Hi Jun, thanks for your mail.
Thank you for your questions!
I think they are really good and tackle the core of the problem I see.
I will answer inline, mostly but still want to set the tone here.
The core strength of kafka is what Martin once called the
kappa-Architecture. How does this work?
You have everything as a log as in kafka. When you need to change
something.
You create the new version of your application and leave it running in
parallel.
Once the new version is good you switch your users to use the new
application.
The online reshuffling effectively breaks this architecture and I think
the switch in thinking here is more harmful
than any details about the partitioning function to allow such a change. I
feel with my suggestion we are the closest to
the original and battle proven architecture and I can only warn to move
away from it.
I might have forgotten something, sometimes its hard for me to getting all
the thoughts captured in a mail, but I hope the comments inline will
further make my concern clear, and put some emphasis on why I prefer my
solution ;)
One thing we should all be aware of when discussing this, and I think Dong
should have mentioned it (maybe he did).
We are not discussing all of this out of thin air but there is an effort
in the Samza project.
https://cwiki.apache.org/confluence/display/SAMZA/SEP-5%3A+
Enable+partition+expansion+of+input+streams
https://issues.apache.org/jira/browse/SAMZA-1293
To be clear. I think SEP-5 (state of last week, dont know if it adapted to
this discussion) is on a way better path than KIP-253, and I can't really
explain why.
Best Jan,
nice weekend everyone
On 09.03.2018 03:36, Jun Rao wrote:
Hi, Jan,
Thanks for the feedback. Just some comments on the earlier points that you
mentioned.
50. You brought up the question of whether existing data needs to be
copied
during partition expansion. My understand of your view is that avoid
copying existing data will be more efficient, but it doesn't work well
with
compacted topics since some keys in the original partitions will never be
cleaned. It would be useful to understand your use case of compacted
topics
a bit more. In the common use case, the data volume in a compacted topic
may not be large. So, I am not sure if there is a strong need to expand
partitions in a compacted topic, at least initially.
I do agree. State is usually smaller. Update rates might be also
competitively high.
Doing Log-compaction (even beeing very efficient and configurable) is also
a more expensive operation than
just discarding old segments. Further if you want to use more consumers
processing the events
you also have to grow the number of partitions. Especially for use-cases
we do (KIP-213) a tiny state full
table might be very expensive to process if it joins against a huge table.
I can just say we have been in the spot of needing to grow log compacted
topics. Mainly for processing power we can bring to the table.
Further i am not at all concerned about the extra spaced used by "garbage
keys". I am more concerned about the correctness of innocent consumers. The
logic becomes complicated. Say for streams one would need to load the
record into state but not forward it the topology ( to have it available
for shuffeling). I rather have it simple and a topic clean regardless if it
still has its old partition count. Especially with multiple partitions
growth's I think it becomes insanely hard to to this shuffle correct. Maybe
Streams and Samza can do it. Especially if you do "hipster stream
processing" <https://www.confluent.io/blog/introducing-kafka-streams-
stream-processing-made-simple/>. This makes kafka way to complicated.
With my approach I think its way simpler because the topic has no "history"
in terms of partitioning but is always clean.
51. "Growing the topic by an integer factor does not require any state
redistribution at all." Could you clarify this a bit more? Let's say you
have a consumer app that computes the windowed count per key. If you
double
the number of partitions from 1 to 2 and grow the consumer instances from
1
to 2, we would need to redistribute some of the counts to the new consumer
instance. Regarding to linear hashing, it's true that it won't solve the
problem with compacted topics. The main benefit is that it redistributes
the keys in one partition to no more than two partitions, which could help
redistribute the state.
You don't need to spin up a new consumer in this case. every consumer
would just read every partition with the (partition % num_task) task. it
will still have the previous data right there and can go on.
This sounds contradictory to what I said before, but please bear with me.
52. Good point on coordinating the expansion of 2 topics that need to be
joined together. This is where the 2-phase partition expansion could
potentially help. In the first phase, we could add new partitions to the 2
topics one at a time but without publishing to the new patitions. Then, we
can add new consumer instances to pick up the new partitions. In this
transition phase, no reshuffling is needed since no data is coming from
the
new partitions. Finally, we can enable the publishing to the new
partitions.
I think its even worse than you think. I would like to introduce the Term
transitive copartitioning. Imagine
2 streams application. One joins (A,B) the other (B,C) then there is a
transitive copartition requirement for
(A,C) to be copartitioned aswell. This can spread significantly and
require many consumers to adapt at the same time.
