitives not data structures. Simply because a lazy seq looks like a channel, doesn't mean that they should be treated as such. 

In all the core.async code I've written I've never had to put a nil in a channel, so I'm left with the uncomfortable conclusion that most complaints on this subject are contrived. I could be wrong, but I just haven't seen a valid use case yet. 


This all being said, there really isn't a technical reason to not allow nils, it just simplifies much of the design and that probably translates to better performance. So the restriction could be lifted if a rock solid reason could be found, but as of yet, I haven't seen it. 

Timothy Baldridge

--

--
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to clo...@googlegroups.com
Note that posts from new members are moderated - please be patient with your first post.
To unsubscribe from this group, send email to
clojure+u...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
---
You received this message because you are subscribed to the Google Groups "Clojure" group.
To unsubscribe from this group and stop receiving emails from it, send an email to clojure+u...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

--
“One of the main causes of the fall of the Roman Empire was that–lacking zero–they had no way to indicate successful termination of their C programs.”
(Robert Firth)

[Timothy Baldridge]

All your arguments come down to this:

"I have an arbitrary seq of things I want to send down a channel". It's exactly that concept I that I push against. Everything you've mentioned thus far is a data structure. Channels are not data structures they are concurrency management primitives, treat them as such and I doubt you'll ever have a need for nils in a channel. 

If we treat channels as ways of co-ordinating concurrent processes, then nil doesn't have a use case. In every use of channels I've had thus far, nil is better expressed as an empty collection, false, 0, :tick, or some other "ground value".

It's these Rx style programming methods that make people think they need this feature. 

Timothy 

On Tue, Aug 27, 2013 at 8:51 AM, Mike Anderson <mike.r.an...@gmail.com> wrote:

On 27 August 2013 20:45, Timothy Baldridge <tbald...@gmail.com> wrote:

The reason for not allowing nils isn't a complex one, and basically boils down to the following:

a) to avoid race conditions, we need a single value to signal "the channel is closed". As mentioned, nil is the obvious choice for this as it matches lazy seqs and fits well with the rest of clojure:

Agreed that you want a single sentinel value.

It doesn't match lazy-seqs at all though: lazy seqs can contain nils just fine. There's a big difference between (next some-lazy-seq) [which could be nil, indicating an empty sequence] and the actual values in the seq [which could also be nil but don't indicate the end of the seq].

 

(when-let [v (<! c)]
  (process v))

If we chose a different value, this becomes much more ugly:

(let [v (<! c)]
  (when-not (= v :async/closed)
    (process v)))

This can be solved easily by providing a macro or some other predicate that knows how to check for the sentinel value correctly. e.g.

(when-more [v (<! c)]

  (process v))

 

b) I question if there are any valid uses for putting nil in a channel. With all due respect to all who have written here, thus far, every complaint about nils and channels boils down to a conversion from seqs to channels. This is the wrong way to look at the problem. Channels are co-ordination primitives not data structures. Simply because a lazy seq looks like a channel, doesn't mean that they should be treated as such. 

In all the core.async code I've written I've never had to put a nil in a channel, so I'm left with the uncomfortable conclusion that most complaints on this subject are contrived. I could be wrong, but I just haven't seen a valid use case yet. 

To me it's all about consistency with other Clojure constructs. You can safely put nils in sequences, vectors, lists, sets etc.. nil is a valid "value" just like anything else. So why can't you put them in a channel?

Two use cases I have encountered that motivate this:

a) what if you want to send a sequence through a channel? Since nil as a value represents the empty sequence, you have to put in some extra special case handling with the current core.async model. 

 

b) what if you want to write generic code to send all the values in an arbitrary collection through a channel? you would have to wrap/unwrap nils at either end to make this work currently.

Both of these, I think, are reasonable and common enough use cases that it's worth supporting them elegantly rather than forcing users to implement their own nil-wrapping functionality.

This all being said, there really isn't a technical reason to not allow nils, it just simplifies much of the design and that probably translates to better performance. So the restriction could be lifted if a rock solid reason could be found, but as of yet, I haven't seen it. 

I don't believe there is any noticeable performance difference between checking for nil and checking if a value is identical? to some sentinel value (which would presumably be static, final, immutable and hence very well optimised by the JVM). In addition, not allowing nils just means you have to do extra work to wrap/unwrap nils as a user - which is almost certainly a net loss on overall performance.

