Again, without misclick sending (sorry for the spam).

Hi Justin,

First, one important thing: the data you are trying to get is VERY sensitive data in term of potential personal information (pageview + country at very granular time level).

I'd like to understand better how you'll keep it (how long, what protection, who will have access), and also how you plan to publish about it.

Now that administrative things are said, some technical stuff:

• The hive table is partitioned by year / month / day / hour
  For testing purposes, having those four values set makes the volume of data to scan much smaller. (~350G per month, therefore ~11G per day, 0.5G per hour)
• Whatever request you make, the columns needed for your request (for filtering, group by, or selection) will be scanned
  In that respect, going for only one query getting all the articles or countries at once data is much cheaper than repeated queries filtering per article or country
• Total Ordering costs a lot in that it implies all the data flowing through a single reducer - If you can do without, it'll be cheaper :)
• Finally your query has a bug: since there are other dimensions in the table than the one you are filtering by / selecting, you'll have many more rows than expected.
  You need to group by the selected columns and sum the view_count

Here is how I'd do it (with some time / data size considerations):

INSERT OVERWRITE DIRECTORY "/tmp/joal/test_pageviews_countries"

    SELECT

        year,

        month,

        day,

        hour,

        page_title,

        country_code,

        SUM(view_count) as views

    FROM

        wmf.pageview_hourly

    WHERE

        year = 2016

        AND month = 1

        AND agent_type = 'user'

    GROUP BY

        year,

        month,

        day,

        hour,

        page_title,

        country_code

    ;

I have run this query with more time restrictions for testing purposes (as discussed above).

• For one hour of data (day = 1, hour = 0):
• ouput data size ~ 220M
• actual CPU time 3mins 2secs
• waiting time 66 secs

• For one hour of data (day = 1, hour = 0) with sorting (ORDER BY views DESC LIMIT 1000000000):
• ouput data size ~ 220M
• actual CPU time 4mins 30secs (overhead 50% of original time, but since it's not parallelizable, would be much more for bigger data)
• waiting time 114 secs

• For one day of data (day = 1):
• ouput data size ~ 6.5G
• actual CPU time 1h 28 mins
• waiting time 132 secs

Another thing to consider: one month of the data generated with that query would be about 200G. It will take some time to copy over.

Joseph

On Sat, Apr 23, 2016 at 12:57 AM, Justin Clark <jclark@cyber.law.harvard.edu> wrote:

Hi all,

I'm a researcher at the Berkman Center for Internet & Society at Harvard doing some work on anomaly detection against Wikipedia article request volumes.

I'd like to create time series of request volumes for as many article-country pairs as is possible. I'm using a number of different data sets, but the most useful for our purposes is the pageview_hourly table. I understand that this is a tremendous amount of data, and we are in the process of prioritizing the article-country pairs, but my question is: what is the best/fastest way to query this data from hive? Writing a query that gets at the data is not a problem, but I'm curious about possible strategies that could speed up the process.

Here is a reference query that shows the kind of data I'm looking for:

SELECT view_count FROM pageview_hourly WHERE
year = 2015 AND
month = 1 AND
page_title = 'World_War_II' AND
country_code = 'CA' AND
agent_type = 'user'
ORDER BY day, hour;

A couple options that come to mind:
  * a year > 0 query vs many yearly, monthly, daily, or hourly queries
  * batching articles with page_title IN (...)
  * dropping country_code to get all countries at once (or batch like above)
  * ordering posthoc to avoid the map-reduce overhead

Because there's so much data and a query like the above takes ~10 minutes, experimenting with these is a long process. I was hoping someone more familiar could share any magic that might speed things up (or tell me there's no magic bullet and everything will take about as long as everything else). If no one can say quickly off the top of their head, I can just do that experimentation, but more options to try are totally welcome.

Thanks,
  Justin

_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics

--

Joseph Allemandou

Data Engineer @ Wikimedia Foundation

IRC: joal

--

Joseph Allemandou

Data Engineer @ Wikimedia Foundation

IRC: joal