It is also not entirely clear to me how you not need reshuffling in this
case. If A has a record that never gets updated after the expansion and the
coresponding B record moves to a new partition. How shall they meet w/o
shuffle?
53. "Migrating consumer is a step that might be made completly unnecessary
if - for example streams - takes the gcd as partitioning scheme instead of
enforcing 1 to 1." Not sure that I fully understand this. I think you mean
that a consumer application can run more instances than the number of
partitions. In that case, the consumer can just repartitioning the input
data according to the number of instances. This is possible, but just has
the overhead of reshuffling the data.
No what I meant is ( that is also your question i think Mathias) that if
you grow a topic by a factor.
Even if your processor is statefull you can can just assign all the
multiples of the previous partition to
this consumer and the state to keep processing correctly will be present
w/o any shuffling.
Say you have an assignment
Statefull consumer => partition
0 => 0
1 => 1
2 => 2
and you grow you topic by 4 you get,
0 => 0,3,6,9
1 => 1,4,7,10
2 => 2,5,8,11
Say your hashcode is 8. 8%3 => 2 before so consumer for partition 2 has
it.
Now you you have 12 partitions so 8%12 => 8, so it goes into partition 8
which is assigned to the same consumer
who had 2 before and therefore knows the key.
Userland reshuffeling is there as an options. And it does exactly what I
suggest. And I think its the perfect strategie. All I am suggestion is
broker side support to switch the producers to the newly partitioned topic.
Then the old (to few partition topic) can go away. Remember the list of
steps in the beginning of this thread. If one has broker support for all
where its required and streams support for those that aren’t necessarily.
Then one has solved the problem.
I repeat it because I think its important. I am really happy that you
brought that up! because its 100% what I want just with the differences to
have an option to discard the to small topic later (after all consumers
adapted). And to have order correct there. I need broker support managing
the copy process + the produces and fence them against each other. I also
repeat. the copy process can run for weeks in the worst case. Copying the
data is not the longest task migrating consumers might very well be.
Once all consumers switched and copying is really up to date (think ISR
like up to date) only then we stop the producer, wait for the copy to
finish and use the new topic for producing.
After this the topic is perfect in shape. and no one needs to worry about
complicated stuff. (old keys hanging around might arrive in some other
topic later.....). can only imagine how many tricky bugs gonna arrive after
someone had grown and shrunken is topic 10 times.
54. "The other thing I wanted to mention is that I believe the current
suggestion (without copying data over) can be implemented in pure userland
with a custom partitioner and a small feedbackloop from ProduceResponse =>
Partitionier in coorporation with a change management system." I am not
sure a customized partitioner itself solves the problem. We probably need
some broker side support to enforce when the new partitions can be used.
We
also need some support on the consumer/kstream side to preserve the per
key
ordering and potentially migrate the processing state. This is not trivial
and I am not sure if it's ideal to fully push to the application space.
Broker support is defenitly the preferred way here. I have nothing against
broker support.
I tried to say that for what I would preffer - copying the data over, at
least for log compacted topics -
I would require more broker support than the KIP currently offers.
Jun
On Tue, Mar 6, 2018 at 10:33 PM, Jan Filipiak <jan.filip...@trivago.com>
wrote:
Hi Dong,
are you actually reading my emails, or are you just using the thread I
started for general announcements regarding the KIP?
I tried to argue really hard against linear hashing. Growing the topic by
an integer factor does not require any state redistribution at all. I
fail
to see completely where linear hashing helps on log compacted topics.
If you are not willing to explain to me what I might be overlooking: that
is fine.
But I ask you to not reply to my emails then. Please understand my
frustration with this.
Best Jan
On 06.03.2018 19:38, Dong Lin wrote:
Hi everyone,
Thanks for all the comments! It appears that everyone prefers linear
hashing because it reduces the amount of state that needs to be moved
between consumers (for stream processing). The KIP has been updated to
use
linear hashing.
Regarding the migration endeavor: it seems that migrating producer
library
to use linear hashing should be pretty straightforward without
much operational endeavor. If we don't upgrade client library to use
this
KIP, we can not support in-order delivery after partition is changed
anyway. Suppose we upgrade client library to use this KIP, if partition
number is not changed, the key -> partition mapping will be exactly the
same as it is now because it is still determined using murmur_hash(key)
%
original_partition_num. In other words, this change is backward
compatible.
Regarding the load distribution: if we use linear hashing, the load may
be
unevenly distributed because those partitions which are not split may
receive twice as much traffic as other partitions that are split. This
issue can be mitigated by creating topic with partitions that are
several
times the number of consumers. And there will be no imbalance if the
partition number is always doubled. So this imbalance seems acceptable.
