Hey Daniyar, Thanks for looking at it!
Something like your screenshot is more along the lines of what I was thinking. Sorry, but I didn't follow what you mean, how would that not be "vanilla java"? Unfortunately the deserializer needs more information, though. For example, what if the inner type is a Map<String,String>? The serde could only be used to produce a LinkedList<Map>, thus, we'd still need an inner serde, like you have in the KIP (Serde<T> innerSerde). Something more like Serde<LinkedList<MyRecord>> = Serdes.listSerde( /**list type**/ LinkedList.class, /**inner serde**/ new MyRecordSerde() ) And in configuration, it's something like: default.key.serde: org...ListSerde default.key.list.serde.type: java.util.LinkedList default.key.list.serde.inner: com.mycompany.MyRecordSerde What do you think? Thanks, -John On Thu, Jun 20, 2019 at 2:46 PM Development <d...@yeralin.net> wrote: > Hey John, > > I gave read about TypeReference. It could work for the list serde. > However, it is not directly supported: > https://github.com/FasterXML/jackson-databind/issues/1490 > The only way is to pass an actual class object into the constructor, > something like: > > It could be an option, but not a pretty one. What do you think of my > approach to use vanilla java and canonical class name? (As described > previously) > > Best, > Daniyar Yeralin > > On Jun 20, 2019, at 2:45 PM, Development <d...@yeralin.net> wrote: > > Hi John, > > Thank you for your input! Yes, my idea looks a little bit over engineered > :) > > I also wanted to see a feedback from Mathias as well since he gave me an > idea about storing fixed/variable size entries. > > Best, > Daniyar Yeralin > > On Jun 18, 2019, at 6:06 PM, John Roesler <j...@confluent.io> wrote: > > Hi Daniyar, > > That's a very clever solution! > > One observation is that, now, this is what we might call a polymorphic > serde. That is, you're detecting the actual concrete type and then > promising to produce the exact same concrete type on read. There are > some inherent problems with this approach, which in general require > some kind of schema registry (not necessarily Schema Registry, just > any registry for schemas) to solve. > > Notice that every serialized record has quite a bit of duplicated > information: the concrete type as well as a byte to indicate whether > the value type is a fixed size, and, if so, an integer to indicate the > actual size. These constitute a schema, of sorts, because they tell us > later how exactly to deserialize the data. Unfortunately, this > information is completely redundant. In all likelihood, the > information will be exactly the same for every record in the topic. > This problem is essentially the core motivation for serializations > like Avro: to move the schema outside of the serialization itself, so > that the records won't contain so much redundant information. > > In this light, I'm wondering if it makes sense to go back to something > like what you had earlier in which you don't support perfectly > preserving the concrete type for _this_ serde, but instead just > support deserializing to _some_ List. Then, you could defer full, > perfect, type preservation to serdes that have an external system in > which to register their type information. > > There does exist an alternative, if we really do want to preserve the > concrete type (which does seem kind of nice). You can add a > configuration option specifically for the serde to configure what the > list type will be, and maybe what the element type is, as well. > > As far as "related work" goes, you might be interested to take a look > at how Jackson can be configured to deserialize into a specific, > arbitrarily nested, generically parameterized class structure. > Specifically, you might find > > https://fasterxml.github.io/jackson-core/javadoc/2.0.0/com/fasterxml/jackson/core/type/TypeReference.html > interesting. > > Thanks, > -John > > On Mon, Jun 17, 2019 at 12:38 PM Development <d...@yeralin.net> wrote: > > > bump > > > >