On Mon, Dec 14, 2015 at 4:15 PM, David Whiting <[email protected]> wrote:
> 1) Not at all, just some leftover working names for stuff. > > 2) Not for a totally minimal implementation, but some of the features I > would like to include would rely on Java 8 things, for example adapting the > GroupedTable stuff to use Streams rather than Iterables because of a) the > extra expressivity and b) the implied once-only traversal. We could have a > filterMap which applies a Function<S, Optional<T>> (my most common use case > for a DoFn instead of a MapFn at the moment). We can also potentially > utilise Collectors for collapsing values in reduce-side stuff and finally, > it'll make the implementation of it a fair bit easier. The maven overhead > is pretty low, so I guess it's just the existence of an extra artifact to > consider. The way I see it is that it's a push to make the API feel more > like Java streams and be more immediately usable by someone who knows Java > streams but not necessarily big data, so the more we can replicate that > feel by integrating with other familiar Java 8 features, the better. > Makes sense to me. +1 for a new crunch-lambda module. > > On 15 December 2015 at 00:51, Josh Wills <[email protected]> wrote: > > > I think I lean towards the collections approach, but that's probably > > because of my Scrunch experience. Two questions: > > > > 1) Is mapToTable necessary? I would think map(SFunction, PTableType) > would > > be distinguishable from map(SFunction, PType) by the compiler in the same > > way it is for parallelDo. > > 2) Does the collections approach need a separate maven target at all, or > > could it just be part of crunch-core as a replacement for the IFn stuff? > Or > > is there Java 8-only stuff we'll want to add in to its API? > > > > On Mon, Dec 14, 2015 at 3:13 PM, David Whiting <[email protected]> wrote: > > > > > Ok, so I've implemented a few iterations of this. I went forward with > the > > > "wrap the functions" method, which seemed to work alright, but finding > > good > > > names for functions which essentially just wrap functions but which > > aren't > > > ambiguous in erasure and read nicely was a real challenge. I showed > some > > > sample code to some of my fellow data engineers and the consensus > seemed > > to > > > be that it was definitely better than anonymous inner classes, but it > > still > > > felt kind of awkward and strange to use. > > > > > > So here's a 3rd option: wrap the collection types rather than the > > function > > > types, and present an API which feels truly Java 8 native whilst still > > > being able to dig back to the underlying PCollections (doing pretty > much > > > what Scrunch does, but with less implicit Scala magic). > > > > > > Here's a super-minimal proof-of-concept for that: > > > https://gist.github.com/DavW/7efe484ea0c00cf6e66b > > > > > > and a comparison of the two approaches in usage: > > > https://gist.github.com/DavW/997a92b31d55c5317fb7 > > > > > > > > > On 13 December 2015 at 16:14, Gabriel Reid <[email protected]> > > wrote: > > > > > > > This looks very cool. As long as we can keep things compatible with > > > > Java 7 using whatever kind of maven voodoo that's necessary, I'm all > > > > for it. > > > > > > > > I'd say no real reason to keep the IFn stuff if this goes in. > > > > > > > > - Gabriel > > > > > > > > On Fri, Dec 11, 2015 at 11:18 PM, Josh Wills <[email protected]> > > > wrote: > > > > > It seems like a net positive over the IFn stuff, so I could make an > > > > > argument for replacing it, but if there's anyone out there in love > > > > w/IFns, > > > > > they should speak up now. :) > > > > > > > > > > J > > > > > > > > > > On Fri, Dec 11, 2015 at 2:17 PM, David Whiting <[email protected]> > > > wrote: > > > > > > > > > >> I *think* you can set language level and target jdk on a > per-module > > > > basis, > > > > >> so it should be relatively