kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From John Roesler <j...@confluent.io>
Subject Re: Can kafka internal state be purged ?
Date Fri, 28 Jun 2019 23:20:37 GMT
Ok, good, that's what I was hoping. I think that's a good strategy, at
the end of the "real" data, just write a dummy record with the same
keys with a high timestamp to flush everything else through.

For the most part, I'd expect a production program to get a steady
stream of traffic with increasing timestamps, so windows would be
constantly be getting flushed out as stream time moves forward.

Some folks have reported that they don't get enough traffic through
their program to flush out the suppressed results on a regular basis,
though. Right now, the best solution is to have the same Producer that
writes data to the input topics also write "heartbeat/dummy" records
periodically when there is no data to send, just to keep stream time
moving forward. But this isn't a perfect solution, as you have pointed
out in this thread; you really want the heartbeat records to go to all
partitions, and also go to all re-partitions if there are repartition
topics in the topology.

I agree that there seems to be a need for some first-class support for
keeping stream time moving reliably. I think that the ideal would be
to allow Producers to automatically send the heartbeats and to
implement a variant of the Chandy-Lamport distributed snapshot
algorithm to push them though the whole topology while skipping any
actual computations (so your business logic wouldn't have to see
them). I'd really love to see this feature in Streams; I just haven't
written it up yet because I haven't had time.

During the design for Suppress, we did consider having some kind of
timer, but the problem is that it's not possible for this to be
deterministic. If you want the "until window closes" version, you're
supposed to get a guarantee that you'll really only see one, final,
result for each window/key. If we were to use the system clock on the
Streams machine to decide it's probably been long enough and emit the
"final" result, but it turns out that we actually had just stalled
(maybe waiting for quorum during a broker upgrade or something, or
just a run-of-the-mill networking problem) and the next record we poll
was supposed to be in the window, what can we do? We already emitted
the "final" result, so we can say "oops, that wasn't the final result,
_this_ one is the final result", but that seems to render the words
"final result" kind of meaningless. On the other hand, we can just
drop that record and ignore it, but that's a bummer because the only
reason we couldn't include it in the result was some ephemeral
environmental problem. If we run the same data through the same
program again, we'd get a different result.

So for "until window closes" mode, where we guarantee you _only_ see
the final results, we only offer stream time expiration. Anything else
would violate correctness one way or another. On the other hand, you
have "until time limit" mode. In that case, it's just "buffer for a
while, but multiple results are still ok" semantics. For that case, we
have https://cwiki.apache.org/confluence/display/KAFKA/KIP-424%3A+Allow+suppression+of+intermediate+events+based+on+wall+clock+time
, to indeed use a timer to emit events if too much time passes on the
Streams side. It's just not implemented yet.

