[
https://issues.apache.org/jira/browse/GIRAPH-277?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13427736#comment-13427736
]
Jaeho Shin commented on GIRAPH-277:
-----------------------------------
Thanks Eli.
It would be great to entirely eliminate the need of writing I/O code and let
users pick from a large collection, but I think that is unlikely if complex
data structure is used in the input graph. We can easily exhaust the possible
combinations of int, double, Text I/O formats, but Map and List were the types
of values and messages mostly in my cases. I admit some parts could look over
engineered, and don't really reduce the net amount of code, but the big picture
is clearly to make the blanks in the code obvious for users and only ask them
to implement what they really need to.
I didn't have chance to test these text formats on large scale, but I'm using
identically structured format codes for reading and writing Hive tables of
>100M rows, ~15GB without problem. The holes user needs to fill in this case
are how to get vertex id, value and edge maps from a particular row in the
table, instead of a single text line.
About the checkstyle xml, I was reluctant to make the change because it might
encourage us to abuse suppress comments, but thought it was also silly to blow
up the xml itself. I'm in no way a checkstyle expert, and just found [good
hints|http://stackoverflow.com/a/4023351/390044] from the web :)
> Text Vertex Input/Output Format base classes overhaul
> -----------------------------------------------------
>
> Key: GIRAPH-277
> URL: https://issues.apache.org/jira/browse/GIRAPH-277
> Project: Giraph
> Issue Type: Improvement
> Components: examples, lib
> Reporter: Jaeho Shin
> Attachments: GIRAPH-277.patch
>
>
> The current way of implementing {{VertexInputFormat}} and {{VertexReader}}
> had bad smell. It required users to understand how these two classes are
> glued together, and forced similar codes to be duplicated in every new input
> format. (Similarly for the VertexOutputFormat and VertexWriter.) Anyone who
> wants to create a new format should create an underlying record reader or
> writer at the right moment and delegate some calls to it, which seemed
> unnecessary detail being exposed. Besides, type parameters had to appear all
> over every new format code, which was extremely annoying for both reading
> existing code and writing a new one. I was very frustrated writing my first
> format code especially when I compared it to writing a new vertex code. I
> thought writing a new input/output format should be as simple as vertex.
> So, I have refactored {{TextVertexInputFormat}} and {{OutputFormat}} into new
> forms that have no difference in their interfaces, but remove a lot of burden
> for subclassing. Instead of providing static VertexReader base classes, I
> made it a non-static inner-class of its format class, which helps eliminate
> the repeated code for gluing these two, already tightly coupled classes.
> This has additional advantage of eliminating all the Generics type variables
> on the VertexReader side, which makes overall code much more concise. I
> added several useful TextVertexReader base classes that can save efforts for
> implementing line-oriented formats.
> Please comment if you see my proposed change have any impact on other
> aspects. I'm unsure of how these additional layers of abstraction could
> affect performance.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
