Hi Pavel, Interesting findings :) Agree that we should not use the CompletableFuture - it clearly has a different purpose.
I think that the approach taken by Redis makes more sense. I don't like that it requires a custom interface, but I think we can live with that. I would be glad to hear other opinions though. -Val On Fri, Aug 21, 2020 at 1:02 AM Pavel Tupitsyn <ptupit...@apache.org> wrote: > Val, > > The problems with CompletableFuture in public API are: > * It is a class, not an interface > * It is completable - anyone can call .complete(), which is not what we > want > > There seems to be no clear guidance in Java world on async API design; > however, it is often recommended to return CompletionStage instead of > CompletableFuture > from the public APIs [1] [2], and some products follow this [3]. > > Other products return their own future interface that extends both Future > and CompletionStage, > which seems to be a better alternative to me [4]. > > Thoughts? > > [1] > > https://stackoverflow.com/questions/47571117/what-is-the-difference-between-completionstage-and-completablefuture > [2] > > https://stackoverflow.com/questions/34930840/should-i-return-completablefuture-or-future-when-defining-api > < > https://stackoverflow.com/questions/34930840/should-i-return-completablefuture-or-future-when-defining-api#:~:text=by%20returning%20a%20CompletableFuture%2C%20you,API%2C%20which%20is%20not%20good > .> > [3] > > https://docs.hazelcast.org/docs/latest/javadoc/com/hazelcast/cache/ICache.html > [4] > > https://lettuce.io/lettuce-4/release/api/com/lambdaworks/redis/RedisFuture.html > > On Fri, Aug 21, 2020 at 10:28 AM Alex Plehanov <plehanov.a...@gmail.com> > wrote: > > > Pavel, > > > > Thanks for the discussion, I've also faced with the necessity of having > > async calls while implementing POC for thin client data streamer [1] and > > solve it similarly (but in my case it's required only for internal > > implementation, so I've only changed the internal API). > > > > I want to note that described in IEP approach (and implemented in POC) is > > not fully async, since "send" is still used in the user's thread. To make > > it fully async we need additional sending thread (since blocking IO is > used > > for communication with the server). If partition awareness is enabled > there > > will be 2 threads per each server connection, perhaps we should think > about > > moving to NIO and introducing some kind of communication thread pool. > > > > [1]: https://github.com/apache/ignite/pull/8175 > > > > пт, 21 авг. 2020 г. в 03:35, Valentin Kulichenko < > > valentin.kuliche...@gmail.com>: > > > > > Sounds good. I've added this to the 3.0 roadmap: > > > https://cwiki.apache.org/confluence/display/IGNITE/Apache+Ignite+3.0 > > > > > > Unless there are any objections from others, let's stick with the > > > CompletableFuture for any future development, including the thin > client. > > > > > > -Val > > > > > > On Thu, Aug 20, 2020 at 9:30 AM Pavel Tupitsyn <ptupit...@apache.org> > > > wrote: > > > > > > > Val, no objections from my side. > > > > As noted above, the only benefit of IgniteFuture is consistency > across > > > > thin/thick APIs, > > > > which is probably not so important. > > > > > > > > On Thu, Aug 20, 2020 at 6:28 PM Valentin Kulichenko < > > > > valentin.kuliche...@gmail.com> wrote: > > > > > > > > > Hi Pavel, > > > > > > > > > > Are there any benefits of IgniteFuture over CompletableFuture? > > > > > > > > > > IgniteFuture was created long ago, during the time when > > > CompletableFuture > > > > > did not exist. There is a big chance that IgniteFuture actually > > became > > > > > redundant at the moment we transitioned to Java8. If that's the > > case, I > > > > > would prefer using CompletableFuture in the thin client and getting > > rid > > > > of > > > > > IgniteFuture altogether in 3.0. > > > > > > > > > > What do you think? > > > > > > > > > > -Val > > > > > > > > > > On Thu, Aug 20, 2020 at 7:19 AM Pavel Tupitsyn < > ptupit...@apache.org > > > > > > > > wrote: > > > > > > > > > > > Igniters, > > > > > > > > > > > > I've prepared an IEP [1], please review and let me know what you > > > think. > > > > > > > > > > > > In particular, I'd like to discuss the Future interface to be > used: > > > > > > * IgniteFuture is the first candidate - Thin APIs will be > > consistent > > > > with > > > > > > Thick APIs, probably better for existing Ignite users. > > > > > > * CompletableFuture is the standard for async Java APIs. Many > users > > > may > > > > > > prefer that instead of a custom IgniteFuture. > > > > > > > > > > > > [1] > > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/IGNITE/IEP-51%3A+Java+Thin+Client+Async+API > > > > > > > > > > > > > > > > > > > > >