Does this all seem about right to you?
-john

On Wed, Jun 26, 2019 at 12:57 PM Parthasarathy, Mohan <mparthas@hpe.com> wrote:
>
> Initially it started in the testing. QA reported problems where "events" were not detected
after they finished their testing. After this discussion, my proposal was to send a few more
records to cause the windows to flush so that the suppressed event would show up. Now it looks
to me, these few dummy records have to match the "key" of the pending windows. Then it would
be flushed.
>
> In practice, it may not be a problem always. But then the real time nature of the problem
might require us that there is not a huge delay between the processing of the event and the
flush. How does one solve this issue in production ? I am wondering why the design did not
accommodate a timer to flush the windows ?
>
> Thanks
> Mohan
>
>
> ´╗┐On 6/26/19, 8:18 AM, "John Roesler" <john@confluent.io> wrote:
>
>     Hi Mohan,
>
>     I see where you're going with this, and it might indeed be a
>     challenge. Even if you send a "dummy" message on all input topics, you
>     won't have a guarantee that after the repartition, the dummy message
>     is propagated to all partitions of the repartition topics. So it might
>     be difficult to force the suppression buffer to flush if it's after a
>     repartition.
>
>     Can we take a step back and discuss the motivation for forcing the
>     records to flush out? Is this for testing your app, or is it to drive
>     some production logic?
>
>     Thanks,
>     -John
>
>
>     On Mon, Jun 24, 2019 at 7:26 PM Parthasarathy, Mohan <mparthas@hpe.com> wrote:
>     >
>     > John,
>     >
>     > Thanks for the nice explanation. When the repartitioning happens, does the window
get associated with the new partition i.e., now does a message with new timestamp has to appear
on the repartition topic for the window to expire ? It is possible that there is new stream
of messages coming in but post-map operation, the partitions in the repartitioned topic does
not see the same thing.
>     >
>     > Thanks
>     > Mohan
>     >
>     > On 6/24/19, 7:49 AM, "John Roesler" <john@confluent.io> wrote:
>     >
>     >     Hey, this is a very apt question.
>     >
>     >     GroupByKey isn't a great example because it doesn't actually change
>     >     the key, so all the aggregation results are actually on records from
>     >     the same partition. But let's say you do a groupBy or a map (or any
>     >     operation that can change the key), followed by an aggregation. Now
>     >     it's possible that the aggregation would need to process records from
>     >     two different partitions. In such a case (key-changing operation
>     >     followed by a stateful operation), Streams actually round-trips the
>     >     data through an intermediate topic, called a repartition topic, before
>     >     the aggregation. This has the effect, similar to the "shuffle" phase
>     >     of map-reduce, of putting all the data into its *new* right partition,
>     >     so then the aggregation can still process each of its partitions
>     >     independently.
>     >
>     >     Regarding the latter statement, even though you only have one
>     >     instance, Streams _still_ processes each partition independently. The
>     >     "unit of work" responsible for processing a partition is called a
>     >     "task". So if you have 4 partitions, then your one instance actually
>     >     has 4 state stores, one for each task, where each task only gets
>     >     records from a single partition. The tasks can't see anything about
>     >     each other, not their state nor other metadata like their current
>     >     stream time. Otherwise, the results would depend on which tasks happen
>     >     to be co-located with which other tasks. So, having to send your
>     >     "purge" event to all partitions is a pain, but in the end, it buys you
>     >     a lot, as you can add another instance to your cluster at any time,
>     >     and Streams will scale up, and you'll know that the program is
>     >     executing exactly the same way the whole time.
>     >
>     >     -John
>     >
>     >     On Sat, Jun 22, 2019 at 4:37 PM Parthasarathy, Mohan <mparthas@hpe.com>
wrote:
>     >     >
>     >     > I can see the issue. But it raised other questions. Pardon my ignorance.
Even though partitions are processed independently, windows can be aggregating state from
records read from many partitions. Let us say there is a groupByKey followed by aggregate.
In this case how is the state reconciled across all the application instances ? Is there a
designated instance for a particular key ?
>     >     >
>     >     > In my case, there was only one instance processing records from all
partitions and it is kind of odd that windows did not expire even though I understand why
now.
>     >     >
>     >     > Thanks
>     >     > Mohan
>     >     >
>     >     >
>     >     > On 6/21/19, 2:25 PM, "John Roesler" <john@confluent.io> wrote:
>     >     >
>     >     >     No problem. It's definitely a subtlety. It occurs because each
>     >     >     partition is processed completely independently of the others,
