KAFKA-13264: fix inMemoryWindowStore backward fetch not in reversed order by showuon · Pull Request #11292 · apache/kafka

showuon · 2021-09-02T08:12:39Z

When introducing backward iterator for WindowStroe in #9138, we forgot to make "each segment" in reverse order (i.e. in descendingMap) in InMemoryWindowStore. Fix it and add integration tests for it.

Currently, in Window store, we store records in [segments -> [records] ].

For example:
window size = 500,
input records:

key: "a", value: "aa", timestamp: 0 ==> will be in [0, 500] window
key: "b", value: "bb", timestamp: 10 ==> will be in [0, 500] window
key: "c", value: "cc", timestamp: 510 ==> will be in [500, 1000] window

So, internally, the "a" and "b" will be in the same segment, and "c" in another segments.
segments: [0 /* window start */, records], [500, records].
And the records for window start 0 will be "a" and "b".
the records for window start 500 will be "c".

Before this change, we did have a reverse iterator for segments, but not in "records". So, when doing backwardFetchAll, we'll have the records returned in order: "c", "a", "b", which should be "c", "b", "a" obviously.

So, back to the question: why did the original test cases not catch this issue?
It's because the test input are all in different window start timestamp, which will have different different segments:

 private void putFirstBatch(final WindowStore<Integer, String> store,
                               @SuppressWarnings("SameParameterValue") final long startTime,
                               final InternalMockProcessorContext context) {
        context.setRecordContext(createRecordContext(startTime));
        store.put(0, "zero", startTime);
        store.put(1, "one", startTime + 1L);
        store.put(2, "two", startTime + 2L);
        store.put(3, "three", startTime + 2L);  // <-- this is the new record I added, to test multiple records in the same segment case

        store.put(4, "four", startTime + 4L);
        store.put(5, "five", startTime + 5L);
    }

I added an additional record for AbstractWindowBytesStoreTest test. InWindowStoreFetchTest, we will produce records in timestamp:0, 1, 500, 501, 502, which will be put into window: [0, 500] * 2 and [500, 1000] * 3. And we try to fetch them forward/backward, to see if the results are as expected, i.e.: in reverse order should be 502, 501, 500, 1, 0.

The behavior works as expected in RocksDBWindowStore.

Committer Checklist (excluded from commit message)

Verify design and implementation
Verify test coverage and CI build status
Verify documentation (including upgrade notes)

showuon · 2021-09-02T09:22:08Z

        store.put(0, "zero", startTime);
        store.put(1, "one", startTime + 1L);
        store.put(2, "two", startTime + 2L);
+        store.put(3, "three", startTime + 2L);


add 2 records at the same timestamp to test the forward and backward fetch cases.

showuon · 2021-09-02T09:24:43Z

-            if (allKeys) {
-                return currentSegment.getValue().entrySet().iterator();
+            final ConcurrentNavigableMap<Bytes, byte[]> subMap = allKeys ?
+                currentSegment.getValue() :
+                currentSegment.getValue().subMap(keyFrom, true, keyTo, true);
+
+            if (forward) {
+                return subMap.entrySet().iterator();
            } else {
-                return currentSegment.getValue().subMap(keyFrom, true, keyTo, true).entrySet().iterator();
+                return subMap.descendingMap().entrySet().iterator();


Before this change, when setting records iterator, we only consider the allKey case, not the forward/backward cases. Fix it.

showuon · 2021-09-02T09:27:56Z

@jeqo @ableegoldman @guozhangwang , please help review this PR. Thank you.

showuon · 2021-09-03T02:03:59Z

+        final List<StoreType> types = Arrays.asList(StoreType.InMemory, StoreType.RocksDB, StoreType.Timed);
+        final List<Boolean> logging = Arrays.asList(true, false);
+        final List<Boolean> caching = Arrays.asList(true, false);
+        final List<Boolean> forward = Arrays.asList(true, false);


Test all kinds of combination.

jeqo

LGTM. Great catch @showuon, thanks!