Regarding storing the partition strategy as per-topic config: It seems
not
necessary since we can still use murmur_hash as the default hash
function
and additionally apply the linear hashing algorithm if the partition
number
has increased. Not sure if there is any use-case for producer to use a
different hash function. Jason, can you check if there is some use-case
that I missed for using the per-topic partition strategy?
Regarding how to reduce latency (due to state store/load) in stream
processing consumer when partition number changes: I need to read the
Kafka
Stream code to understand how Kafka Stream currently migrate state
between
consumers when the application is added/removed for a given job. I will
reply after I finish reading the documentation and code.
Thanks,
Dong
On Mon, Mar 5, 2018 at 10:43 AM, Jason Gustafson <ja...@confluent.io>
wrote:
Great discussion. I think I'm wondering whether we can continue to leave
Kafka agnostic to the partitioning strategy. The challenge is
communicating
the partitioning logic from producers to consumers so that the
dependencies
between each epoch can be determined. For the sake of discussion,
imagine
you did something like the following:
1. The name (and perhaps version) of a partitioning strategy is stored
in
topic configuration when a topic is created.
2. The producer looks up the partitioning strategy before writing to a
topic and includes it in the produce request (for fencing). If it
doesn't
have an implementation for the configured strategy, it fails.
3. The consumer also looks up the partitioning strategy and uses it to
determine dependencies when reading a new epoch. It could either fail
or
make the most conservative dependency assumptions if it doesn't know
how
to
implement the partitioning strategy. For the consumer, the new
interface
might look something like this:
// Return the partition dependencies following an epoch bump
Map<Integer, List<Integer>> dependencies(int
numPartitionsBeforeEpochBump,
int numPartitionsAfterEpochBump)
The unordered case then is just a particular implementation which never
has
any epoch dependencies. To implement this, we would need some way for
the
consumer to find out how many partitions there were in each epoch, but
maybe that's not too unreasonable.
Thanks,
Jason
On Mon, Mar 5, 2018 at 4:51 AM, Jan Filipiak <jan.filip...@trivago.com
wrote:
Hi Dong
thank you very much for your questions.
regarding the time spend copying data across:
It is correct that copying data from a topic with one partition
mapping
to
a topic with a different partition mapping takes way longer than we can
stop producers. Tens of minutes is a very optimistic estimate here.
Many
people can not afford copy full steam and therefore will have some
rate
limiting in place, this can bump the timespan into the day's. The good
part
is that the vast majority of the data can be copied while the producers
are
still going. One can then, piggyback the consumers ontop of this
timeframe,
by the method mentioned (provide them an mapping from their old offsets
to
new offsets in their repartitioned topics. In that way we separate
migration of consumers from migration of producers (decoupling these
is
what kafka is strongest at). The time to actually swap over the
producers
should be kept minimal by ensuring that when a swap attempt is started
the
consumer copying over should be very close to the log end and is
expected
to finish within the next fetch. The operation should have a time-out
and
should be "reattemtable".
Importance of logcompaction:
If a producer produces key A, to partiton 0, its forever gonna be
there,
unless it gets deleted. The record might sit in there for years. A new
producer started with the new partitions will fail to delete the
record
in
the correct partition. Th record will be there forever and one can not
reliable bootstrap new consumers. I cannot see how linear hashing can
solve
this.
Regarding your skipping of userland copying:
100%, copying the data across in userland is, as far as i can see,
only
a
usecase for log compacted topics. Even for logcompaction + retentions
it
should only be opt-in. Why did I bring it up? I think log compaction
is
a
very important feature to really embrace kafka as a "data plattform".
The
point I also want to make is that copying data this way is completely
inline with the kafka architecture. it only consists of reading and
writing
to topics.
I hope it clarifies more why I think we should aim for more than the
current KIP. I fear that once the KIP is done not much more effort
will
be
taken.
On 04.03.2018 02:28, Dong Lin wrote:
Hey Jan,
In the current proposal, the consumer will be blocked on waiting for
other
consumers of the group to consume up to a given offset. In most cases,
all
consumers should be close to the LEO of the partitions when the
partition
expansion happens. Thus the time waiting should not be long e.g. on
the
order of seconds. On the other hand, it may take a long time to wait
for
the entire partition to be copied -- the amount of time is
proportional
to
the amount of existing data in the partition, which can take tens of
minutes. So the amount of time that we stop consumers may not be on
the
same order of magnitude.
If we can implement this suggestion without copying data over in
purse
userland, it will be much more valuable. Do you have ideas on how
this
can
be done?
