[
https://issues.apache.org/jira/browse/BEAM-1581?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15926687#comment-15926687
]
Aviem Zur edited comment on BEAM-1581 at 3/15/17 6:24 PM:
----------------------------------------------------------
I think we should avoid exposing a contract to the user which promises writing
JSONs but accepts strings.
This is a loose contract which will leave JSON validity up to the user. If the
user does not create valid JSON Strings errors can occur.
Errors which might be detected very late in the process, possibly only upon an
attempt to consume the data in another process (which may belong to a different
user as JSON is often used for integration).
We definitely need concrete {{JsonSink}} and {{JsonSource}} classes which
extend the existing abstract {{FileBasedSink}} and {{FileBasedSource}}. But
these should not be used directly by the user.
In order to avoid exposing classes which deal with Strings to the user we need
concrete {{PTransform}} classes which deal with objects.
The problem is these probably can't exist in a {{JsonIO}} class since it cannot
have the transformations from object to JSON string (since there are several
ways to implement this).
Should these transforms be in a separate class such as JacksonIO?All common
JSON file logic regarding how the file should be constructed (As [~jkff]
mentioned this should be better defined) will be in the abstract sink and
source, including all file writing/reading related code (Inherited from
{{FileBasedSink}} and {{FileBasedSource}}).
So none of these are actually an abstract {{PTransform}}.
The {{PTransform}} that the user will use will be concrete, found in
{{JacksonIO}}
was (Author: aviemzur):
I think we should avoid exposing a contract to the user which promises writing
JSONs but accepts strings.
This is a loose contract which will leave JSON validity up to the user. If the
user does not create valid JSON Strings errors can occur.
Errors which might be detected very late in the process, possibly only upon an
attempt to consume the data in another process (which may belong to a different
user as JSON is often used for integration).
I agree that {{JsonIO}} itself should not be abstract and should not implement
any interface (same as other {{XxxIO}} classes that exist today).
It will be the enclosing class of {{JsonSink}} and {{JsonSource}} classes which
extend the existing abstract {{FileBasedSink}} and {{FileBasedSource}}. These
will not be exposed to the user.
These will be implemented by concrete classes such as {{JacksonSink}} and
{{JacksonSource}}.
All common JSON file logic regarding how the file should be constructed (As
[~jkff] mentioned this should be better defined) will be in the abstract sink
and source, including all file writing/reading related code (Inherited from
{{FileBasedSink}} and {{FileBasedSource}}).
So none of these are actually an abstract {{PTransform}}.
The {{PTransform}} that the user will use will be concrete, found in
{{JacksonIO}}
> JSON sources and sinks
> ----------------------
>
> Key: BEAM-1581
> URL: https://issues.apache.org/jira/browse/BEAM-1581
> Project: Beam
> Issue Type: New Feature
> Components: sdk-java-extensions
> Reporter: Aviem Zur
> Assignee: Aviem Zur
>
> JSON source and sink to read/write JSON files.
> Similarly to {{XmlSource}}/{{XmlSink}}, these be a {{JsonSource}}/{{JonSink}}
> which are a {{FileBaseSource}}/{{FileBasedSink}}.
> Consider using methods/code (or refactor these) found in {{AsJsons}} and
> {{ParseJsons}}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
