reactor/reactor-core v3.2.0.RELEASE

on GitHub

This is the first release of Reactor 3.2, part of the Californium-RELEASE Release Train.

This major release contains a few significant new features, in addition to a heap of improvements and bug fixes.

⚠️ Update considerations and deprecations

• Compared to 3.1.x, the Scannable#operatorName() method has been renamed to stepName() (#1156)
• Disposables.composite() now has list semantics rather than undocumented set semantics, and thus should be more performant (#1237)
• The constant Schedulers.DEFAULT_POOL_SIZE is now used internally (notably by Schedulers.[new]parallel()), but unlike before it doesn't enforce a minimum of 4 anymore. It can be tuned via system property reactor.schedulers.defaultPoolSize now though (#1243, #1246)
• repeat(N) is now aligned with retry(N), as well as repeat(N, Predicate)/retry(N, Predicate): the number indicates the amount of repetitions of the original sequence, ie 0 mirrors the original sequence while 1 mirrors the original sequence + an additional repetition of it. Negative values are rejected at assembly time. (#1252)
  • look out for usage of repeat(long) in your code and decrement the passed parameter by one
  • repeat(0) should be replaced with Flux/Mono.empty()
• windowUntil and windowWhile don't have a final empty window at the end of the sequence (#1033)
• Default behavior for Operators.onErrorDropped is now to both bubble and throw the error
  • Mono.fromCompletionStage now drops fatal exceptions rather than hanging (#1118)
• Blocking APIs (like blockLast(), block(), iterator()) called inside a parallel or single Scheduler trigger an exception (#1102)
  • This kind of blocking call are harmful as they impact limited resources, with a high risk of freezing the application
• concatMapDelayError default behavior is now to delay the errors until the END (#1101)
  • This aligns with other *DelayError operators
• StepVerifier.withVirtualTime now use a global lock, making virtual time verifications mutually exclusive (#648)
  • Virtual time impacts schedulers at the application level. Parallelisation of tests could lead to unforeseen side effects due to this (initializing operators with a VTS and have the StepVerifier see and manipulate the wrong VTS)
• create uses Operators.onOperatorError hook instead of onErrorDropped (#1078)
• (deprecation) Scannable#operatorName() is deprecated in favor of stepName() (#1115, #1140)
• (internal) Switch to a simplified implementation for Mono.publish(Function) (#437)
• (internal) Simplify MonoProcessor internals in preparation for #1114 and further 3.2 changes

⭐ Release focus features

• Add metrics() to Flux (#1183, #1123, 34e0d4b, #1245, #1242, #1250)
  • This operator does nothing if Micrometer is not on the classpath
  • It exposes metrics from upstream signals visible at the operator's position in the chain
• Add exponential backoff retry with jitter to core: retryWithBackoff (#1122)
  • This version of retry reflects what we think is the industry best practice in terms of retries.
  • It is a good middle ground between the too simple retry(n), the complex retryWhen(Function) and the more configurable RetryFunction from reactor-addons
• New operators for transactional reactive use cases: usingWhen (#1220, b87ea6d, #1233, #1259)
  • Like using, but the resource is provided asynchronously through a Publisher
  • Can have separate async "cleanup" for complete, error and cancel terminations
  • Cleanups are asynchronous as well (Function<Resource, Publisher>) and only delay the propagation of the terminal signals, NOT the onNext signals.
  • Cancelled early before the resource is emitted, the Publisher<Resource> itself is cancelled.
• New doOnDiscard(Class, Consumer) operator (#999, #1343, #1345, #1347)
• Installs a local Context hook that will let upstream operators in the chain clean up elements that get either filtered out of the downstream sequence or prefetched then discarded
• New onErrorContinue() mode that influences the upstream chain of operator (#629)
  • Supporting operators up in the chain will request more from their source in case of error, instead of terminating the sequence.
  • The Throwable is passed to a handler instead of downstream.

✨ New features and improvements

• Include sequence name in timeout()'s message (#1349)
• Replace String.replaceAll with a static Pattern and Matcher.replaceAll in Scannable (#1315)
• Added a way to strictly limit the rate of requests with limitRate(n, 0) (#1317)
  • previously the lowTide and highTide parameter could still be changed by internal prefetching strategies
  • now with a lowTide of 0, the requests will strictly adhere to highTide
• Blocking code detection and rejection has been improved for toStream() and toIterable(), and now only reject when the iteration is performed (#1313)
  • it would previously reject right where the Iterable was materialized, but what actually blocks is the act of iterating.
• Improve concat error message on null Publisher (#1321)
• Calling Mono.from(Flux.from(mono)) is now immediately returning the mono instance, as this is conceptually a no-op. (#1329)
  • the inverse operation (Flux.from(Mono.from(flux))) is NOT a no-op as it transforms the original Flux into a Mono, so it still returns a different instance.
• Use more descriptive parameter names for buffer durations (#1331)
• The default Scheduler pool size is now configurable via a system property (#1243)
• In the case where a QueueSubscription is logged via log() during an onNext, its toString method is explicitly called (#1270)
  • typically this would happen when using window() immediately followed by log(), which is not really useful but was even less useful since it would trigger an exception
• Tuple2..Tuple8 now have mapT1..mapT8 (as relevant) methods that allow to change a single part of the Tuple (e4a9aee, discussed in #1058)
• cache(Duration) now have an overload that takes a Scheduler, added Mono.cache TTL-generator variant (#1189, #1125)
• Operators.lift now has an alternative that exposes the raw Publisher: liftPublisher and supports Fuseable (#1205, #1206)
• Added zip variants with 7 and 8 arguments (#1210
• Support 0 delay/period in Flux.interval (#1178)
• Add immutable empty() queue Supplier to Queues, support empty and one cases in Queues.capacity() (#1161, d9d76ab)
• doOnEach now support Fuseable and ConditionalSubscriber (#1003)
• Flux.take(0) now eagerly cancels on subscription (#1158)
• Lazy Mono#fromFuture and Mono#fromCompletionStage added, with Supplier param (#1131)
• Add an error(Supplier) variant to Flux and Mono (#1100)
• Add Scannable#steps(), which produces a Stream<String> of all the stepNames both upstream and downstream (including the current Scannable) (#1115, #1140)
• All operators now implement basic Scannable (this avoids problems for users that use instanceof checks in hooks, #1136)
• [kotlin] Add Flux.split extension to convert Flux<List<T>> to Flux<T> (#1089)
• [reactor-test] Add verifyThenAssertThat(Duration) to allow assertions and verification timeout (#1124)
• [reactor-test] Allow naming of a whole StepVerifier scenario through options (#1077)
• [reactor-test] Allow to change the LoggerFactory with a custom implementation, provide a TestLogger one that can be used to assert log contents (8f3a8fa)
• [reactor-test] Add a cold version of the TestPublisher than can be set up before eg. subscribing in a StepVerifier (#1236)

