Simple benchmark added - see JmhCacheAsyncListenBenchmark in the PR. There is a 6-8% drop (1 client, 2 servers, 1 machine, int key/val). I expect this difference to become barely observable on real-world workloads.
On Thu, Mar 18, 2021 at 12:35 PM Pavel Tupitsyn <ptupit...@apache.org> wrote: > Denis, > > For a reproducer, please see CacheAsyncContinuationExecutorTest.java in > the linked PoC [1] > > [1] > https://github.com/apache/ignite/pull/8870/files#diff-c788c12013622093df07d8f628a6e8c5fb0c15007f0787f2d459dbb3e377fc5aR54 > > On Thu, Mar 18, 2021 at 1:58 AM Raymond Wilson <raymond_wil...@trimble.com> > wrote: > >> We implemented the ContinueWith() suggestion from Pavel, and it works well >> so far in testing, though we do not have data to support if there is or is >> not a performance penalty in our use case.. >> >> To lend another vote to the 'Don't do continuations on the striped thread >> pool' line of thinking: Deadlocking is an issue as is starvation. In some >> ways starvation is more insidious because by the time things stop working >> the cause and effect distance may be large. >> >> I appreciate the documentation does make statements about not performing >> cache operations in a continuation due to deadlock possibilities, but that >> statement does not reveal why this is an issue. It is less a case of a >> async cache operation being followed by some other cache operation (an >> immediate issue), and more a general case of the continuation of >> application logic using a striped pool thread in a way that might mean >> that >> thread is never given back - it's now just a piece of the application >> infrastructure until some other application activity schedules away from >> that thread (eg: by ContinueWith or some other async operation in the >> application code that releases the thread). To be fair, beyond structures >> like ContinueWith(), it is not obvious how that continuation thread should >> be handed back. This will be the same behaviour for dedicated continuation >> pools, but with far less risk in the absence of ContinueWith() constructs. >> >> In the .Net world this is becoming more of an issue as fewer .Net use >> cases >> outside of UI bother with synchronization contexts by default. >> >> Raymond. >> >> >> On Thu, Mar 18, 2021 at 9:56 AM Valentin Kulichenko < >> valentin.kuliche...@gmail.com> wrote: >> >> > Hi Denis, >> > >> > I think Pavel's main point is that behavior is unpredictable. For >> example, >> > AFAIK, putAsync can be executed in the main thread instead of the >> striped >> > pool thread if the operation is local. The listener can also be >> executed in >> > the main thread - this happens if the future is completed prior to >> listener >> > invocation (this is actually quite possible in the unit test environment >> > causing the test to pass). Finally, I'm pretty sure there are many cases >> > when a deadlock does not occur right away, but instead it will reveal >> > itself under high load due to thread starvation. The latter type of >> issues >> > is the most dangerous because they are often reproduced only in >> production. >> > Finally, there are performance considerations as well - cache operations >> > and listeners share the same fixed-sized pool which can have negative >> > effects. >> > >> > I'm OK with the change. Although, it might be better to introduce a new >> > fixed-sized pool instead of ForkJoinPool for listeners, simply because >> this >> > is the approach taken throughout the project. But this is up to a >> debate. >> > >> > -Val >> > >> > On Wed, Mar 17, 2021 at 11:31 AM Denis Garus <garus....