lct45 commented on a change in pull request #9157:
URL: https://github.com/apache/kafka/pull/9157#discussion_r483685325
##########
File path:
streams/src/main/java/org/apache/kafka/streams/kstream/internals/KStreamSlidingWindowAggregate.java
##########
@@ -192,29 +213,125 @@ public void processInOrder(final K key, final V value,
final long timestamp) {
//create left window for new record
if (!leftWinAlreadyCreated) {
final ValueAndTimestamp<Agg> valueAndTime;
- //there's a right window that the new record could create -->
new record's left window is not empty
- if (latestLeftTypeWindow != null) {
+ // if there's a right window that the new record could create
&& previous record falls within left window -> new record's left window is not
empty
+ if (previousRecordTimestamp != null &&
leftWindowNotEmpty(previousRecordTimestamp, timestamp)) {
valueAndTime = ValueAndTimestamp.make(leftWinAgg.value(),
timestamp);
} else {
valueAndTime = ValueAndTimestamp.make(initializer.apply(),
timestamp);
}
final TimeWindow window = new TimeWindow(timestamp -
windows.timeDifferenceMs(), timestamp);
putAndForward(window, valueAndTime, key, value, closeTime,
timestamp);
}
- //create right window for new record
if (!rightWinAlreadyCreated && rightWindowIsNotEmpty(rightWinAgg,
timestamp)) {
- final TimeWindow window = new TimeWindow(timestamp + 1,
timestamp + 1 + windows.timeDifferenceMs());
- final ValueAndTimestamp<Agg> valueAndTime =
ValueAndTimestamp.make(getValueOrNull(rightWinAgg),
Math.max(rightWinAgg.timestamp(), timestamp));
+ createCurrentRecordRightWindow(timestamp, rightWinAgg, key);
+ }
+ }
+
+ /**
+ * Created to handle records where 0 < timestamp < timeDifferenceMs.
These records would create
+ * windows with negative start times, which is not supported. Instead,
they will fall within the [0, timeDifferenceMs]
+ * window, and we will update or create their right windows as new
records come in later
+ */
+ private void processEarly(final K key, final V value, final long
timestamp, final long closeTime) {
+ // A window from [0, timeDifferenceMs] that holds all early records
+ KeyValue<Windowed<K>, ValueAndTimestamp<Agg>> combinedWindow =
null;
+ ValueAndTimestamp<Agg> rightWinAgg = null;
+ boolean rightWinAlreadyCreated = false;
+ final Set<Long> windowStartTimes = new HashSet<>();
+
+ Long previousRecordTimestamp = null;
+
+ try (
+ final KeyValueIterator<Windowed<K>, ValueAndTimestamp<Agg>>
iterator = windowStore.fetch(
+ key,
+ key,
+ Math.max(0, timestamp - 2 * windows.timeDifferenceMs()),
+ // to catch the current record's right window, if it
exists, without more calls to the store
+ timestamp + 1)
+ ) {
+ KeyValue<Windowed<K>, ValueAndTimestamp<Agg>> next;
+ while (iterator.hasNext()) {
+ next = iterator.next();
+ windowStartTimes.add(next.key.window().start());
+ final long startTime = next.key.window().start();
+ final long windowMaxRecordTimestamp =
next.value.timestamp();
+
+ if (startTime == 0) {
+ combinedWindow = next;
+ if (windowMaxRecordTimestamp < timestamp) {
+ // If maxRecordTimestamp > timestamp, the current
record is out-of-order, meaning that the
+ // previous record's right window would have been
created already by other records. This
+ // will always be true for early records, as they
all fall within [0, timeDifferenceMs].
Review comment:
Going back and re-reading this comments in the context of Matthias's
later comment, it seems like maybe this one is unnecessary and could be
combined with the later one. Correct me if I'm wrong, but it feels like the
statement `if (windowMaxRecordTimestamp < timestamp) {
previousRecordTimestamp =
windowMaxRecordTimestamp;
}`
is somewhat self explanatory and only needs justification when we set the
`rightWinAgg` later.
----------------------------------------------------------------
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:
us...@infra.apache.org
