Hi, Tom, First, your observation about current InMemorySystem is exactly right and thanks for raising this issue to the community!
The current InMemorySystem came up with a tight coupling with the Samza test framework, which I believe put quite a lot of limitations on its uses, e.g. using InMemorySystem in prototyping. Currently we are working on ways to improve it so it can be used by normal user code. For your case, it seems to me the inconsistency of using serde caused the confusion. You point of being more consistent to support Serde in InMemory stream sounds reasonable to me. I have the same impression that InMemory can be treated the same way as other input streams. The initial rollout doesn't have this feature and I create the ticket https://issues.apache.org/jira/browse/SAMZA-2075 to track this. We will see whether we can get it there in the next release. Thanks, Xinyu On Thu, Jan 17, 2019 at 6:11 AM Tom Davis <t...@recursivedream.com> wrote: > Hey Sanil, thanks for the reply. I eventually figured that not > supporting serdes for in-memory streams was an intentional restriction, > I was just pointing out that it is inconsistent with earlier versions > since it was relatively easy to supply stream serdes directly before the > Descriptor API. > > I can't really send a test case along because it's all in Clojure and > uses a Clojure-based API wrapper I wrote for interacting with Samza. In > theory, the easiest test would be one where a Config contains the > property I mentioned; with that, you should be able to run a simple > pipeline that shows -- despite the NoOpSerde forced by > InMemorySystemDescriptor -- the input is serialized using that serde. > > Anyway, I'm not sure if it's worth the trouble. I get why you'd forgo > serialization for the in-memory system, it was just a handy way to test > my entire pipeline which contains a few non-trivial custom serdes. > > > Sanil Jain <sanil.jai...@gmail.com> writes: > > > Hi Tom, > > > > InMemorySystem is a system that is supposed to only support NoOpSerde > since > > all the associated steams for this system are maintained in memory. In > > addition to this, if your test is using the Samza's Test Framework, it > will > > override any explicit serde configs specified for streams to NoOp. > > > > > > You are expected to supply deserialized objects to the in-memory system. > > > > > > In addition to that in your email you mentioned: > > > > > > {unformat} > > > > I had still specified in my config: > > > > streams.in-0.samza.msg.serde=integer > > > > > > Apparently, that *was* respected by some part of the system because > > integers were > > deserialized properly! Removing this configuration value results in my > > operator > > receiving a byte array since the in-memory system only uses NoOpSerde. > > > > {unformat} > > > > > > Can you send me a snippet of test you were trying to fix so that I can > > understand the problem better? > > > > > > Thanks > > > > Sanil > > > > On Tue, 8 Jan 2019 at 17:28, Tom Davis <t...@recursivedream.com> wrote: > > > >> I am in the process of updating a project to 1.0 and spent today > debugging > >> a > >> rather odd test failure. When using input/output streams with > IntegerSerde, > >> things worked fine -- however, using LongSerde, every message value was > 0! > >> I > >> eventually found that InMemorySystemDescriptor#getInputDescriptor > ignores > >> the > >> serde passed to it. However, I had still specified in my config: > >> > >> streams.in-0.samza.msg.serde=integer > >> > >> Apparently that *was* respected by some part of the system because > >> integers were > >> deserialized properly! Removing this configuration value results in my > >> operator > >> receiving a byte array since the in-memory system only uses NoOpSerde. > >> > >> This behavior appears inconsistent with the previous version of Samza. > The > >> old > >> `getInputStream` was passed a serde that was always used, but since the > new > >> version receives a Descriptor that has already discarded the serde, I am > >> forced > >> into assuming NoOpSerde everywhere, at least for testing purposes. > >> > >> Not the end of the world, but it does introduce an inconsistency between > >> the > >> in-memory system and any other -- one that requires a fair bit of domain > >> knowledge to avoid. > >> > >> As always, thanks for the great project! > >> >
