For all Hive users using stat1002/1004, you might have seen a deprecation
warning when you launch the hive client - that claims it's being replaced
with Beeline. The Beeline shell has always been available to use, but it
required supplying a database connection string every time, which was
pretty annoying. We now have a wrapper
setup to make this easier. The old Hive CLI will continue to exist, but we
encourage moving over to Beeline. You can use it by logging into the
stat1002/1004 boxes as usual, and launching `beeline`.
There is some documentation on this here:
If you run into any issues using this interface, please ping us on the
Analytics list or #wikimedia-analytics or file a bug on Phabricator
(If you are wondering stat1004 whaaat - there should be an announcement
coming up about it soon!)
The Analytics team would like to announce that we have migrated the
reportcard to a new domain:
The migrated reportcard includes both legacy and current pageview data,
daily unique devices and new editors data. Pageview and devices data is
updated daily but editor data is still updated ad-hoc.
The team is working at this time on revamping the way we compute edit data
and we hope to be able to provide monthly updates for the main edit metrics
this quarter. Some of those will be visible in the reportcard but the new
wikistats will have more detailed reports.
You can follow the new wikistats project here:
Wikimedia is releasing a new service today: EventStreams
<https://wikitech.wikimedia.org/wiki/EventStreams>. This service allows us
to publish arbitrary streams of JSON event data to the public. Initially,
the only stream available will be good ol’ RecentChanges
<https://www.mediawiki.org/wiki/Manual:RCFeed>. This event stream overlaps
functionality already provided by irc.wikimedia.org and RCStream
<https://wikitech.wikimedia.org/wiki/RCStream>. However, this new service
has advantages over these (now deprecated) services.
We can expose more than just RecentChanges.
Events are delivered over streaming HTTP (chunked transfer) instead of
IRC or socket.io. This requires less client side code and fewer special
routing cases on the server side.
Streams can be resumed from the past. By using EventSource, a
disconnected client will automatically resume the stream from where it left
off, as long as it resumes within one week. In the future, we would like
to allow users to specify historical timestamps from which they would like
to begin consuming, if this proves safe and tractable.
I did say deprecated! Okay okay, we may never be able to fully deprecate
irc.wikimedia.org. It’s used by too many (probably sentient by now) bots
out there. We do plan to obsolete RCStream, and to turn it off in a
reasonable amount of time. The deadline iiiiiis July 7th, 2017. All
services that rely on RCStream should migrate to the HTTP based
EventStreams service by this date. We are committed to assisting you in
this transition, so let us know how we can help.
Unfortunately, unlike RCStream, EventStreams does not have server side
event filtering (e.g. by wiki) quite yet. How and if this should be done
is still under discussion <https://phabricator.wikimedia.org/T152731>.
The RecentChanges data you are used to remains the same, and is available
at https://stream.wikimedia.org/v2/stream/recentchange. However, we may
have something different for you, if you find it useful. We have been
internally producing new Mediawiki specific events
for a while now, and could expose these via EventStreams as well.
Take a look at these events, and tell us what you think. Would you find
them useful? How would you like to subscribe to them? Individually as
separate streams, or would you like to be able to compose multiple event
types into a single stream via an API? These things are all possible.
I asked for a lot of feedback in the above paragraphs. Let’s try and
centralize this discussion over on the mediawiki.org EventStreams talk page
<https://www.mediawiki.org/wiki/Talk:EventStreams>. In summary, the
What RCStream clients do you maintain, and how can we help you migrate
to EventStreams? <https://www.mediawiki.org/wiki/Topic:Tkjkee2j684hkwc9>
Is server side filtering, by wiki or arbitrary event field, useful to
Would you like to consume streams other than RecentChanges?
available events are described here
- Andrew Otto
I recently started working at Wikimedia Germany. My focus is new editor
retention work in the german community.
In that role my team created a tool to count views of videos the potential
new editor viewed on a wiki-page. If you're interested you can find it
here: https://tools.wmflabs.org/commons-video-clicks/ .
