* deduplicate logic for IODispatcher #24604
* introduce a resolveDispatcher helper in ActorAttributes
* mention akka.stream.materializer.blocking-io-dispatcher instead of akka.stream.blocking-io-dispatcher in scaladocs
* fix a flaky test
* cosmetic changes in the touched files
* move resolveDispather helper to the Dispatcher companion object under a new name resolve
* filter out mima warning
* fix mima excludes after the 2.5.11 release
* address review comments
* update stream-io.md with the correct dispatcher config key
* mark ActorAttributes.Dispatcher#resolve as internal API
* use the dispatche config key in ActorMaterializer
* add private[akka] to the resolve methods
inside ActorContext, in which case it still is child actors as usual
This makes sense as they're "internal", so more like system actors
anyway, but the major reason for the change is Akka Typed, in which we
do not control the user guardian, and as such can not attach things
from the side into it
+str #24229 remove protobuf changes, which do not need to be made in this PR
docs
moved things
config object
subscription timeout confifmed working, also, attributes
document attributes for sub timeout
tests for the source also failing when it should
additional demand test
implemented protection from materializing "in cycles"; would be nice in
types but that breaks the niceness of use of the types
SinkRef/SourceRef...
cleanup
no idle timeout built in, can use the Timeout stages
more docs
simplest change to prevent exposing SinkRef => SourceRef => SinkRef cycle
Things to decide:
* is it ok to require using `getSource` / `getSink` as Java API, is there better naming?
* where should the constructors go? I'd say just in regular javadsl/scaladsl `Source`/ `Sink` objects
move constructors to {javadsl,scaladsl}.{Source,Sink} companion objects
Remove now useless "canMaterialize" field
Separate stage (implementation) from ref (wrapped actor ref) to make it clearer what is serialized
Clarify that partner refs are not optional in on-the-wire interfaces
minor cleanup in SourceRefStage
Renamed the stages but questionable if that really helps ;)
cleanups, better docs
cleanup, fix docs compilation
fix mima
got rid of Futures in the materialized values of stream refs
Instead of renaming - deprecating and documenting
Renaming defaultInitialAttributes to defaultAttributes and adding docs
Tests passing
More work on using attributes for the attribute settings in ActorMaterializerSettings
Mark some attributes as mandatory and provide a smoother access w/o allocation for those
Another warning notice
MiMa filters
Part of the review adressed
Incorrect rebase fixed + formatting
Review feedback adressed
Formatting ...
* This is an optimization of TcpStreamLogic to accumulating bytes in a buffer while waiting for
acknoledgment of pending write. This improves throughput for small messages (frames)
without sacrificing latency. While waiting for the ack the stage will eagerly pull
from upstream until the buffer limit is exceeded. Accumulated bytes are written
immediately when ack is received.
* Noticed 20x throughput improvement with Artery MaxThroughputSpec thanks to this buffer
when working on the Artery TCP implementation. The small message (100 bytes) benchmark
improved from 30k msg/s to 600k msg/s.
* +str #18793 StageLogging that allows logger access in stages
Also, non ActorMaterializers can opt-into providing a logger here.
* +str #18794 add javadsl for StageLogging
* fix missing test method on compile only class
* StreamSupervisor as system actor so that it is
stopped after ordinary actors
* when transport is shutdown send flush message to all
outbound associations (over control stream) and wait for ack
or timeout
A SubFlow (or SubSource) is not a Graph, it is an unfinished builder
that accepts transformations. This allows us to capture the substreams’
transformations before materializing the flow, which will be very
helpful in fully fusing all operators.
Another change is that groupBy now requires a maxSubstreams parameter in
order to bound its resource usage. In exchange the matching merge can be
unbounded. This trades silent deadlock for explicit stream failure.
This commit also changes all uses of Predef.identity to use `conforms`
and removes the HTTP impl.util.identityFunc.
main work by @drewhk with contributions from @2m and @rkuhn
This work uncovered many well-hidden bugs in existing Stages, in
particular StatefulStage. These were hidden by the behavior of
OneBoundedInterpreter that normally behaves more orderly than it
guarantees in general, especially with respect to the timeliness of
delivery of upstream termination signals; the bugs were then that
internal state was not flushed when onComplete arrived “too early”.
