There are two webrequest tables available on Hive. The
wmf_raw.webrequest table is JSON data directly imported from Kafka.
wmf.webrequest contains the 'refined' webrequest data. The raw table is purged more frequently than 90 days.
For many purposes, it is preferable to query the much smaller pageview_hourly and projectview_hourly tables. These are derived from webrequest by filtering to keep only pageviews and aggregating over a predefined set of dimensions.
The stream originates from the Varnish caching servers, which emit log data in their cache log format. A varnishkafka instance (Puppet code) pipes the Varnish log data into Kafka, in the
Gobblin then loads the data from Kafka into HDFS, where it is stored as Snappy-compressed SequenceFiles in JSON format. A Hive table with hourly partitions is mapped on top of these data files. This is
An Airflow job (code) then checks for missing or duplicated data and creates a "refined" copy which has additional fields and is stored as a Parquet file. Parquet is a column-based data storage format, which is much more efficient to query than the "raw" JSON data (see Data Format Experiments#Results for more info). This refined copy is
When the webrequest job fails, it is important to find the problem and rerun it as quickly as possible, because it is upstream from many other pipelines and it uses more resources than other jobs. Following are common problems and rerun procedures:
It's when a small number of rows of a webrequest hour have an incorrect format and break the refinement process (for instance the
x_analytics header, which is parsed into a map). You need to build requests to the data manually to find the rows which break the process. The identifying fields for a single row over an hour are the pair
(hostname, sequence). If the number of rows is small (less than 100), you can then rerun the webrequest failed hour in Airflow with a specific airflow variable set: in the
refine_webrequest_hourly_text_config variable (for the text partition), add a json value as follow:
"excluded_row_ids":"'hostname1', sequence1, 'hostname2', sequence2..." . This will remove the faulty rows from the refined process. Important: Don't forget to remove the
excluded_row_ids part of the variable once the run has succeeded, to not remove rows from later hours.
If you want to query the webrequest stream, you almost always want to use the "refined" version available in
wmf.webrequest, which been enriched with extra fields and stored in a format that is much faster to query. (NOTE: descriptions below are a little brief on purpose, for precise definitions you can look at the refine step, which is a simple Hive query invoking UDFs, and follow that through by searching refinery-source)
|hostname||string||Source node hostname|
|sequence||bigint||Per-host sequence number|
|dt||string||Timestamp at cache in ISO 8601 format|
|time_firstbyte||double||Time to first byte in seconds (a sample value is .000243). The time it took from when the req was first received by varnish until it started sending response bytes back to the client.|
|ip||string||IP of packet at cache|
|http_status||string||HTTP status of response|
|http_method||string||HTTP method of request|
|uri_host||string||Host of request|
|uri_path||string||Path of request|
|uri_query||string||Query of request – including the leading "?" if set|
|content_type||string||Content-Type header of response|
|referer||string||Referer header of the request. "Unknown" value indicates that url in referrer was not parseable or defective, "-" indicates no referrer was received.|
|x_forwarded_for||string||X-Forwarded-For header of request|
|user_agent||string||User-Agent header of request|
|accept_language||string||Accept-Language header of request|
|x_analytics||string||X-Analytics header of response, looks like: |
|range||string||Range header of response|
|is_pageview||boolean||Indicates if this record was marked as a pageview during refinement|
|record_version||string||Keeps track of changes in the table content definition|
|client_ip||string||Best guess client IP address from the X-Client-IP header, set by varnish using X-Forwarded-For and other information.|
|geocoded_data||map<string,string>||Geocoded data computed during refinement using client_ip and MaxMind database contains: continent, country_code, country, subdivision, city but has nulls where information is not available|
|x_cache||string||X-Cache header of response|
|user_agent_map||map<string,string>||User-agent map with browser_family, browser_major, device_family, os_family, os_major, os_minor and wmf_app_version keys and associated values|
|x_analytics_map||map<string,string>||The X-Analytics header broken up into key-value pairs|
|ts||timestamp||Unix timestamp in milliseconds extracted from dt|
|access_method||string||Client type used to access the site (|
|agent_type||string||Categorise the agent making the webrequest as either user or spider (automatas to be added).|
|is_zero||boolean||Indicates if the webrequest is accessed through a zero provider|