Still, I think consistency is more significant than the performance argument in this case.

 

You received this message because you are subscribed to a topic in the Google Groups "Clojure" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/clojure/pF9FEP7b77U/unsubscribe.
To unsubscribe from this group and all its topics, send an email to clojure+u...@googlegroups.com.

For more options, visit https://groups.google.com/groups/opt_out.

--
--
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to clo...@googlegroups.com
Note that posts from new members are moderated - please be patient with your first post.
To unsubscribe from this group, send email to
clojure+u...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
---
You received this message because you are subscribed to the Google Groups "Clojure" group.
To unsubscribe from this group and stop receiving emails from it, send an email to clojure+u...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[David Nolen]

On Tue, Aug 27, 2013 at 10:51 AM, Mike Anderson <mike.r.an...@gmail.com> wrote:

To me it's all about consistency with other Clojure constructs. You can safely put nils in sequences, vectors, lists, sets etc.. nil is a valid "value" just like anything else. So why can't you put them in a channel?

Channels are *not* data structures nor are they a "place" to put something. 

 

a) what if you want to send a sequence through a channel? Since nil as a value represents the empty sequence, you have to put in some extra special case handling with the current core.async model.

You're not going to put random sequences into channels. Channels are conduits for meaningful messages - some well considered coordination protocol.

 

Both of these, I think, are reasonable and common enough use cases that it's worth supporting them elegantly rather than forcing users to implement their own nil-wrapping functionality.

If you're putting arbitrary sequences into a channel and need to wrap nils, you probably need to redesign your coordination protocol.

David

[Brandon Bloom]

> In every use of channels I've had thus far, nil is better expressed as an empty collection, false, 0, :tick, or some other "ground value". 

I agree completely. But I'll note that you mention false being useful...

If you're writing completely general operators, like map, which are *sometimes* quite useful, then you have no choice but to do something like if-recv or explicitly test against nil.

> It's these Rx style programming methods that make people think they need this feature. 

I built my little Rx with channels library (asyncx) without intention to use it directly, but because I wanted to learn how to work with channels. I rapidly learned that the techniques are a lot more different than they look. In particular, it's more difficult to write channel & process combinators precisely because they are more powerful. However, in practice, each new reusable channel/process combinator yields more complexity than it tends to save. I'd rather intentionally choose strictly less powerful primitives where appropriate and enforce that with encapsulation.

With that in mind, if I ever revisit asyncx, I'll probably define "push sequences" or "streams" in terms of protocols and deftypes. I'd use core.async to implement them, but only for the lowest level primitives. I'd provide ways to get in to or out of the stream subsystem for interop with channels, but the public interface would take IStream objects.

[Timothy Baldridge]

Right, the use of false is a special case. I'm thinking of a mouse event stream, may have a button channel that sends true or false based on the state of the mouse button. Even saying that though, I would probably opt for :clicked and :unclicked or somethig of that nature. 

Timothy

[Mike Anderson]

On 27 August 2013 20:45, Timothy Baldridge <tbald...@gmail.com> wrote:

The reason for not allowing nils isn't a complex one, and basically boils down to the following:

a) to avoid race conditions, we need a single value to signal "the channel is closed". As mentioned, nil is the obvious choice for this as it matches lazy seqs and fits well with the rest of clojure:

Agreed that you want a single sentinel value.

It doesn't match lazy-seqs at all though: lazy seqs can contain nils just fine. There's a big difference between (next some-lazy-seq) [which could be nil, indicating an empty sequence] and the actual values in the seq [which could also be nil but don't indicate the end of the seq].

 

(when-let [v (<! c)]

  (process v))

If we chose a different value, this becomes much more ugly:

(let [v (<! c)]
  (when-not (= v :async/closed)
    (process v)))

This can be solved easily by providing a macro or some other predicate that knows how to check for the sentinel value correctly. e.g.

(when-more [v (<! c)]

  (process v))

 

b) I question if there are any valid uses for putting nil in a channel. With all due respect to all who have written here, thus far, every complaint about nils and channels boils down to a conversion from seqs to channels. This is the wrong way to look at the problem. Channels are co-ordination primitives not data structures. Simply because a lazy seq looks like a channel, doesn't mean that they should be treated as such. 

