Research talk:Measuring edit productivity/Work log/2014-12-5

From Meta, a Wikimedia project coordination wiki

Friday, December 5, 2014[edit]

It's been a little while. I have been working with hadoop and trying to build up some robust strategies for working with diff and persistence data. I made some substantial progress in working out how I can chop Wiki actions by different dimensions. See my blog on the subject.

A descriptive diagram of in-memory vs. streaming data processing is presented. Streaming processing is much more memory efficient, but it does come with limitations. Given the size of the datasets we're working with, I'll need to operate within these limitations.
Streaming vs. In-memory processing. A descriptive diagram of in-memory vs. streaming data processing is presented. Streaming processing is much more memory efficient, but it does come with limitations. Given the size of the datasets we're working with, I'll need to operate within these limitations.
A conceptual diagram of data processing dimensions of edits is presented. Note how the page and user dimensions are orthogonal -- yet they intersect.
Data processing dimensions. A conceptual diagram of data processing dimensions of edits is presented. Note how the page and user dimensions are orthogonal -- yet they intersect.

TL;DR: I can do unix style stream processing in Hadoop and I can use Hadoop's secondary sort to cut up data by dimensions. Woot.

Now I'm working on building up a library of streaming utilities that will let me slice and dice MediaWiki database dumps. See the repo.

Right now, the pipeline for getting to persistence stats looks like this: 

cat mw_xml_dump.xml | 
    dump2json | \
    json2tsv page.id timestamp id - | \
    sort -k1,1n -k2,2 -k3,3n | \
    cut -f4 | \
    json2diff | \
    <token_persistence> | \
    json2tsv id - | \
    sort -k1,1n | \
    <revision_stats>

I've put <token_stats> and <revision_stats> in brackets because those scripts have yet to be written at this point. Writing this out is helping me think about how they will function.

  • dump2json: Converts a mediawiki XML dump to a sequence of json blobs that represent a revision with this schema
  • json2tsv page.id timestamp id -: Extracts fields from a JSON into tsv format. Note that "-" prints the json blob as the 4th column
  • sort -k1,1n -k2,2 -k3,3n: Partitions by <page.id> and sorts revisions by <timestamp> ASC, <id> ASC
  • cut -f4: Trims all of the tsv columns but the JSON one to return as plain ol' JSON
  • json2diff: Generates diff information. This is the *very* CPU intensive part of the work
  • <token persistence>: Generates statistics about how tokens persist in a page. This is a *large* dataset. It will have a record for every token that was ever added to any page.
  • json2tsv id -: Extract the revision ID field from the token persistence JSON blob
  • sort -k1,1n: Partition by revision.
  • <revision stats>: Generates statistics about how a revision's tokens persisted overall. This will enable quality/productivity measures to be made.

My plan is to output a full dataset at every step to allow for debugging and reuse. We might not output the <token persistence> dataset because it will be too big. Then again we have a whole cluster to hold the data -- and it would be really cool to be able to re-process it. --Halfak (WMF) (talk) 20:33, 5 December 2014 (UTC)[reply]

Retrieved from "https://meta.wikimedia.org/w/index.php?title=Research_talk:Measuring_edit_productivity/Work_log/2014-12-5&oldid=10703124"