Pulling this in from Slack:

Looking at all the RDD filters, they're all basically the same implementation; there's a field, do this custom filter on it. So, a DF and RDD re-implementation could be very similar. Basically what you proposed, the filter UDF taking in two parameters. So, we could do something like this for both RDD and DF:

.filter($"col".isInUrlPatterns(Set(".*index.*".r)))

...and, if we play our cards right, we could just have one implementation for both 🤷‍♂

we could just have one implementation for both

That would be great in the short term, but not necessary for the long term, IMO. Eventually, the DF functionality would be a superset of the RDD functionality, since we have no intention of backporting new DF features to RDD.

Thinking about this more, I'm not seeing the use of moving in this direction since it appears to be a slightly more complicated version of just using filter.

For example:

import io.archivesunleashed._

RecordLoader.loadArchives("/home/nruest/Projects/au/sample-data/geocities",sc)
  .webpages()
  .count()

res0: Long = 125579

import io.archivesunleashed._

val languages = Set("th","de","ht")

RecordLoader.loadArchives("/home/nruest/Projects/au/sample-data/geocities",sc)
  .webpages()
  .keepLanguagesDF(languages)
  .count()

res1: Long = 3536

import io.archivesunleashed._

val languages = Set("th","de","ht")

RecordLoader.loadArchives("/home/nruest/Projects/au/sample-data/geocities",sc)
  .webpages()
  .filter($"language".isinCollection(languages))
  .count()

res7: Long = 3536

With that, I'd argue we keep what we have now, or remove all the DataFrame filters as they exist now in DataFrames, and resolve this issue by updating the current documentation with the pure Spark DF implementation of filters.

...and if we go with the latter, that'll solve a sizable chunk of the Python implementation 😃

Can I propose yet another alternative design?

import io.archivesunleashed._

RecordLoader.loadArchives("/home/nruest/Projects/au/sample-data/geocities",sc)
  .webpages()
  .filter(hasLanguage("th","de","ht"))
  .count()

This saves the scholar from having to know about the schema explicitly? The UDF should be able to figure it out...

And similarly, we can have:

import io.archivesunleashed._

RecordLoader.loadArchives("/home/nruest/Projects/au/sample-data/geocities",sc)
  .webpages()
  .filter(urlMatches("""regex"""))
  .count()

I'm just thinking that a humanities scholar might get confused/scared about the $ notation and would be unfamiliar with the . method notation being applied with weird $ thingys?

Yeah, I like that better @lintool. Then we should be able get the negation with ! in Scala and ~ in Python/PySpark.

• keepImagesDF -> hasImages
• keepHttpStatusDF -> hasHTTPStatus
• keepDateDF -> hasDates
• keepUrlsDF -> hasUrls
• keepDomainsDF -> hasDomains
• keepMimeTypesTikaDF -> hasTikaMimeTypes
• keepMimeTypesDF -> hasMimeTypes
• keepContentDF -> contentMatches
• keepUrlPatternsDF -> urlMatches
• keepLanguagesDF -> hasLanguages

Feel free to suggest better names if you have them.