In all the core.async code I've written I've never had to put a nil in a channel, so I'm left with the uncomfortable conclusion that most complaints on this subject are contrived. I could be wrong, but I just haven't seen a valid use case yet. 

To me it's all about consistency with other Clojure constructs. You can safely put nils in sequences, vectors, lists, sets etc.. nil is a valid "value" just like anything else. So why can't you put them in a channel?

Two use cases I have encountered that motivate this:

a) what if you want to send a sequence through a channel? Since nil as a value represents the empty sequence, you have to put in some extra special case handling with the current core.async model. 

 

b) what if you want to write generic code to send all the values in an arbitrary collection through a channel? you would have to wrap/unwrap nils at either end to make this work currently.

Both of these, I think, are reasonable and common enough use cases that it's worth supporting them elegantly rather than forcing users to implement their own nil-wrapping functionality.

This all being said, there really isn't a technical reason to not allow nils, it just simplifies much of the design and that probably translates to better performance. So the restriction could be lifted if a rock solid reason could be found, but as of yet, I haven't seen it. 

I don't believe there is any noticeable performance difference between checking for nil and checking if a value is identical? to some sentinel value (which would presumably be static, final, immutable and hence very well optimised by the JVM). In addition, not allowing nils just means you have to do extra work to wrap/unwrap nils as a user - which is almost certainly a net loss on overall performance.

Still, I think consistency is more significant than the performance argument in this case.

 

Timothy Baldridge

You received this message because you are subscribed to a topic in the Google Groups "Clojure" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/clojure/pF9FEP7b77U/unsubscribe.
To unsubscribe from this group and all its topics, send an email to clojure+u...@googlegroups.com.

[Softaddicts]

+1

We built a distributed software sending/receiving *messages* based on
different protocols.

All our protocols wrap data in an envelope. The receiver can then decide how to
handle the message based on the envelope. Obviously, nil makes a bad envelope.

A nil message on a channel never had any significance to us four years ago
and rethinking about it we reach the same conclusion today :)


Luc P.

> --
> --
> You received this message because you are subscribed to the Google
> Groups "Clojure" group.
> To post to this group, send email to clo...@googlegroups.com
> Note that posts from new members are moderated - please be patient with your first post.
> To unsubscribe from this group, send email to
> clojure+u...@googlegroups.com
> For more options, visit this group at
> http://groups.google.com/group/clojure?hl=en
> ---
> You received this message because you are subscribed to the Google Groups "Clojure" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to clojure+u...@googlegroups.com.
> For more options, visit https://groups.google.com/groups/opt_out.
>
--

Softaddicts<lprefo...@softaddicts.ca> sent by ibisMail from my ipad!

[Ben Wolfson]

On Tue, Aug 27, 2013 at 7:51 AM, Mike Anderson <mike.r.an...@gmail.com> wrote:

On 27 August 2013 20:45, Timothy Baldridge <tbald...@gmail.com> wrote:

The reason for not allowing nils isn't a complex one, and basically boils down to the following:

a) to avoid race conditions, we need a single value to signal "the channel is closed". As mentioned, nil is the obvious choice for this as it matches lazy seqs and fits well with the rest of clojure:

Agreed that you want a single sentinel value.

It doesn't match lazy-seqs at all though: lazy seqs can contain nils just fine. There's a big difference between (next some-lazy-seq) [which could be nil, indicating an empty sequence] and the actual values in the seq [which could also be nil but don't indicate the end of the seq].

And the when-first macro correctly handles this! It's not just `(when-let [~x (first ~xs)] ~@body). when-first nicely hides away what would otherwise be "much more ugly".

 

[Brandon Bloom]

As long as we don't go full Haskell mode:

data Message a = Value a | Done

[Mike Anderson]

I still don't see why you would want to to arbitrarily limit what you can put down a channel. FWIW, plenty of other concurrency management primitives allow nils as values (java.util.concurrent.AtomicReference, Clojure atoms / refs / agents to name but a few).