@gmail.com> >> wrote: >> > >> > > Pavel, >> > > I tried this: >> > > >> > > @Test >> > > public void test() throws Exception { >> > > IgniteCache<Integer, String> cache = >> > > startGrid().getOrCreateCache("test_cache"); >> > > >> > > cache.putAsync(1, "one").listen(f -> cache.replace(1, "two")); >> > > >> > > assertEquals("two", cache.get(1)); >> > > } >> > > >> > > and this test is green. >> > > I believe that an user can make listener that leads to deadlock, but >> > > the example in the IEP does not reflect this. >> > > >> > > >> > > >> > > ср, 17 мар. 2021 г. в 17:36, Вячеслав Коптилин < >> slava.kopti...@gmail.com >> > >: >> > > >> > > > Hi Pavel, >> > > > >> > > > > Not a good excuse really. We have a usability problem, you have to >> > > admit >> > > > it. >> > > > Fair enough. I agree that this is a usability issue, but I have >> doubts >> > > that >> > > > the proposed approach to overcome it is the best one. >> > > > >> > > > > Documentation won't help - no one is going to read the Javadoc >> for a >> > > > trivial method like putAsync >> > > > That is sad... However, I don't think that this is a strong argument >> > > here. >> > > > >> > > > This is just my opinion. Let's see what other community members >> have to >> > > > say. >> > > > >> > > > Thanks, >> > > > S. >> > > > >> > > > >> > > > ср, 17 мар. 2021 г. в 17:01, Pavel Tupitsyn <ptupit...@apache.org>: >> > > > >> > > > > > the user should use the right API >> > > > > >> > > > > Not a good excuse really. We have a usability problem, you have to >> > > admit >> > > > > it. >> > > > > "The brakes did not work on your car - too bad, you should have >> known >> > > > that >> > > > > on Sundays only your left foot is allowed on the pedal" >> > > > > >> > > > > This particular use case is too intricate. >> > > > > Even when you know about that, it is difficult to decide what can >> run >> > > on >> > > > > the striped pool, >> > > > > and what can't. It is too easy to forget. >> > > > > And most people don't know, even among Ignite developers. >> > > > > >> > > > > Documentation won't help - no one is going to read the Javadoc >> for a >> > > > > trivial method like putAsync. >> > > > > >> > > > > >> > > > > So I propose to have a safe default. >> > > > > Then document the performance tuning opportunity on [1]. >> > > > > >> > > > > Think about how many users abandon a product because it >> mysteriously >> > > > > crashes and hangs. >> > > > > >> > > > > [1] >> > > > > >> > > > > >> > > > >> > > >> > >> https://ignite.apache.org/docs/latest/perf-and-troubleshooting/general-perf-tips >> > > > > >> > > > > >> > > > > On Wed, Mar 17, 2021 at 4:21 PM Вячеслав Коптилин < >> > > > > slava.kopti...@gmail.com> >> > > > > wrote: >> > > > > >> > > > > > Hi Pavel, >> > > > > > >> > > > > > Well, I think that the user should use the right API instead of >> > > > > introducing >> > > > > > uncontested overhead for everyone. >> > > > > > For instance, the code that is provided by IEP can changed as >> > > follows: >> > > > > > >> > > > > > IgniteFuture fut = cache.putAsync(1, 1); >> > > > > > fut.listenAync(f -> { >> > > > > > // Executes on Striped pool and deadlocks. >> > > > > > cache.replace(1, 2); >> > > > > > }, ForkJoinPool.commonPool()); >> > > > > > >> > > > > > Of course, it does not mean that this fact should not be >> properly >> > > > > > documented. >> > > > > > Perhaps, I am missing something. >> > > > > > >> > > > > > Thanks, >> > > > > > S. >> > > > > > >> > > > > > ср, 17 мар. 2021 г. в 16:01, Pavel Tupitsyn < >> ptupit...@apache.org >> > >: >> > > > > > >> > > > > > > Slava, >> > > > > > > >> > > > > > > Your suggestion is to keep things as is and discard the IEP, >> > right? >> > > > > > > >> > > > > > > > this can lead to significant overhead >> > > > > > > Yes, there is some overhead, but the cost of accidentally >> > starving >> > > > the >> > > > > > > striped pool is worse, >> > > > > > > not to mention the deadlocks. >> > > > > > > >> > > > > > > I believe that we should favor correctness over performance in >> > any >> > > > > case. >> > > > > > > >> > > > > > > >> > > > > > > On Wed, Mar 17, 2021 at 3:34 PM Вячеслав Коптилин < >> > > > > > > slava.kopti...