+1 to Denis, Near caches are on-heap, as well, so I guess we need it. BTW, TeamCity Bot uses Guava on-heap caching above Ignite (Durable Memory). This is because keeping the same instance of Java object cached brings a visible performance boost for really hot code points. At least, it reduces GC pressure. Offheap->onheap unmarshalling gives new object from JVM point of view.
ср, 19 июн. 2019 г. в 00:35, Denis Magda <dma...@apache.org>: > Ivan, > > I believe that, yes, those caches are still extremely useful for > low-latency use cases. Companies are ready to allocate more RAM in favor of > lower latencies because off-heap access is still slower than the on-heap > one. There are not that many use case of this kind, but I can recall > several companies that exploit on-heap caching a lot. > > - > Denis > > > On Tue, Jun 18, 2019 at 11:42 AM Павлухин Иван <vololo...@gmail.com> > wrote: > > > Do we still need onheap caches? > > > > вт, 18 июн. 2019 г. в 21:30, Denis Magda <dma...@apache.org>: > > > > > > +1 > > > > > > Thick (aka. standard clients) provide comprehensive compute APIs with > > > peer-class-loading. That's a huge differentiator for Ignite. Until thin > > > clients support compute and ML API at the same level as the standard > > client > > > does, I would not consider the standard clients' discontinuation. Plus, > > as > > > Alex outlined, a functional gap is even wider. > > > > > > - > > > Denis > > > > > > > > > On Mon, Jun 17, 2019 at 6:28 AM Alexey Goncharuk < > > alexey.goncha...@gmail.com> > > > wrote: > > > > > > > Nikolay, > > > > > > > > I had this thought too, but I am not too eager to implement it yet. > The > > > > reason is transaction protocol complexity/performance issues with > thin > > > > clients. > > > > > > > > A thick client can communicate with each primary node and coordinate > > > > prepare/commit phases. Thin client can only communicate with one > node, > > so > > > > the change will mean an additional network hop. Of course, we can > make > > thin > > > > clients implement the same protocol, but it will immediately increase > > the > > > > protocol complexity for all platforms. > > > > > > > > Plus, we do not have near cache on thin clients, we do not support > p2p > > > > class deployment, etc. Since thin clients are positioned as > > > > platform-agnostic, I do not think it makes sense to expose all > feature > > set > > > > of Igntie to thin clients. > > > > > > > > Instead, we can significantly simplify client node configuration - it > > > > currently requires the same config as a regular Ignite node, however, > > in > > > > most cases, the configuration can be reduced almost to a several > > host:port > > > > pairs. > > > > > > > > пн, 17 июн. 2019 г. в 15:58, Nikolay Izhikov <nizhi...@apache.org>: > > > > > > > > > Alexey. > > > > > > > > > > I want to share a thought (just don't drop it out in one moment :) > ). > > > > > > > > > > Do we really need "client nodes"? > > > > > > > > > > We have thin client protocol that is a very convenient point to > > interact > > > > > with Ignite. > > > > > So, why, we need one more entity and work mode such as "client > node"? > > > > > > > > > > From my point of view, client nodes were required in the time > > without a > > > > > thin client. > > > > > Now, we have it. > > > > > > > > > > Let's simplify Ignite codebase and drop client nodes! > > > > > > > > > > How does it sound? > > > > > > > > > > > > > > > В Пн, 17/06/2019 в 15:52 +0300, Alexey Goncharuk пишет: > > > > > > Nikolay, > > > > > > > > > > > > Local caches and scalar are already in the list :) Added the > > outdated > > > > > > metrics point. > > > > > > > > > > > > пн, 17 июн. 2019 г. в 15:32, Nikolay Izhikov < > nizhi...@apache.org > > >: > > > > > > > > > > > > > * Scalar. > > > > > > > * LOCAL caches. > > > > > > > * Deprecated metrics. > > > > > > > > > > > > > > В Пн, 17/06/2019 в 15:18 +0300, Alexey Goncharuk пишет: > > > > > > > > Igniters, > > > > > > > > > > > > > > > > Even though we are still planning the Ignite 2.8 release, I > > would > > > > > like to > > > > > > > > kick-off a discussion related to Ignite 3.0, because the > > efforts > > > > for > > > > > AI > > > > > > > > > > > > > > 3.0 > > > > > > > > will be significantly larger than for AI 2.8, better to start > > > > early. > > > > > > > > > > > > > > > > As a first step, I would like to discuss the list of things > to > > be > > > > > removed > > > > > > > > in Ignite 3.0 (partially this thread is inspired by Denis > > Magda's > > > > > IGFS > > > > > > > > removal thread). I've separated all to-be-removed points from > > > > > existing > > > > > > > > Ignite 3.0 wishlist [1] to a dedicated block and also added a > > few > > > > > more > > > > > > > > things that look right to be dropped. > > > > > > > > > > > > > > > > Please share your thoughts, probably, there are more outdated > > > > things > > > > > we > > > > > > > > need to add to the wishlist. > > > > > > > > > > > > > > > > As a side question: I think it makes sense to create tickets > > for > > > > such > > > > > > > > improvements, how do we track them. Will the 3.0 version > > suffice or > > > > > > > > > > > > > > should > > > > > > > > we add a separate label? > > > > > > > > > > > > > > > > > -- > > Best regards, > > Ivan Pavlukhin > > >