easy. I'll experiment at some point > over > > > the > > > > >> weekend. Would this complement or replace the I*Fn stuff do you > > think? > > > > 14.0 > > > > >> is not yet released, so I guess it's not too late to change if we > > want > > > > to. > > > > >> > > > > >> On 11 December 2015 at 22:57, Josh Wills <[email protected]> > > > wrote: > > > > >> > > > > >> > That's the sexiest thing I've seen in some time. +1 for a lambda > > > > module, > > > > >> > but how does that work in Maven-fu? Is it like a conditional > > compile > > > > or > > > > >> > something? > > > > >> > > > > > >> > On Fri, Dec 11, 2015 at 1:20 PM, David Whiting <[email protected] > > > > > > wrote: > > > > >> > > > > > >> > > Oops, my bad. Here's a Gist: > > > > >> > > https://gist.github.com/DavW/e2588e42c45ad8c06038 > > > > >> > > > > > > >> > > On 11 December 2015 at 18:43, Josh Wills < > [email protected]> > > > > wrote: > > > > >> > > > > > > >> > > > I think it's kind of awesome, but the attachment didn't go > > > > through- > > > > >> PR > > > > >> > or > > > > >> > > > gist? > > > > >> > > > On Fri, Dec 11, 2015 at 7:42 AM David Whiting < > > [email protected]> > > > > >> wrote: > > > > >> > > > > > > > >> > > > > While fixing the bug where the IFn version of mapValues on > > > > >> > > PGroupedTable > > > > >> > > > > was missing, I got thinking that this is quite an > > inefficient > > > > way > > > > >> of > > > > >> > > > > including support for lambdas and method references, and > it > > > > still > > > > >> > > didn't > > > > >> > > > > actually support quite a few of the features that would > make > > > it > > > > >> easy > > > > >> > to > > > > >> > > > > code against. > > > > >> > > > > > > > > >> > > > > Negative parts of existing lambda implementation: > > > > >> > > > > 1) Explosion of already-crowded PCollection, PTable and > > > > >> PGroupedTable > > > > >> > > > > interfaces, and having to implement those methods in all > > > > >> > > implementations. > > > > >> > > > > 2) Not supporting flatMap to Optional or Stream types. > > > > >> > > > > 3) Not exposing convenient types for reduce-type > operations > > > > (Stream > > > > >> > > > > instead of Iterable, for example). > > > > >> > > > > > > > > >> > > > > Something that would solve all three of these is to build > > > lambda > > > > >> > > support > > > > >> > > > > as a separate artifact (so we can use all java8 types), > and > > > > instead > > > > >> > of > > > > >> > > > the > > > > >> > > > > API being directly on the PSomething interfaces, we just > > have > > > > >> > > convenient > > > > >> > > > > ways to wrap up lambdas into DoFns or MapFns via > > > > >> statically-imported > > > > >> > > > > methods. > > > > >> > > > > > > > > >> > > > > The usage then becomes > > > > >> > > > > import static org.apache.crunch.Lambda.*; > > > > >> > > > > ... > > > > >> > > > > someCollection.parallelDo(flatMap(d -> someFnOf(d)), pt) > > > > >> > > > > ... > > > > >> > > > > otherGroupedTable.mapValue(reduce(seq -> seq.mapToInt(i -> > > > > >> i).sum()), > > > > >> > > > > ints()) > > > > >> > > > > > > > > >> > > > > Where flatMap and reduce are static methods on Lambda, and > > > > Lambda > > > > >> > goes > > > > >> > > in > > > > >> > > > > it's own artifact (to preserve compatibility with 6 and 7 > > for > > > > the > > > > >> > rest > > > > >> > > of > > > > >> > > > > Crunch). > > > > >> > > > > I've attached a basic proof-of-concept implementation > which > > > I've > > > > >> > > tested a > > > > >> > > > > few things with, and I'm very happy to sketch out a more > > > > >> substantial > > > > >> > > > > implementation if people here think it's a good idea in > > > general. > > > > >> > > > > > > > > >> > > > > Thoughts? Ideas? Suggestions? Please tell me if this is > > crazy. > > > > >> > > > > > > > > >> > > > > > > > > >> > > > > > > > >> > > > > > > >> > > > > > >> > > > > > > > > > >