@gmail.com> >> > > > > > > wrote: >> > > > > > > >> > > > > > > > Well, the specified method already exists :) >> > > > > > > > >> > > > > > > > /** >> > > > > > > > * Registers listener closure to be asynchronously >> notified >> > > > > > whenever >> > > > > > > > future completes. >> > > > > > > > * Closure will be processed in specified executor. >> > > > > > > > * >> > > > > > > > * @param lsnr Listener closure to register. Cannot be >> > {@code >> > > > > > null}. >> > > > > > > > * @param exec Executor to run listener. Cannot be >> {@code >> > > > null}. >> > > > > > > > */ >> > > > > > > > public void listenAsync(IgniteInClosure<? super >> > > > IgniteFuture<V>> >> > > > > > > lsnr, >> > > > > > > > Executor exec); >> > > > > > > > >> > > > > > > > Thanks, >> > > > > > > > S. >> > > > > > > > >> > > > > > > > ср, 17 мар. 2021 г. в 15:25, Вячеслав Коптилин < >> > > > > > slava.kopti...@gmail.com >> > > > > > > >: >> > > > > > > > >> > > > > > > > > Hello Pavel, >> > > > > > > > > >> > > > > > > > > I took a look at your IEP and pool request. I have the >> > > following >> > > > > > > > concerns. >> > > > > > > > > First of all, this change breaks the contract of >> > > > > > > > IgniteFuture#listen(lsnr) >> > > > > > > > > >> > > > > > > > > /** >> > > > > > > > > * Registers listener closure to be asynchronously >> > notified >> > > > > > > whenever >> > > > > > > > > future completes. >> > > > > > > > > * Closure will be processed in thread that completes >> > this >> > > > > future >> > > > > > > or >> > > > > > > > > (if future already >> > > > > > > > > * completed) immediately in current thread. >> > > > > > > > > * >> > > > > > > > > * @param lsnr Listener closure to register. Cannot be >> > > {@code >> > > > > > > null}. >> > > > > > > > > */ >> > > > > > > > > public void listen(IgniteInClosure<? super >> > IgniteFuture<V>> >> > > > > > lsnr); >> > > > > > > > > >> > > > > > > > > In your pull request, the listener is always called >> from >> > a >> > > > > > > specified >> > > > > > > > > thread pool (which is fork-join by default) >> > > > > > > > > even though the future is already completed at the >> moment >> > > the >> > > > > > > listen >> > > > > > > > > method is called. >> > > > > > > > > In my opinion, this can lead to significant overhead - >> > > > > submission >> > > > > > > > > requires acquiring a lock and notifying a pool thread. >> > > > > > > > > >> > > > > > > > > It seems to me, that we should not change the current >> > > > behavior. >> > > > > > > > > However, thread pool executor can be added as an optional >> > > > parameter >> > > > > > of >> > > > > > > > > listen() method as follows: >> > > > > > > > > >> > > > > > > > > public void listen(IgniteInClosure<? super >> > > > IgniteFuture<V>> >> > > > > > > lsnr, >> > > > > > > > > Executor exec); >> > > > > > > > > >> > > > > > > > > Thanks, >> > > > > > > > > S. >> > > > > > > > > >> > > > > > > > > пн, 15 мар. 2021 г. в 19:24, Pavel Tupitsyn < >> > > > ptupit...@apache.org >> > > > > >: >> > > > > > > > > >> > > > > > > > >> Igniters, >> > > > > > > > >> >> > > > > > > > >> Please review the IEP [1] and let me know your thoughts. >> > > > > > > > >> >> > > > > > > > >> [1] >> > > > > > > > >> >> > > > > > > > >> >> > > > > > > > >> > > > > > > >> > > > > > >> > > > > >> > > > >> > > >> > >> https://cwiki.apache.org/confluence/display/IGNITE/IEP-70%3A+Async+Continuation+Executor >> > > > > > > > >> >> > > > > > > > > >> > > > > > > > >> > > > > > > >> > > > > > >> > > > > >> > > > >> > > >> > >> >> >> -- >> <http://www.trimble.com/> >> Raymond Wilson >> Solution Architect, Civil Construction Software Systems (CCSS) >> 11 Birmingham Drive | Christchurch, New Zealand >> raymond_wil...@trimble.com >> >> < >> https://worksos.trimble.com/?utm_source=Trimble&utm_medium=emailsign&utm_campaign=Launch >> > >> >