I have 2 questions to the data used in that tool and hope you can halp me
First, the tool is using the following query to get the data in JSON from
WMF-Database and showing it in the tool: https://tools.wmflabs.org/
A colleague told me that in the consecutive dates in the JSON output there
is stated that the data is updated daily. The problem is, that form June on
there is no data for the views. E. g. the following query gives no data:
knowledge.webm/20170601/20170603.Do you have more information on the update
rates or missing update of the data itself?
Second, do you know in whitch time zone the date in the database is?
I would love to have a feedback if this message reached you and if or when
you could help me with that. Maybe you know someone else who can support me
Big Thanks in advance and best regards,
Projektassistenz Neue Freiwillige
Wikimedia Deutschland e. V. | Tempelhofer Ufer 23-24 | 10963 Berlin
Tel. (030) 219 158 26-0
Stellen Sie sich eine Welt vor, in der jeder Mensch an der Menge allen
Wissens frei teilhaben kann. Helfen Sie uns dabei!
Wikimedia Deutschland - Gesellschaft zur Förderung Freien Wissens e. V.
Eingetragen im Vereinsregister des Amtsgerichts Berlin-Charlottenburg unter
der Nummer 23855 B. Als gemeinnützig anerkannt durch das Finanzamt für
Körperschaften I Berlin, Steuernummer 27/029/42207.
I just ran an experiment that surprised me and I thought folks on this list
would find interesting.
*tl;dr* We found that navigation vector embeddings for articles (as
produced by Ellery Wulcyzn
content-based vector embeddings (word2vec on article text) by 62% vs 37%
accuracy in a task-based user study. I've volunteered to help with the
engineering to productionize navigation embedding
<https://phabricator.wikimedia.org/T158972> and this study reinforces my
eagerness to get navigation vectors out in the world!
*More detail: *The maps we use in Cartograph (cartograph.info) are almost
entirely built on "embedding" vectors for articles. We experimented with
two word2vec-based embeddings: *content vectors* mined from article text
and link structure, and* navigation vectors* mined from user browsing
sessions. For the latter, we used Ellery Wulczyn's navigation vectors
staring at maps, our intuition told us that the navigation vectors seemed
better in "preference spaces" where the human taste space wasn't
necessarily easily encoded into Wikipedia text.
Last weekend we ran a Mechanical Turk experiment to test this intuition. We
created two Cartograph maps of movies: one built on navigation vectors and
one built on content vectors. We identified 40 relatively popular movies
that were not close neighbors in either map (i.e. cities that were not too
close to each other) and ran a Mechanical Turk study using the maps.
For each Turker, we randomly selected 5 seen movies (out of the 30), and
asked them to evaluate maps for each movie. For each movie city, we showed
the map region around the city, but hid the city and asked them to guess
the city from a list of 12 movies they had seen (screenshot below). We
added in trivial validation questions using sequels to ensure Turkers were
working in good faith (show a map for "Rocky II" that had "Rocky" at the
Result: Turkers exhibited 62% accuracy with the navigation vectors and 37%
accuracy with content vectors. We want to conduct several follow-up studies
to understand different subject areas and parameter settings and user
tasks, but the difference in performance was striking.
Our study shows the value of navigation vectors and makes me super excited
to contribute to the engineering needed to get them out to the world on a
regular basis. Imagine if every researcher and practitioner who uses
word2vec now on Wikipedia content switches to navigation vectors. That's a
Feedback and questions welcome!
Shilad W. Sen
Mathematics, Statistics, and Computer Science Dept.
Senior Research Fellow, Target Corporation
the Analytics team is working on some alter tables to the Eventlogging
'log' database on analytics-store (dbstore1002) and analytics-slave
(db1047) as part of https://phabricator.wikimedia.org/T167162.
