Join GitHub today
GitHub is home to over 40 million developers working together to host and review code, manage projects, and build software together.
Sign upMigration of all RDD functionality over to DataFrames #223
Comments
This comment has been minimized.
This comment has been minimized.
|
@jrwiebe I have a note here from our call to "Go through the scala dir and identify all the functions that take in RDD and do not take in DF, then create tickets for JWb." Do you want me to do that granular of a level, or do you want to use this issue to take care of it? |
This comment has been minimized.
This comment has been minimized.
|
This is fine. |
This comment has been minimized.
This comment has been minimized.
|
@ianmilligan1 @lintool this look the basic inventory?
|
This comment has been minimized.
This comment has been minimized.
|
I think we should just leave this as a "catch-all" issue, open. IMO, this should be driven by the documentation update - go through docs, everything that we do with RDDs, we make sure there's a corresponding DF code example. When the docs have everything in both RDD and DF, I think we're done. |
This comment has been minimized.
This comment has been minimized.
|
@SinghGursimran this one is tied to #372, and should become a lot clearer as to what needs to be done to close this one. I think we're pretty close here. |
This comment has been minimized.
This comment has been minimized.
|
From @lintool
This spawned from a Slack convo Jimmy and I had in a non-public channel that never made it to the ticket. I have a branch working through this now. |
We need to migrate all current RDD functionality over to DataFrames. This means porting all matchbox UDFs over to DF UDFs.
There are two possible ways to do this - we can simply take matchbox UDFs and wrap them, or rewrite them from scratch. I suggest we revisit one by one, which will give us an opportunity to refine the UDF we actually want.
For example, the current RDD matchbox
ExtractDomainis implemented a bit differently than the DF version we've been playing with... it, for example, strips the prefixwww, whereas the RDD impl doesn't. I like the newer implementation better, but open to discussion.Also, this is an issue we'll come across sooner or later:
https://stackoverflow.com/questions/33664991/spark-udf-initialization
I have a general question we need to look into from the performance perspective: what's the lifecycle status of a Spark DF UDF? In particular, if there's initialization like compiling regexp, we don't want to do that over and over again... we want to have an init stage?
@TitusAn let's start developing in parallel the DF versions of the apps in #222 and try and work this out?