My motivating use case is the ability to create higher order constructs that communicate via channels, as a way of gluing concurrent processes together. A simplified example:

(defn printer [ch id]

  (go (while true

        (let [v (<! ch)]

          (prn (str "Printer " id " handled value: " v))))))

(defn sender [ch]

     (fn [xs] (go (doseq [x xs] (>! ch x)))))

(let [ch (chan)

      pr1 (printer ch "1")

      pr2 (printer ch "2")

      sendr (sender ch)]

  (sendr ["foo" "a"])

  (sendr ["bar"]))

Using nil as a sentinel appears to prevent such constructs from working with arbitrary Clojure values (or alternatively forces extra wrapping / special case handling that adds complexity and overhead). Furthermore, if different libraries start adopting different protocols / techniques for wrapping nils then the composability of such constructs will be severely restricted.

I may be missing something, but this seems like a reasonable use case for core.async to support?

Of course, if anyone has an actual technical argument why it is necessary/better to use nil as a sentinel value, I would be delighted to learn of it and would consider myself enlightened. 

[guns]

On Wed 28 Aug 2013 at 09:38:06AM +0800, Mike Anderson wrote:

> Of course, if anyone has an actual technical argument why it is
> necessary/better to use nil as a sentinel value, I would be delighted to
> learn of it and would consider myself enlightened.

Forgive me if someone already mentioned this, but isn't this simply a
consequence of building on the existing Java Queue implementations?

Quoting http://docs.oracle.com/javase/7/docs/api/java/util/Queue.html:

Queue implementations generally do not allow insertion of null
elements, although some implementations, such as LinkedList, do not
prohibit insertion of null. Even in the implementations that permit
it, null should not be inserted into a Queue, as null is also used
as a special return value by the poll method to indicate that the
queue contains no elements.

I think most would agree with you that a configurable sentinel value
is best (I like using ::namespaced-keywords myself), but the existing
machinery has already decided on null, so that's what we have.

guns

[Mikera]

Well, that's certainly a good explanation of why core.async works the way it does now - it's a natural and sensible starting point to build on Java Queues.

I don't think that this necessarily implies that we have to follow this model in the future Clojure API though. The Java designers didn't always get everything right :-)

There's no reason I can see why core.async can't internally substitute its own nil-representing value into the underlying Java infrastructure. And this would free users of core.async from the burden of doing this translation every time that the transmission of nil values is desired.

As for "configurable sentinel value": I'm less sure that there's a need for this. In fact, I think it's more useful to guarantee a standard sentinel value so that different channel-handling code can interoperate. I just think that this sentinel shouldn't be nil (or any other regular Clojure value). A simple singleton object that is privately owned by core.async should suffice?

[guns]

On Tue 27 Aug 2013 at 07:42:53PM -0700, Mikera wrote:

> On Wednesday, 28 August 2013 09:54:51 UTC+8, guns wrote:
> >
> > On Wed 28 Aug 2013 at 09:38:06AM +0800, Mike Anderson wrote:
> >
> > > Of course, if anyone has an actual technical argument why it is
> > > necessary/better to use nil as a sentinel value, I would be delighted to
> > > learn of it and would consider myself enlightened.
> >
> > Forgive me if someone already mentioned this, but isn't this simply a
> > consequence of building on the existing Java Queue implementations?
> >
> > Quoting http://docs.oracle.com/javase/7/docs/api/java/util/Queue.html:
> >
> > Queue implementations generally do not allow insertion of null
> > elements, although some implementations, such as LinkedList, do not
> > prohibit insertion of null. Even in the implementations that permit
> > it, null should not be inserted into a Queue, as null is also used
> > as a special return value by the poll method to indicate that the
> > queue contains no elements.
>

> I don't think that this necessarily implies that we have to follow
> this model in the future Clojure API though. The Java designers didn't
> always get everything right :-)

I agree with you that this is an inconvenience. The bit about avoiding
null even in collections that support null elements because poll uses
null as a sentinel value is a classic leaky abstraction.

That said, I haven't found designing around it to be a terrible
nuisance. I often use queues as messaging channels, so it is natural to
simply send maps or tuples instead of atomic values.

> As for "configurable sentinel value": I'm less sure that there's
> a need for this. In fact, I think it's more useful to guarantee a
> standard sentinel value so that different channel-handling code can
> interoperate. I just think that this sentinel shouldn't be nil (or
> any other regular Clojure value). A simple singleton object that is
> privately owned by core.async should suffice?

