RFR: JDK-8205461 Create Collector which merges results of two other collectors

Stuart Marks stuart.marks at oracle.com
Fri Sep 14 17:09:04 UTC 2018

Hi Tagir, thanks for the update.

Also thanks Tomasz for keeping everybody honest on the open issues.

First, naming. I think "duplex" as the root word wins! Using "duplexing" to 
conform to many of other collectors is fine; so, "duplexing" is good.

Unfortunately "duplex" is not really a verb. For some reason "duplexing" sounds 
OK. This probably follows in the long tradition of the tech industry's "verbing" 
of nouns and adjectives. However, I can't quite bring myself to recommend using 
this in the javadoc summary sentence, e.g.,

    Returns a {@code Collector} that duplexes stream elements to two downstream

Ugh. I think the current summary sentence is fine

    Returns a {@code Collector} that is a composite of two downstream collectors.

even though it uses "composite" which is not the name of this collector. But I 
think it's a more descriptive term and it reads more smoothly.

Turning to the issues mentioned by Tomasz:

> 1) Brian Goetz' suggestion of changing "? extends R" into "R":
>   - http://mail.openjdk.java.net/pipermail/core-libs-dev/2018-August/054947.html
Yes, I think this should be done.
> 2) Stuart Marks' suggestion about renaming "c1" and "c2" to "downstream1" and 
> "downstream2":
>   - http://mail.openjdk.java.net/pipermail/core-libs-dev/2018-August/054949.html
Yes. It's a small thing, but the parameter names do appear in the javadoc. The 
text refers to "downstream" collectors, and naming the parameters "downstream" 
strengthens the association to the text. Otherwise, the reader has to think 
"what are c1 and c2? Oh, they're the downstream collectors."
> 3) my suggestion about renaming "merger" to "biFinisher" because "merger" 
> means BinaryOperator everywhere else:
>   - 
> http://mail.openjdk.java.net/pipermail/core-libs-dev/2018-September/055235.html
I'm ok with "merger". I don't feel that "merge" has a strong implication that 
the result must be the same type as the two inputs. I'm ok with the notion of 
"merge" taking two inputs (which might be of different types) and producing a 
single output (which might be a different type from either input). The 
difficulty with "biFinisher" is that it doesn't seem to imply "take two things 
and produce one". A finisher takes one input and produces one output, so would a 
"biFinisher" take two inputs and produce two outputs?
> As to the name, I still think a name related to a well-understood concept 
> (like the composite pattern [1]) would be better. Note that "composite" is 
> also a verb [2], but "Collectors.compositing" looks a bit strange. 
> "Collectors.composing" seems much better to me, but - as far as I understand - 
> there was some concern that the users could misunderstand it as element-wise 
> composition, is that right?
The difficulty I have with "compose" and "composition" is that with function 
composition, one function is applied, and the second function is applied to the 
result of the first. Of course that isn't what happens here. The notion of 
"duplexing" is that things are happening side-by-side, which applies well here, 
I think.


Tagir, overall, looks good! Let me know when you've finished updating the CSR.