Love the parametrized integration tests to validate each combination.

ableegoldman · 2021-09-08T00:53:00Z

+
+    private TimeWindowedKStream<String, String> windowedStream;
+
+    public WindowStoreFetchIntegrationTest(final StoreType storeType, final boolean enableLogging, final boolean enableCaching, final boolean forward) {


Seems more like a unit test than an integration (though a particularly thorough one😜 )

ableegoldman

LGTM, but can you move the parametrized test out of the integration tests? It's not exactly a clear definition, but we usually consider as integration test the more heavy/long-running tests, eg those that start up an actual KafkaStreams, read/write/verify data, etc

guozhangwang

Thanks for the fix @showuon ! It lgtm. I also feel the same as @ableegoldman that it seems a unit test would suffice, which would take much less time as well :)

…ation test

comptooladmin · 2021-09-08T08:40:20Z

+    final KeyValue<Windowed<Integer>, String> zero = windowedPair(0, "zero", defaultStartTime);
+    final KeyValue<Windowed<Integer>, String> one = windowedPair(1, "one", defaultStartTime + 1);
+    final KeyValue<Windowed<Integer>, String> two = windowedPair(2, "two", defaultStartTime + 2);
+    final KeyValue<Windowed<Integer>, String> three = windowedPair(3, "three", defaultStartTime + 2);


Is this supposed to be defaultStartTime + 3?

Thanks for your comment, but here, I make it as defaultStartTime + 2 on purpose, to test the case that when window starts time is the same, the forward/backward query API can return the order as expected. I updated the PR description to make it clear. Thank you.

showuon · 2021-09-08T10:12:12Z

@ableegoldman @guozhangwang , thanks for your comments. I agree that it doesn't need to be integration test. I've moved it out from integration test. Thank you.

Failed tests are unrelated.

    Build / JDK 8 and Scala 2.12 / kafka.server.KRaftClusterTest.testCreateClusterAndPerformReassignment()
    Build / JDK 8 and Scala 2.12 / kafka.server.KRaftClusterTest.testCreateClusterAndCreateListDeleteTopic()
    Build / JDK 11 and Scala 2.13 / kafka.server.DelayedOperationTest.testDelayedFuture()

guozhangwang

Thanks @showuon . I still have a few questions in mind just for my own clarification:

Do you know why the original test cases in AbstractWindowBytesStoreTest, like shouldGetBackwardAll and testBackwardFetchRange did not capture this bug? This test class is leveraged by the in-memory stores as well.
Related to 1), what additional coverage does the new WindowStoreFetchTest provides in addition to the above two test cases?

showuon · 2021-09-09T02:36:17Z

@guozhangwang , those are good questions. Let me answer them below:

1. Do you know why the original test cases in AbstractWindowBytesStoreTest, like `shouldGetBackwardAll` and `testBackwardFetchRange` did not capture this bug? This test class is leveraged by the in-memory stores as well.

That's right, those tests also tested in-memory stores, but it didn't test multiple records in the same window cases. Currently, in Window store, we store records in [segments -> [records] ].

For example:
window size = 500,
input records:

key: "a", value: "aa", timestamp: 0 ==> will be in [0, 500] window
key: "b", value: "bb", timestamp: 10 ==> will be in [0, 500] window
key: "c", value: "cc", timestamp: 510 ==> will be in [500, 1000] window

So, internally, the "a" and "b" will be in the same segment, and "c" in another segments.
segments: [0 /* window start */, records], [500, records].
And the records for window start 0 will be "a" and "b".
the records for window start 500 will be "c".

Before this change, we did have a reverse iterator for segments, but not in "records". So, when doing backwardFetchAll, we'll have the records returned in order: "c", "a", "b", which should be "c", "b", "a" obviously.

So, back to the question: why did the original test cases not catch this issue?
It's because the test input are all in different window start timestamp, which will have different different segments:

 private void putFirstBatch(final WindowStore<Integer, String> store,
                               @SuppressWarnings("SameParameterValue") final long startTime,
                               final InternalMockProcessorContext context) {
        context.setRecordContext(createRecordContext(startTime));
        store.put(0, "zero", startTime);
        store.put(1, "one", startTime + 1L);
        store.put(2, "two", startTime + 2L);
        store.put(3, "three", startTime + 2L);  // <-- this is the new record I added, to test multiple records in the same segment case

        store.put(4, "four", startTime + 4L);
        store.put(5, "five", startTime + 5L);
    }

2. Related to 1), what additional coverage does the new `WindowStoreFetchTest` provides in addition to the above two test cases?

I think I've added above. I added an additional record for AbstractWindowBytesStoreTest test. In WindowStoreFetchTest, we will produce records in timestamp:0, 1, 500, 501, 502, which will be put into window: [0, 500] * 2 and [500, 1000] * 3. And we try to fetch them forward/backward, to see if the results are as expected, i.e.: in reverse order should be 502, 501, 500, 1, 0.

The behavior works as expected in RocksDBWindowStore.