Not sure why the current KIP not help people who depend on log
compaction.
Could you elaborate more on this point?
Thanks,
Dong
On Wed, Feb 28, 2018 at 10:55 PM, Jan Filipiak<Jan.Filipiak@trivago.
com
wrote:
Hi Dong,
I tried to focus on what the steps are one can currently perform to
expand
or shrink a keyed topic while maintaining a top notch semantics.
I can understand that there might be confusion about "stopping the
consumer". It is exactly the same as proposed in the KIP. there
needs
to
be
a time the producers agree on the new partitioning. The extra
semantics I
want to put in there is that we have a possibility to wait until all
the
existing data
is copied over into the new partitioning scheme. When I say stopping
I
think more of having a memory barrier that ensures the ordering. I
am
still
aming for latencies on the scale of leader failovers.
Consumers have to explicitly adapt the new partitioning scheme in
the
above scenario. The reason is that in these cases where you are
dependent
on a particular partitioning scheme, you also have other topics that
have
co-partition enforcements or the kind -frequently. Therefore all your
other
input topics might need to grow accordingly.
What I was suggesting was to streamline all these operations as best
as
possible to have "real" partition grow and shrinkage going on.
Migrating
the producers to a new partitioning scheme can be much more
streamlined
with proper broker support for this. Migrating consumer is a step
that
might be made completly unnecessary if - for example streams - takes
the
gcd as partitioning scheme instead of enforcing 1 to 1. Connect
consumers
and other consumers should be fine anyways.
I hope this makes more clear where I was aiming at. The rest needs to
be
figured out. The only danger i see is that when we are introducing
this
feature as supposed in the KIP, it wont help any people depending on
log
compaction.
The other thing I wanted to mention is that I believe the current
suggestion (without copying data over) can be implemented in pure
userland
with a custom partitioner and a small feedbackloop from
ProduceResponse
=>
Partitionier in coorporation with a change management system.
Best Jan
On 28.02.2018 07:13, Dong Lin wrote:
Hey Jan,
I am not sure if it is acceptable for producer to be stopped for a
while,
particularly for online application which requires low latency. I
am
also
not sure how consumers can switch to a new topic. Does user
application
needs to explicitly specify a different topic for producer/consumer
to
subscribe to? It will be helpful for discussion if you can provide
more
detail on the interface change for this solution.
Thanks,
Dong
On Mon, Feb 26, 2018 at 12:48 AM, Jan
Filipiak<Jan.Filipiak@trivago.
com
wrote:
Hi,
just want to throw my though in. In general the functionality is
very
usefull, we should though not try to find the architecture to hard
while
implementing.
The manual steps would be to
create a new topic
the mirrormake from the new old topic to the new topic
wait for mirror making to catch up.
then put the consumers onto the new topic
(having mirrormaker spit out a mapping from old offsets to
new
offsets:
if topic is increased by factor X there is gonna be a
clean
mapping from 1 offset in the old topic to X offsets in the new
topic,
if there is no factor then there is no chance to
generate a
mapping that can be reasonable used for continuing)
make consumers stop at appropriate points and continue
consumption
with offsets from the mapping.
have the producers stop for a minimal time.
wait for mirrormaker to finish
let producer produce with the new metadata.
Instead of implementing the approach suggest in the KIP which will
leave
log compacted topic completely crumbled and unusable.
I would much rather try to build infrastructure to support the
mentioned
above operations more smoothly.
Especially having producers stop and use another topic is
difficult
and
it would be nice if one can trigger "invalid metadata" exceptions
for
them
and
if one could give topics aliases so that their produces with the
old
topic
will arrive in the new topic.
The downsides are obvious I guess ( having the same data twice for
the
transition period, but kafka tends to scale well with datasize). So
its a
nicer fit into the architecture.
I further want to argument that the functionality by the KIP can
completely be implementing in "userland" with a custom partitioner
that
handles the transition as needed. I would appreciate if someone
could
point
out what a custom partitioner couldn't handle in this case?
With the above approach, shrinking a topic becomes the same steps.
Without
loosing keys in the discontinued partitions.
Would love to hear what everyone thinks.
Best Jan
On 11.02.2018 00:35, Dong Lin wrote:
Hi all,
I have created KIP-253: Support in-order message delivery with
partition
expansion. See
https://cwiki.apache.org/confluence/display/KAFKA/KIP-253%
3A+Support+in-order+message+delivery+with+partition+expansion
.
This KIP provides a way to allow messages of the same key from
the
same
producer to be consumed in the same order they are produced even
if
we
expand partition of the topic.
Thanks,
Dong
