Can someone confirm that each
partition has its own stream time and that the stream time for a partition
only advances when a record is written to the partition after the window
closes?
That's correct.
On 5/21/24 10:11 AM, Chad Preisler wrote:
After reviewing the logs, I think I understand what happens with the
repartition topics. Looks like they will be assigned to one or more
instances. In my example I ran three instances of the application (A, B,
C). Looks like the two repartition topics got assigned to A and B. The six
partitions from the input topics got split evenly across all three running
instances A, B, and C. Since the repartitioned streams are what I'm joining
on, I guess the join will run on two instances, and any input topic
processing will run across all three. Is that correct?
Still would like clarification regarding some records appearing to not get
processed: I think the issue is related to certain partitions not getting
records to advance stream time (because of low volume). Can someone confirm
that each partition has its own stream time and that the stream time for a
partition only advances when a record is written to the partition after the
window closes?
On Tue, May 21, 2024 at 10:27 AM Chad Preisler <chad.preis...@gmail.com>
wrote:
See one small edit below...
On Tue, May 21, 2024 at 10:25 AM Chad Preisler <chad.preis...@gmail.com>
wrote:
Hello,
I think the issue is related to certain partitions not getting records to
advance stream time (because of low volume). Can someone confirm that each
partition has its own stream time and that the stream time for a partition
only advances when a record is written to the partition after the window
closes?
If I use the repartition method on each input topic to reduce the number
of partitions for those streams, how many instances of the application will
process records? For example, if the input topics each have 6 partitions,
and I use the repartition method to set the number of partitions for the
streams to 2, how many instances of the application will process records?
Thanks,
Chad
On Wed, May 1, 2024 at 6:47 PM Matthias J. Sax <mj...@apache.org> wrote:
How do you know this?
First thing we do is write a log message in the value joiner. We
don't see
the log message for the missed records.
Well, for left/right join results, the ValueJoiner would only be called
when the window is closed... And for invalid input (or late record, ie,
which arrive out-of-order and their window was already closes), records
would be dropped right away. So you cannot really infer that a record
did make it into the join or not, or what happens if it did make it into
the `Processor`.
-> https://kafka.apache.org/documentation/#kafka_streams_task_monitoring
`dropped-records-total` is the name of the metric.
-Matthias
On 5/1/24 11:35 AM, Chad Preisler wrote:
Hello,
We did some testing in our test environment today. We are seeing some
records processes where only one side of the join has a record. So
that's
good. However, we are still seeing some records get skipped. They
never hit
the value joiner (we write a log message first thing in the value
joiner).
During the test we were putting some load on the system, so stream
time was
advancing. We did notice that the join windows were taking much longer
than
30 minutes to close and process records. Thirty minutes is the window
plus
grace.
How do you know this?
First thing we do is write a log message in the value joiner. We don't
see
the log message for the missed records.
I will try pushing the same records locally. However, we don't see any
errors in our logs and the stream does process one sided joins after
the
skipped record. Do you have any docs on the "dropper records" metric?
I did
a Google search and didn't find many good results for that.
Thanks,
Chad
On Tue, Apr 30, 2024 at 2:49 PM Matthias J. Sax <mj...@apache.org>
wrote:
Thanks for the information. I ran the code using Kafka locally.
After
submitting some records inside and outside of the time window and
grace,
the join performed as expected when running locally.
That gives some hope :)
However, they never get into the join.
How do you know this?
Did you check the metric for dropper records? Maybe records are
considers malformed and dropped? Are you using the same records in
production and in your local test?
Are there any settings for the stream client that would affect the
join?
Not that I can think of... There is one more internal config, but as
long as data is flowing, it should not impact the result you see.
Are there any settings on the broker side that would affect the
join?
No. The join is computed client side. Broker configs should not have
any
impact.
f I increase the log level for the streams API would that
shed some light on what is happening?
I don't think it would help much. The code in question is
org.apache.kafka.streams.kstream.internals.KStreamKStreamJoin -- but
it
does not do any logging except WARN for the already mentioned
"dropping
malformed" records that is also recorded via JMX.
WARN: "Skipping record due to null key or value. "
If you can identify a specific record from the input which would
produce
an output, but does not, maybe you can try to feed it into your local
test env and try to re-produce the issue?
-Matthias
On 4/30/24 11:38 AM, Chad Preisler wrote:
Matthias,
Thanks for the information. I ran the code using Kafka locally. After
submitting some records inside and outside of the time window and
grace,
the join performed as expected when running locally.
I'm not sure why the join is not working as expected when running
against
our actual brokers. We are peeking at the records for the streams
and we
are seeing the records get pulled. However, they never get into the
join.
It's been over 24 hours since the expected records were created, and
there
has been plenty of traffic to advance the stream time. Only records
that
have both a left and right side match are getting processed by the
join.
Are there any settings for the stream client that would affect the
join?
Are there any settings on the broker side that would affect the join?
The outer join is just one part of the topology. Compared to running
it
locally there is a lot more data going through the app when running
on
our
actual servers. If I increase the log level for the streams API would
that
shed some light on what is happening? Does anyone know if there are
specific packages that I should increase the log level for? Any
specific
log message I can hone in on to tell me what is going on?
Basically, I'm looking for some pointers on where I can start
looking.
Thanks,
Chad
On Tue, Apr 30, 2024 at 10:26 AM Matthias J. Sax <mj...@apache.org>
wrote:
I expect the join to
execute after the 25 with one side of the join containing a
record and
the
other being null
Given that you also have a grace period of 5 minutes, the result
will
only be emitted after the grace-period passed and the window is
closed
(not when window end time is reached).
One has a
naming convention of "KSTREAM_OUTERSHARED". I see a record there,
but
I'm
not sure how to decode that message to see what is in it. What is
the
purpose of those messages?
It's an internal store, that stores all records which are subject
to be
emitted as left/right join result, ie, if there is no inner join
result.
The format used is internal:
https://github.com/apache/kafka/blob/trunk/streams/src/main/java/org/apache/kafka/streams/state/internals/LeftOrRightValueSerde.java
Also note: time is based on event-time, ie, if the input stream
stops to
send new records, "stream-time" will stop to advance and the result
might not be emitted because the window does not get closed.
(Last, there is some internal wall-clock time delay of one second to
emit results for performance reasons...)
HTH.
-Matthias
On 4/30/24 6:51 AM, Chad Preisler wrote:
Hello,
I have a KStream to KStream outer join with a time difference of 25
minutes
and 5 minutes of grace. When I get a record for one side of the
join,
but
don't get a record on the other side of the join, I expect the
join to
execute after the 25 with one side of the join containing a record
and
the
other being null. Is that correct? If it is correct, it's not
working
for
me.
I was poking around on the broker and saw some internal topics. I
see
the
key I expected to execute the join on some of those topics. One
has a
naming convention of "KSTREAM_OUTERSHARED". I see a record there,
but
I'm
not sure how to decode that message to see what is in it. What is
the
purpose of those messages? If I decode the message will it help me
see
when
the join should have been executed?
I also see the key on a topic with the naming convention
"KSTREAM_OUTERTHIS".
Are there any other topics that I should be looking at to
troubleshoot
this
issue?
Thanks,
Chad