so
>     >     >     "stream time" is tracked per partition, and there's no way to look
>     >     >     across at the other partitions to find out what stream time they
have.
>     >     >
>     >     >     In general, it's not a problem because you'd expect all partitions
to
>     >     >     receive updates over time, but if you're specifically trying to
send
>     >     >     events that cause stuff to get flushed from the buffers, it can
mess
>     >     >     with you. It's especially notable in tests. So, for most tests,
I just
>     >     >     configure the topics to have one partition.
>     >     >
>     >     >     -John
>     >     >
>     >     >     On Fri, Jun 21, 2019 at 3:56 PM Parthasarathy, Mohan <mparthas@hpe.com>
wrote:
>     >     >     >
>     >     >     > That change "In the same partition" must explain what we are
seeing. Unless you see one message per partition, all windows will not expire. That is an
interesting twist. Thanks for the correction ( I will go back and confirm this.
>     >     >     >
>     >     >     > -mohan
>     >     >     >
>     >     >     >
>     >     >     > On 6/21/19, 12:40 PM, "John Roesler" <john@confluent.io>
wrote:
>     >     >     >
>     >     >     >     Sure, the record cache attempts to save downstream operators
from
>     >     >     >     unnecessary updates by also buffering for a short amount
of time
>     >     >     >     before forwarding. It forwards results whenever the cache
fills up or
>     >     >     >     whenever there is a commit. If you're happy to wait at
least "commit
>     >     >     >     interval" amount of time for updates, then you don't need
to do
>     >     >     >     anything, but if you're on the edge of your seat, waiting
for these
>     >     >     >     results, you can set cache.max.bytes.buffering to 0 to
disable the
>     >     >     >     record cache entirely. Note that this would hurt throughput
in
>     >     >     >     general, though.
>     >     >     >
>     >     >     >     Just a slight modification:
>     >     >     >     * a new record with new timestamp > (all the previous
timestamps +
>     >     >     >     grace period) will cause all the old windows *in the same
partition*
>     >     >     >     to close
>     >     >     >     * yes, expiry of the window depends only on the event
time
>     >     >     >
>     >     >     >     Hope this helps!
>     >     >     >     -John
>     >     >     >
>     >     >     >     On Thu, Jun 20, 2019 at 11:42 AM Parthasarathy, Mohan
<mparthas@hpe.com> wrote:
>     >     >     >     >
>     >     >     >     > Could you tell me a little more about the delays
about the record caches and how I can disable it ?
>     >     >     >     >
>     >     >     >     >  If I could summarize my problem:
>     >     >     >     >
>     >     >     >     > -A new record with a new timestamp > all records
sent before, I expect *all* of the old windows to close
>     >     >     >     > -Expiry of the windows depends only on the event
time and not on the key
>     >     >     >     >
>     >     >     >     > Are these two statements correct ?
>     >     >     >     >
>     >     >     >     > Thanks
>     >     >     >     > Mohan
>     >     >     >     >
>     >     >     >     > On 6/20/19, 9:17 AM, "John Roesler" <john@confluent.io>
wrote:
>     >     >     >     >
>     >     >     >     >     Hi!
>     >     >     >     >
>     >     >     >     >     In addition to setting the grace period to zero
(or some small
>     >     >     >     >     number), you should also consider the delays
introduced by record
>     >     >     >     >     caches upstream of the suppression. If you're
closely watching the
>     >     >     >     >     timing of records going into and coming out of
the topology, this
>     >     >     >     >     might also spoil your expectations. You could
always disable the
>     >     >     >     >     record cache to make the system more predictable
(although this would
>     >     >     >     >     hurt throughput in production).
>     >     >     >     >
>     >     >     >     >     Thanks,
>     >     >     >     >     -John
>     >     >     >     >
>     >     >     >     >     On Wed, Jun 19, 2019 at 3:01 PM Parthasarathy,
Mohan <mparthas@hpe.com> wrote:
>     >     >     >     >     >
>     >     >     >     >     > We do explicitly set the grace period to
zero. I am going to try the new version
>     >     >     >     >     >
>     >     >     >     >     > -mohan
>     >     >     >     >     >
>     >     >     >     >     >
>     >     >     >     >     > On 6/19/19, 12:50 PM, "Parthasarathy, Mohan"
<mparthas@hpe.com> wrote:
>     >     >     >     >     >
>     >     >     >     >     >     Thanks. We will give it a shot.
>     >     >     >     >     >
>     >     >     >     >     >     On 6/19/19, 12:42 PM, "Bruno Cadonna"
<bruno@confluent.io> wrote:
>     >     >     >     >     >
>     >     >     >     >     >         Hi Mohan,
>     >     >     >     >     >
>     >     >     >     >     >         I realized that my previous statement
was not clear. With a grace
