You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@avro.apache.org by Wai Yip Tung <wy...@tungwaiyip.info> on 2015/02/11 02:01:46 UTC
Schema design guideline, strict v.s. lenient
During our development of schema based data pipeline, we often run into
a debate. Should we make the schema tight and strict so that all the
application error can be tested and caught early? Or should we design
the schema to be lenient, because inevitably the schema is going to be
evolved and the data we have found in our system often contains
variations despite our effort constraint it.
Slowly I observed that the difference in school of thought is largely
related to their role. The data producer, mainly the application
developers, wants the schema to be strict (e.g. required attribute, no
union of 'null'). They see this as a debugging tool. They expect errors
to be caught by the encoder during unit test. They expect the production
system to raise alarm loudly if a bad build break things.
The consumers, mainly the data backend developers and the analysts, want
the schema to be lenient. The backend developers often have to reprocess
historical data. Strict schema is often incompatible and cause big
problem in reading historical data. They aruge having some data, even if
slightly broken, is better than having no data.
We have been having difficulty to strike a balance. It leads me to think
perhaps we need more than a single schema in operation. Perhaps an
application developer will create a strict schema. And the backend
application will derive a lenient version from it in order to load all
historical data successfully.
I am wondering if others have seen this kind of tension. Any thought on
how to address this?
Wai Yip
Re: Schema design guideline, strict v.s. lenient
Posted by Andrew Ehrlich <an...@aehrlich.com>.
I have noticed that data consuming people will prefer flat records
because they are easier to query. I have yet to find a good tool to
query unstructured records like JSON. A large amount of time and effort
therefore goes into the ETL process.
Maybe one could fork the data flow and send raw records to an "raw" bin
and send the the other fork through a process that conforms each records
to a schema in a schema library.
On 2/10/15 5:01 PM, Wai Yip Tung wrote:
> During our development of schema based data pipeline, we often run
> into a debate. Should we make the schema tight and strict so that all
> the application error can be tested and caught early? Or should we
> design the schema to be lenient, because inevitably the schema is
> going to be evolved and the data we have found in our system often
> contains variations despite our effort constraint it.
>
> Slowly I observed that the difference in school of thought is largely
> related to their role. The data producer, mainly the application
> developers, wants the schema to be strict (e.g. required attribute, no
> union of 'null'). They see this as a debugging tool. They expect
> errors to be caught by the encoder during unit test. They expect the
> production system to raise alarm loudly if a bad build break things.
>
> The consumers, mainly the data backend developers and the analysts,
> want the schema to be lenient. The backend developers often have to
> reprocess historical data. Strict schema is often incompatible and
> cause big problem in reading historical data. They aruge having some
> data, even if slightly broken, is better than having no data.
>
> We have been having difficulty to strike a balance. It leads me to
> think perhaps we need more than a single schema in operation. Perhaps
> an application developer will create a strict schema. And the backend
> application will derive a lenient version from it in order to load all
> historical data successfully.
>
> I am wondering if others have seen this kind of tension. Any thought
> on how to address this?
>
> Wai Yip