The list of alter tables are the following:
This should be a transparent change but I thought it would have been better
to keep all of you informed in case of unintended regressions or
side-effects. The context of the alter tables is in T167162 but the TL;DR
is that we need nullable attributes across all the EL tables (except fields
like id, uuid and timestamp) to be able to sanitize data with our new
eventlogging_cleaner script (https://phabricator.wikimedia.org/T156933).
Please let me know if you encounter any issue with this change.
Thanks in advance!
*Who: *This mostly applies to people who have access to the stat1002 and
stat1003 statistics machines on the production cluster, and publish
datasets as static files.
*What:* We are no longer using datasets.wikimedia.org to serve static
datasets. We have set up a redirect, so requests like
https://datasets.wikimedia.org/ $1 will be sent to
https://analytics.wikimedia.org/datasets/archive/ $1. Most importantly,
publishing datasets is now much easier. Any files you put in
published-datasets on either machine:
Are going to be merged together and served together on:
One request as we all enjoy this much simpler process: let's use README
files in these directories to let future versions of us know what the
datasets are all about. That will make the repository more fun for others
to browse and ease future cleanups. Thank you!
If something of yours got lost, let us know, we have backups. If you had
stuff that we might have cleaned up, we put it in
/a/otto-to-delete-datasets-cleanup. Take a look there and you can move
files as you see fit into published-datasets
For a long time, publishing files from stat1002 and stat1003 was quite
painful. There were three folders, some on both boxes, some only on one
box, symlinks, rsyncs, it was bad. We talked to everyone who had files in
these folders and gathered consensus for this deprecation. If this message
catches you by surprise, please let us know what channel we should reach
you in next time and we'll add it to our communication plan.
This work is tracked in T159409 <https://phabricator.wikimedia.org/T159409>
The next Research Showcase will be live-streamed this Wednesday, June 21,
2017 at 11:30 AM (PST) 18:30 UTC.
YouTube stream: https://www.youtube.com/watch?v=i2jpKRwPT-Q
As usual, you can join the conversation on IRC at #wikimedia-research. And,
you can watch our past research showcases here
This month's presentations:
Title: Problematizing and Addressing the Article-as-Concept Assumption in
By *Allen Yilun Lin*
Abstract: Wikipedia-based studies and systems frequently assume that each
article describes a separate concept. However, in this paper, we show that
this article-as-concept assumption is problematic due to editors’ tendency
to split articles into parent articles and sub-articles when articles get
too long for readers (e.g. “United States” and “American literature” in the
English Wikipedia). In this paper, we present evidence that this issue can
have significant impacts on Wikipedia-based studies and systems and
introduce the subarticle matching problem. The goal of the sub-article
matching problem is to automatically connect sub-articles to parent
articles to help Wikipedia-based studies and systems retrieve complete
information about a concept. We then describe the first system to address
the sub-article matching problem. We show that, using a diverse feature set
and standard machine learning techniques, our system can achieve good
performance on most of our ground truth datasets, significantly
outperforming baseline approaches.
Title: Understanding Wikidata Queries
By *Markus Kroetzsch*
Abstract: Wikimedia provides a public service that lets anyone answer
complex questions over the sum of all knowledge stored in Wikidata. These
questions are expressed in the query language SPARQL and range from the
most simple fact retrievals ("What is the birthday of Douglas Adams?") to
complex analytical queries ("Average lifespan of people by occupation").
The talk presents ongoing efforts to analyse the server logs of the
millions of queries that are answered each month. It is an important but
difficult challenge to draw meaningful conclusions from this dataset. One
might hope to learn relevant information about the usage of the service and
Wikidata in general, but at the same time one has to be careful not to be
misled by the data. Indeed, the dataset turned out to be highly
heterogeneous and unpredictable, with strongly varying usage patterns that
make it difficult to draw conclusions about "normal" usage. The talk will
give a status report, present preliminary results, and discuss possible
Sarah R. Rodlund
Senior Project Coordinator-Product & Technology, Wikimedia Foundation