🪲 Bug fixes

• Fix ClassCastException in OneQueue/ZeroQueue + Java9 build (#1326)
• When using doOnSuccess and subscribing without an error handler, errors would be swallowed (#1337)
  • This now generates a ErrorCallbackNotImplemented exception like in other cases, unless a doOnTerminate is also used
  • The best practice is still to always define an error handler when using subscribe(...)
• Fixed an ArrayIndexOutOfBoundsException when using lift with a ParallelFlux operator that is fuseable (can be triggered by using Sleuth, #1293)
• FluxPublish should trigger an extra poll on sync fusion (#1290, #1291)
• Request fusion with THREAD_BARRIER in concatMap* and publish (#1302)
• When using a non-compliant TestPublisher, calling .mono() would turn it into a compliant one. This has been fixed and the mono will now continue to be a bad boy (#1244)
• A few bufferTimeout and bufferWhen issues have been corrected around cancellation and drain race conditions (#1247)
• Flux.last() now correctly throws a NoSuchElementException on empty source even when said source is a Callable Mono(#1226)
• delayUntil correctly request Long.MAX_VALUE and not Integer.MAX_VALUE on subscribe (#1214)
• Operators.lift() properly maintain GroupedFlux/ConnectableFlux interfaces (#1204)
• Request is now accounted for in `FluxOnBackpressureBufferTimeout (#1194)
• onSubscribe should be done before add in MonoCacheTime.subscribe (#1190)
• Multi-subscriber operators Context resolution resolves to 1st context (#1114)
• Fix dangling thread when calling WorkQueueProcessor.forceShutdown (#1142)
• Let VirtualTimeScheduler defer advanceTime if queue is empty (#783)
  • This avoids a situation where, due to a subscribeOn, the operators delaying values end up reading the clock AFTER the virtual time has been moved forward by thenAwait, hanging the test.
• Tweak FluxPublish dispose to clear the existing connection (#1173)
• Avoid interrupting WorkerTask Future if cancelled race (#1107)
• FluxRefCount DisconnectException on cancellation race (#1088)
• (internal cleanup) use passed queue in FluxGroupBy.checkTerminated (#1094)
• Only terminate ExecutorScheduler if underlying is shut down (#1080)
• Mono.fromCompletionStage now drops fatal exceptions rather than hanging (#1118)

📖 Documentation, Tests and Build

• [doc] The reference documentation has been extensively reviewed and revised for 3.2 (#1325)
• [doc] Point between Flux#create and Flux#push, show multithreading diff (3834537)
• [test] Do not fail HooksTestStaticInit test in IDE (#1322)
• [build] Add LGTM.com code quality badges (#1339)
• [doc] Fixed various javadoc marble diagram issues (#1043, #1008, #1167, #1266, #1079, #1294)
• [test] Added unit test for SpscArrayQueue (35519ab)
• Some reference documentation error fallback snippets were wrong (#1269)
• Java 9 build has been fully reactivated now that every build component is Java 9 compatible (right on time for Java 9 EoL 😁 see 2b4bf5e, 209a45f, #896, e239d65)
• Kotlin has been bumped to 1.2.51 (#1274)
• Document (im)mutability of Tuple2 toList() vs iterator() (#1135)
• Point to reference doc operator matrix in Flux/Mono javadoc (f2bc74f)
• Warn on doOnSubscribe usage (#1090)
• various javadoc typos and minor rephrasing (#1240, #1222, 89a8e0e, #1149, #1199, 51dff9f, 4a1f691, 4e55122, #1095, 7a6a8a1, 35cb6fd, #1346)
• Various test fixes (2f74309, 0254f45, #1139, 1c27a15, 44dc73c, 8d942e1, 38f8edb, #1218)
• Add meaningful toString to FluxSink/MonoSink implementations (#1130)

🎁 Contribution process change

• Contributing Guidelines have been totally reworked and reflect the new organisation, notably in terms of labels and branching (see details below, d166557, #1225)

A notable change in the way we work with the repository and contributions that make sense for maintenance releases: as of commit 0d8e16e, there will be no more port #xxx backport commits in maintenance branches, but rather issues will be fixed first in the maintenance branch THEN a forward-merge will happen to include the fix in the master branch. See #1225.

👍 Thanks to the following contributors that also participated to this release

@alex-diez, @baptistemesta, @charlesmuchene, @genie-youn, @igoperikov, @madgnome, @MarkusJais, @nurkiewicz, @OlegDokuka, @pdudkiewicz, @smiklos, @szpak, @utwyko, @xcorail, @yamkazy, @0xflotus