>     >     >     >     >     >         period of 12 hour, suppress would
wait for late events until stream
>     >     >     >     >     >         time has advanced 12 hours before
a result would be emitted.
>     >     >     >     >     >
>     >     >     >     >     >         Best,
>     >     >     >     >     >         Bruno
>     >     >     >     >     >
>     >     >     >     >     >         On Wed, Jun 19, 2019 at 9:21 PM
Bruno Cadonna <bruno@confluent.io> wrote:
>     >     >     >     >     >         >
>     >     >     >     >     >         > Hi Mohan,
>     >     >     >     >     >         >
>     >     >     >     >     >         > if you do not set a grace period,
the grace period defaults to 12
>     >     >     >     >     >         > hours. Hence, suppress would
wait for an event that occurs 12 hour
>     >     >     >     >     >         > later before it outputs a result.
Try to explicitly set the grace
>     >     >     >     >     >         > period to 0 and let us know
if it worked.
>     >     >     >     >     >         >
>     >     >     >     >     >         > If it still does not work,
upgrade to version 2.2.1 if it is possible
>     >     >     >     >     >         > for you. We had a couple of
bugs in suppress recently that are fixed
>     >     >     >     >     >         > in that version.
>     >     >     >     >     >         >
>     >     >     >     >     >         > Best,
>     >     >     >     >     >         > Bruno
>     >     >     >     >     >         >
>     >     >     >     >     >         > On Wed, Jun 19, 2019 at 8:37
PM Parthasarathy, Mohan <mparthas@hpe.com> wrote:
>     >     >     >     >     >         > >
>     >     >     >     >     >         > > No, I have not set any
grace period. Is that mandatory ? Have you seen problems with suppress and windows expiring
?
>     >     >     >     >     >         > >
>     >     >     >     >     >         > > Thanks
>     >     >     >     >     >         > > Mohan
>     >     >     >     >     >         > >
>     >     >     >     >     >         > > On 6/19/19, 12:41 AM,
"Bruno Cadonna" <bruno@confluent.io> wrote:
>     >     >     >     >     >         > >
>     >     >     >     >     >         > >     Hi Mohan,
>     >     >     >     >     >         > >
>     >     >     >     >     >         > >     Did you set a grace
period on the window?
>     >     >     >     >     >         > >
>     >     >     >     >     >         > >     Best,
>     >     >     >     >     >         > >     Bruno
>     >     >     >     >     >         > >
>     >     >     >     >     >         > >     On Tue, Jun 18, 2019
at 2:04 AM Parthasarathy, Mohan <mparthas@hpe.com> wrote:
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >     > On further debugging,
what we are seeing is that windows are expiring rather randomly as new messages are being
processed. . We tested with new key for every new message. We waited for the window time before
replaying new messages. Sometimes a new message would come in and create state. It takes several
messages to make some of the old windows to be closed (go past suppress to the next stage).
We have also seen where one of them never closed even but several other older ones expired.
 Then we explicitly sent a message with the same old key and then it showed up. Also, for
every new message, only one of the previous window expires even though there are several pending.
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >     > If we don't use
suppress, then there is never an issue. With suppress, the behavior we are seeing is weird.
We are using 2.1.0 version in DSL mode. Any clues on what we could be missing ? Why isn't
there an order in the way windows are closed ? As event time progresses by the new messages
arriving, the older ones should expire. Is that right understanding or not ?
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >     > Thanks
>     >     >     >     >     >         > >     > Mohan
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >     > On 6/17/19, 3:43
PM, "Parthasarathy, Mohan" <mparthas@hpe.com> wrote:
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >     >     Hi,
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >     >     We are using
suppress in the application. We see some state being created at some point in time. Now there
is no new data for a day or two. We send new data but the old window of data (where we see
the state being created) is not closing i.e not seeing it go through suppress and on to the
next stage. It is as though the state created earlier was purged. Is this possible ?
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >     >     Thanks
>     >     >     >     >     >         > >     >     Mohan
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >     >
>     >     >     >     >     >         > >
>     >     >     >     >     >         > >
>     >     >     >     >     >
>     >     >     >     >     >
>     >     >     >     >     >
>     >     >     >     >     >
>     >     >     >     >
>     >     >     >     >
>     >     >     >
>     >     >     >
>     >     >
>     >     >
>     >
>     >
>
>

Mime
View raw message