kafka-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Matthias J. Sax" <matth...@confluent.io>
Subject Re: KIP-213- [DISCUSS] - Three follow-up discussion points - topic partitioning, serializers, hashers
Date Tue, 19 Mar 2019 00:13:21 GMT
Just my 2 cents. Not sure if others see it differently:

1) it seems that we can lift the restriction on having the same number
of input topic partitions, and thus we should exploit this IMHO; don't
see why we should enforce an artificial restriction


2) for the value serde it's a little bit more tricky; in general, Apache
Kafka should not be concerned with third party tools. It seems that
https://issues.apache.org/jira/browse/KAFKA-7777 might provide a
solution though -- but it's unclear if KIP-213 and 7777 would be shipped
with the same release...

> To me, this is a shortcoming of the Confluent Avro Serde
>> that will likely need to be fixed on that side. 

I agree (good to know...)


3) I am not an expert on hashing, but 128-bit murmur3 sounds reasonable
to me



Btw: I think we can have this discussion on the PR -- no need to concern
the mailing list (it's a lot of people that are subscribed).



-Matthias

On 3/17/19 5:20 PM, Adam Bellemare wrote:
> Hey folks
> 
> I have been implementing the KIP as outlined in
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-213+Support+non-key+joining+in+KTable,
> and I have run into a few points to consider that we did not include in the
> original.
> 
> *1) Do all input topics need to have the same partitions or not?* Currently
> I have it designed such that it must, to be consistent with other joins.
> However, consider the following:
> TableA - 5 partitions
> TableB - 2 partitions
> Pre-subscribe Repartition Topic = 2 partitions, 2 RHS processors
> Post-Subcribe Repartition Topic = 5 partitions, 5 LHS processors
> 
> Would this not be possible? Is there a value in flexibility to this? I have
> not looked deeper into the restrictions of this approach, so if there is
> something I should know I would appreciate a heads up.
> 
> *2) Is it appropriate to use the KTable valueSerde during the computation
> of the hash?* To compute the hash I need to obtain an array of bytes, which
> is immediately possible by  using the valueSerde. However, the Confluent
> Kafka Schema Registry serializer fails when it is being used in this way:
> In the hash generating code, I set topic to null because the data is not
> dependent on any topic value. I simply want the serialized bytes to input
> into the hash function.
> *    byte[] preHashValue = serializer.serialize(topic = null, data)*
> 
> Any KTable that is Consumed.with(Confluent-Key-Serde,
> Confluent-Value-Serde) will automatically try to register the schema to
> topic+"-key" and topic+"-value". If I pass in null, it tries to register to
> "-key" and "-value" each time the serializer is called, regardless of the
> class. In other words, it registers the schemas to a null topic and fails
> any subsequent serializations that aren't of the exact same schema. Note
> that this would be the case across ALL applications using the confluent
> schema registry. To me, this is a shortcoming of the Confluent Avro Serde
> that will likely need to be fixed on that side. However, it does bring up
> the question - is this an appropriate way to use a serializer? Alternately,
> if I should NOT use the KTable value-serde to generate the byte array, does
> anyone have a better idea?
> 
> *3) How big of a hash value do we need? Does the Foreign Key even matter
> for resolving?*
> I am currently looking at fast, non-cryptologically-secure hash options. We
> use murmur2 already in Kafka, but it is only 32 bits. I have been looking
> at murmur3hash as implemented in the Apache Hive project (
> https://github.com/apache/hive/blob/master/storage-api/src/java/org/apache/hive/common/util/Murmur3.java)
> - it supports 128-bit hashes and is allegedly more performant than MD5.
> With a 128-bit hash. The birthday problem indicates that we would have a
> 50% chance of a collision with 2^64 = 1.8446744e+19 entries. I believe that
> this is sufficiently small, especially for our narrow time window, to
> expect a collision for a singly-keyed event. I think that there is no
> benefit in including the foreign key, but again, please let me know if this
> is wrong.
> 
> 
> Thanks All
> 


Mime
View raw message