pnowojski commented on a change in pull request #10029: [FLINK-14553][runtime]
Respect non-blocking output in StreamTask#processInput
URL: https://github.com/apache/flink/pull/10029#discussion_r340033249
##########
File path:
flink-streaming-java/src/main/java/org/apache/flink/streaming/runtime/tasks/StreamTask.java
##########
@@ -281,10 +289,41 @@ protected void processInput(DefaultActionContext
context) throws Exception {
if (status == InputStatus.END_OF_INPUT) {
context.allActionsCompleted();
}
- else if (status == InputStatus.NOTHING_AVAILABLE) {
+
+ CompletableFuture<?> jointFuture =
getInputOutputJointFuture(status);
+ if (jointFuture != null) {
SuspendedMailboxDefaultAction suspendedDefaultAction =
context.suspendDefaultAction();
-
inputProcessor.getAvailableFuture().thenRun(suspendedDefaultAction::resume);
+ jointFuture.thenRun(suspendedDefaultAction::resume);
+ }
+ }
+
+ /**
+ * @return a combination of input and output futures if at-least one
future of them is not
+ * completed, otherwise return null if all input and outputs are
available.
+ */
+ private CompletableFuture<?> getInputOutputJointFuture(InputStatus
status) {
+ if (status == InputStatus.MORE_AVAILABLE &&
isOutputAvailable()) {
+ return null;
+ }
+
+ int length = recordWriters.size();
+ for (int i = 0; i < length; i++) {
+ inputOutputFutures[i] =
recordWriters.get(i).getAvailableFuture();
+ }
+ inputOutputFutures[length] =
inputProcessor.getAvailableFuture();
+ return CompletableFuture.allOf(inputOutputFutures);
+ }
+
+ /**
+ * @return true if all the record writers are available.
+ */
+ private boolean isOutputAvailable() {
+ for (RecordWriter recordWriter : recordWriters) {
+ if (!recordWriter.isAvailable()) {
Review comment:
I think there is a bug in `LocalBufferPool#requestMemorySegment` in this
code block:
```
if (isBlocking && isUnavailable()) {
availabilityHelper.resetUnavailable();
}
```
`availabilityHelper` should be marked unavailable always, not only if this
blocking request. Or am I missing something? (If not, we need a test case for
this)
Also the non synchronised `recordWriter.isAvailable()` works here only if
the memory request that made the buffer pool unavailable came from the
mailbox/task thread (or established some other accidental happen before
relationship by taking/releasing some locks)? This is just a question to think
through, as probably such behaviour is fine.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