Oh, I was confused; I was thinking about sentinel values in user code.
Yes, I imagine a single private (Object.) would work just fine, with
very little overhead.

guns

[Alan Busby]

On Wed, Aug 28, 2013 at 12:18 PM, guns <se...@sungpae.com> wrote:

Oh, I was confused; I was thinking about sentinel values in user code.
Yes, I imagine a single private (Object.) would work just fine, with
very little overhead.

First, I'd hope that sentinel values would be handled by the back-end 

implementation, as we're seeing core.sync implemented on other 

systems like ZeroMQ already.

Second, as (Object.) doesn't play nicely over the wire, a random UUID

or similar value would be much preferred. 

Third, I'd recommend reviewing,

http://clojure.com/blog/2013/06/28/clojure-core-async-channels.html

to understand why core.async is not just a better queue.

Fourth, if you dislike how core.async works, just wrap it in your own

library so it works the way you'd like. 

It looks like "core.async-with-nil" is available on Clojars. ;)

With nil, without nil, it's just bike shedding. Clojure gives you the

freedom to do it the way you want.

[guns]

On Wed 28 Aug 2013 at 12:50:19PM +0900, Alan Busby wrote:

> On Wed, Aug 28, 2013 at 12:18 PM, guns <se...@sungpae.com> wrote:
>
> > Oh, I was confused; I was thinking about sentinel values in user
> > code. Yes, I imagine a single private (Object.) would work just
> > fine, with very little overhead.
>

> Third, I'd recommend reviewing,
> http://clojure.com/blog/2013/06/28/clojure-core-async-channels.html to
> understand why core.async is not just a better queue.

In my mind, core.async is a nice marriage of BlockingQueues, async
processing, and a supercharged POSIX select(2) that works on queue
objects. Nothing about these ideas necessarily requires that null be
reserved by the channel implementation.

> Fourth, if you dislike how core.async works, just wrap it in
> your own library so it works the way you'd like. It looks like
> "core.async-with-nil" is available on Clojars. ;)
>
> With nil, without nil, it's just bike shedding. Clojure gives you the
> freedom to do it the way you want.

Rich Hickey, from: http://clojure.com/blog/2013/06/28/clojure-core-async-channels.html

> While the library is still in an early state, we are ready for people
> to start trying it out and giving us feedback.

I think mikera is trying to be constructive.

For my own part, I am quite ambivalent since I am already used to
avoiding pushing nulls onto Queues. I am quite happy to accept that this
is an implementation detail and move on, but I can also see why it might
be worth it to support nil channel values to avoid confusing users that
are not familiar with this quirk of java.util.Queue.

guns

[Mike Anderson]

On 28 August 2013 11:50, Alan Busby <theb...@gmail.com> wrote:

On Wed, Aug 28, 2013 at 12:18 PM, guns <se...@sungpae.com> wrote:

Oh, I was confused; I was thinking about sentinel values in user code.
Yes, I imagine a single private (Object.) would work just fine, with
very little overhead.

First, I'd hope that sentinel values would be handled by the back-end 

implementation, as we're seeing core.sync implemented on other 

systems like ZeroMQ already.

Second, as (Object.) doesn't play nicely over the wire, a random UUID

or similar value would be much preferred. 

Hi Alan,

Agreed on all the above. The issue is not so much what sentinel value is used internally, but what sentinel value gets exposed to user code.

 

Third, I'd recommend reviewing,

http://clojure.com/blog/2013/06/28/clojure-core-async-channels.html

to understand why core.async is not just a better queue.

Fourth, if you dislike how core.async works, just wrap it in your own

library so it works the way you'd like. 

It looks like "core.async-with-nil" is available on Clojars. ;)

That's precisely what I'm trying to avoid, and the reason why I've been raising the topic here - the last thing we want in the ecosystem is more fragmentation with incompatible subsystems and protocols. That's the Lisp Curse all over again. I think we should aspire to better in the Clojure community - which means working together to make the best implementation possible and rallying around it.

Sending nil as a value over channels is clearly a significant issue if people are willing to fork or create a new wrapper for core.async in order to achieve it. Better, in my view, to make a breaking change to core.async now to fix this issue rather than encouraging a free-for-all.

I'm reminded of Rich Hickey's keynote "The Language of the System", which emphasised composing systems out of simple services that communicate via values. core.async would IMHO be a much more useful tool for realising this sort of vision if it is able transmit arbitrary Clojure values (i.e. including nil).