Hope that's clear.
I also updated in the PR description.
Thank you.

guozhangwang · 2021-09-13T21:39:23Z

Hope that's clear.
I also updated in the PR description.
Thank you.

Thanks!

guozhangwang · 2021-09-13T21:41:00Z

Merged to trunk.

…rder (apache#11292) When introducing backward iterator for WindowStroe in apache#9138, we forgot to make "each segment" in reverse order (i.e. in descendingMap) in InMemoryWindowStore. Fix it and add integration tests for it. Currently, in Window store, we store records in [segments -> [records] ]. For example: window size = 500, input records: key: "a", value: "aa", timestamp: 0 ==> will be in [0, 500] window key: "b", value: "bb", timestamp: 10 ==> will be in [0, 500] window key: "c", value: "cc", timestamp: 510 ==> will be in [500, 1000] window So, internally, the "a" and "b" will be in the same segment, and "c" in another segments. segments: [0 /* window start */, records], [500, records]. And the records for window start 0 will be "a" and "b". the records for window start 500 will be "c". Before this change, we did have a reverse iterator for segments, but not in "records". So, when doing backwardFetchAll, we'll have the records returned in order: "c", "a", "b", which should be "c", "b", "a" obviously. Reviewers: Jorge Esteban Quilcate Otoya <quilcate.jorge@gmail.com>, Anna Sophie Blee-Goldman <ableegoldman@apache.org>, Guozhang Wang <wangguoz@gmail.com>

showuon added 2 commits September 2, 2021 16:09

KAFKA-13264: fix inMemoryStateStore backward fetch not in reversed order

89310f9

KAFKA-13264: add tests for the records at the same time

0cd5a0a

showuon changed the title ~~[WIP] KAFKA-13264: fix inMemoryWindowStore backward fetch not in reversed order~~ KAFKA-13264: fix inMemoryWindowStore backward fetch not in reversed order Sep 2, 2021

showuon commented Sep 2, 2021

View reviewed changes

showuon mentioned this pull request Sep 2, 2021

KAFKA-13211: add support for infinite range query for WindowStore #11227

Merged

3 tasks

showuon commented Sep 3, 2021

View reviewed changes

jeqo approved these changes Sep 3, 2021

View reviewed changes

ableegoldman reviewed Sep 8, 2021

View reviewed changes

ableegoldman approved these changes Sep 8, 2021

View reviewed changes

guozhangwang reviewed Sep 8, 2021

View reviewed changes

showuon added 2 commits September 8, 2021 14:35

Merge branch 'trunk' of https://github.com/apache/kafka into KAFKA-13264

18ed310

KAFKA-13264: move the WindowStoreFetchIntegrationTest out from integr…

4e66e8f

…ation test

comptooladmin reviewed Sep 8, 2021

View reviewed changes

guozhangwang reviewed Sep 8, 2021

View reviewed changes

guozhangwang merged commit 9628c12 into apache:trunk Sep 13, 2021


		private TimeWindowedKStream<String, String> windowedStream;

		public WindowStoreFetchIntegrationTest(final StoreType storeType, final boolean enableLogging, final boolean enableCaching, final boolean forward) {

Conversation

showuon commented Sep 2, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Committer Checklist (excluded from commit message)

Uh oh!

showuon Sep 2, 2021

Choose a reason for hiding this comment

Uh oh!

showuon Sep 2, 2021

Choose a reason for hiding this comment

Uh oh!

showuon commented Sep 2, 2021

Uh oh!

showuon Sep 3, 2021

Choose a reason for hiding this comment

Uh oh!

jeqo left a comment

Choose a reason for hiding this comment

Uh oh!

ableegoldman Sep 8, 2021

Choose a reason for hiding this comment

Uh oh!

ableegoldman left a comment

Choose a reason for hiding this comment

Uh oh!

guozhangwang left a comment

Choose a reason for hiding this comment

Uh oh!

comptooladmin Sep 8, 2021

Choose a reason for hiding this comment

Uh oh!

showuon Sep 8, 2021

Choose a reason for hiding this comment

Uh oh!

showuon commented Sep 8, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

guozhangwang left a comment

Choose a reason for hiding this comment

Uh oh!

showuon commented Sep 9, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

guozhangwang commented Sep 13, 2021

Uh oh!

guozhangwang commented Sep 13, 2021

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

showuon commented Sep 2, 2021 •

edited

Loading

showuon commented Sep 8, 2021 •

edited

Loading

showuon commented Sep 9, 2021 •

edited

Loading