> Regards,
> Tomasz Linkowski
> [1] https://en.wikipedia.org/wiki/Composite_pattern
> [2] https://en.wiktionary.org/wiki/composite#Verb
> On Fri, Sep 14, 2018 at 11:23 AM, Peter Levart <peter.levart at gmail.com 
> <mailto:peter.levart at gmail.com>> wrote:
>     Hi Tagir,
>     I like duplexing more than teeingAndThen. If consensus can be established
>     about the name, I think you will then want to update the CSR draft to
>     reflect new name. Then we'll kindly ask Stuart if he has any more advice
>     before submitting the CSR...
>     Regards, Peter
>     On 09/14/2018 10:41 AM, Tagir Valeev wrote:
>         Hello, Stuart and Peter!
>         Thank you for valuable comments. I updated the webrev:
>         http://cr.openjdk.java.net/~tvaleev/webrev/8205461/r4/
>         <http://cr.openjdk.java.net/%7Etvaleev/webrev/8205461/r4/>
>         1. I renamed "teeingAndThen" to "duplexing". Brian insisted that
>         "-ing" suffix shall present and I agree. Hopefully it's final name.
>         2. I updated the spec as Stuart suggested.
>         No changes in implementation since r3 revision. Please check.
>         With best regards,
>         Tagir Valeev.
>         On Tue, Aug 21, 2018 at 12:43 PM Stuart Marks <stuart.marks at oracle.com
>         <mailto:stuart.marks at oracle.com>> wrote:
>             Hi Tagir,
>             Thanks for working on this. This looks really cool! And thanks
>             Peter for
>             agreeing to sponsor it.
>             I can help out with the CSR. My first bit of advice about the CSR
>             process is to
>             hold off until the specification is complete. :-)
>             I think the intent of the API is fine, but I think some details of
>             the returned
>             collector need to be ironed out first.
>             1. The spec doesn't say what the returned collector's supplier,
>             accumulator,
>             combiner, and finisher do. On the one hand, we don't necessarily
>             want to
>             describe the actual implementation. On the other hand, we need to
>             specify how
>             the thing actually behaves. One can certainly deduce the intended
>             behavior from
>             the description, but this really needs to be specified, and it
>             mustn't rely on
>             the reader having to derive the required behaviors. Since the actual
>             implementation is fairly simple, the spec might end up being
>             rather close to the
>             implementation, but that might be unavoidable.
>             I'm envisioning something like this:
>                - supplier: creates a result container that contains result
>             containers
>             obtained by calling each collector's supplier
>                - accumulator: calls each collector's accumulator with its
>             result container
>             and the input element
>             ... and similar for the combiner and finisher functions.
>             2. Characteristics
>                - UNORDERED: should the returned collector be UNORDERED if
>             *either* of the
>             provided collectors is UNORDERED? (Current draft says *both*.)
>                - CONCURRENT: current draft seems correct
>                - IDENTITY_FINISH: clearly not present; perhaps this should be
>             specified
>             3. Parameter naming
>             The collector parameters are referred to as "specified collectors"
>             or "supplied
>             collectors". Other "higher-order" collectors refer to their underlying
>             collectors as "downstream" collectors. I think it would be useful
>             to work the
>             "downstream" concept into this collector. This would enable the
>             opening summary
>             sentence of the doc to be something like, "Returns a collector
>             that is a
>             composite of two downstream collectors" or some such. (But see
>             naming below.)
>             4. Naming
>             Sigh, naming is hard, and I know there was a fair amount of
>             discussion in the
>             previous thread and earlier in this one, but it seems like there's
>             still some
>             dissatisfaction with the name. (And I'm not particularly thrilled with
>             teeingAndThen myself.) In a few minutes I've managed to come up
>             with a few more
>             names that (mostly) don't seem to have been proposed before, and
>             so here they
>             are for your consideration:
>                - compound
>                - composite
>                - conjoined
>                - bonded
>                - fused
>                - duplex
>             Discussion:
>             A "composite" evokes function composition; this might be good,
>             though it might
>             be odd in that collectors can't be composed in the same way that
>             functions are.
>             "Compound" might be a useful alternative. In chemistry, two
>             substances are
>             combined (or bonded, or fused) to form a single substance, which
>             is a compound.
>             "Conjoin" seems to adequately describe the structure of the two
>             collectors, but
>             it lacks somewhat the connotation of unifying them.
>             In an earlier discussion, Brian had pushed back on names related to
>             split/fork/merge/join since those are currently in use in streams
>             regarding
>             splitting of input elements and merging of results. In describing
>             how the
>             current proposal differs, he said that elements are "multiplexed"
>             to the
>             different collectors. Since we're doing this with two collectors,
>             how about
>             "duplex"? (I note that Jacob Glickman also had suggested "duplex".)
>             s'marks
>             On 8/20/18 1:48 AM, Tagir Valeev wrote:
>                 Hello!
>                 A CSR is created:
>                 https://bugs.openjdk.java.net/browse/JDK-8209685
>                 <https://bugs.openjdk.java.net/browse/JDK-8209685>
>                 (this is my first CSR, hopefully I did it correctly)
>                 With best regards,
>                 Tagir Valeev.
>                 On Mon, Aug 20, 2018 at 2:06 PM Peter Levart
>                 <peter.levart at gmail.com <mailto:peter.levart at gmail.com>> wrote:
>                     Hi Tagir,
>                     I think this looks very good. It just needs a CSR. Will
>                     you file it?
>                     Regards, Peter
>                     On 08/19/2018 11:24 AM, Tagir Valeev wrote:
>                     Hello, Brian!
>                     Of the three phases, teeing is the most important and
>                     least obvious, so
>                     I think something that includes that in the name is going
>                     to be
>                     helpful.  Perhaps "teeingAndThen" is more evocative and
>                     not totally
>                     unwieldy.
>                     Ok, sounds acceptable to me. Renamed pairing to teeingAndThen.
>                     By the way looking into CollectorsTest.java I found some
>                     minor things to
>                     cleanup:
>                     1. `.map(mapper::apply)` and `.flatMap(mapper::apply)` can
>                     be replaced with
>                     simple `.map(mapper)` and `.flatMap(mapper)` respectively
>                     Does IntelliJ have an inspection for eliminating such
>                     locutions?
>                     Sure, that's how I found them. Well, I took the liberty to
>                     fix these two things.
>                     2. In many methods redundant `throws
>                     ReflectiveOperationException` is
>                     declared while exception is never thrown
>                     For test code where a significant fraction of test cases
>                     are going to
>                     throw something, we often do this, since its easier to
>                     just uniformly
>                     tag such methods rather than thinking about which test
>                     methods actually
>                     throw the exception and which don't.  So I think this is
>                     harmless
>                     (though cleaning it up is harmless too.)
>                     I'm not thinking about this, because my IDE thinks for me
>                     :-) Ok, I'll
>                     leave them as is for now.
>                     You may want to optimize the EnumSet mechanics for the
>                     case where
>                     neither collector has interesting characteristics.
>                     Added a special case when reported characteristics for
>                     either of
>                     collectors are empty or IDENTITY_FINISH only.
>                     I think this should be a common case.
>                     The updated webrev is posted here (along with Peter
>                     suggestion to
>                     rename finisher to merger):
>                     http://cr.openjdk.java.net/~tvaleev/webrev/8205461/r3/
>                     <http://cr.openjdk.java.net/%7Etvaleev/webrev/8205461/r3/>
>                     Also copyright year is updated
>                     With best regards,
>                     Tagir Valeev
> -- 
> Pozdrawiam,
> Tomasz Linkowski

More information about the core-libs-dev mailing list