|referer_class||string||The type of site found in the referer header: |
|struct containing project_class/project_familly (such as wikipedia or wikidata for instance - project_class is to be deprecated, use project_family instead), project (such as en or commons), qualifiers (a list of in-between values, such as m and/or zero) and tld (org most often)|
|pageview_info||map<string,string>||Map containing |
|page_id||int||MediaWiki page_id for the page served by this webrequest. This is not set for pages requested through the MediaWiki API. For MediaWiki redirects, this is set to the redirected-to page_id, i.e. the page_id of the content shown to the user, not the page_id of the originally requested title.|
|namespace_id||int||MediaWiki namespace_id for the page served by this webrequest. This is set with the same logic as page_id.|
|tags||array<string>||List containing tags qualifying the request, ex: [portal, wikidata]. Will be used to split webrequest into smaller subsets.|
|isp_data||map<string, string>||Internet Service Provider data in a map with keys |
|accept||string||Accept header of request|
|tls||string||TLS information of request|
|tls_map||string||Map view of TLS information (keys are vers, keyx, auth and ciph)|
|ch_ua||string||Value of the Sec-CH-UA request header|
|ch_ua_mobile||string||Value of the Sec-CH-UA-Mobile request header|
|ch_ua_platform||string||Value of the Sec-CH-UA-Platform request header|
|webrequest_source||string||Which Varnish cluster handled the request. Current values are |
|year||int||Unpadded year of request (partition field)|
|month||int||Unpadded month of request (partition field)|
|day||int||Unpadded day of request (partition field)|
|hour||int||Unpadded hour of request (partition field)|
|referer_data||struct<referer_class:string,referer_name:string>||Struct containing referer_class (indicates if a referer is internal, external, external(media sites), external(search engine) or unknown.) and referer name (name of referer when referer class is external(search engine) or external(media sites))|
hour fields. These are Hive partitions, and are explicit mappings to hourly imports in HDFS. You must include at least one partition predicate in the
where clause of your queries (even if it is just
year > 0). Partitions allow you to reduce the amount of data that Hive must parse and process before it returns you results. For example, if are only interested in data during a particular day, you could add
where year = 2014 and month = 1 and day = 12. This will instruct Hive to only process data for partitions that match that partition predicate. You may use partition fields as you would any normal field, even though the field values are not actually stored in the data files.
The spider/user classification recorded in the
agent_type field is based on the user agent, evaluated against some regexes (the standard spider detection provided by ua-parser augmented by a custom bot regex, see code for detail, or the Automated traffic page).
Queries over the full table will sometimes fail with a
java.io.FileNotFoundException: File does not exist error during execution, in case the purging of the oldest data (after 90 days) happens before the query accesses those partitions. This can be avoided by manually excluding the oldest available dates in the WHERE clause.
SELECT COUNT(DISTINCT ip) AS hits FROM webrequest WHERE year = 2014 AND month = 5 AND day = 5 AND hour = 13 AND uri_host = "en.wikipedia.org" AND http_status = 200;
Top raw referrals for an article
SELECT referer, COUNT(DISTINCT ip) AS hits FROM webrequest WHERE year = 2014 AND month = 5 AND day = 5 AND hour = 13 AND uri_path = "/wiki/London" AND uri_host = "en.wikipedia.org" AND http_status = 200 GROUP BY referer ORDER BY hits DESC LIMIT 50;
Top internal referrals for an article
SELECT SUBSTR(referer,30) AS source, COUNT(DISTINCT ip) AS hits FROM webrequest WHERE year = 2014 AND month = 5 AND day = 5 AND hour = 13 AND uri_path = "/wiki/London" AND uri_host = "en.wikipedia.org" AND referer LIKE "http://en.wikipedia.org/wiki/%" AND http_status = 200 GROUP BY SUBSTR(referer,30) ORDER BY hits DESC LIMIT 50;
Top outbound clicks from an article
SELECT SUBSTR(uri_path,7) AS target, COUNT(DISTINCT ip) AS hits FROM webrequest WHERE year = 2014 AND month = 5 AND day = 5 AND hour = 13 AND uri_host = "en.wikipedia.org" AND referer LIKE "http://en.wikipedia.org/wiki/London%" AND http_status = 200 GROUP BY SUBSTR(uri_path,7) ORDER BY hits DESC LIMIT 50;
Changes and known problems since 2015-03-04
- See also m:Research:Page view#Change log for changes to the page view definition itself, as well as Analytics/Data_Lake/Traffic/Pageview_hourly#Changes and known problems since 2015-06-16
|Date from||Date until||Task||record_version||Details|
|2015-01-01||*||Create table with data from 'text', 'mobile' and 'upload' request sources.|
|2015-01-26||*||Add 'misc' request source.|
|2015-02-12||*||Add 'bits' request source.|
|2015-02-27||task T90725, task T89401||0.0.1||Add record_version, client_ip and geocoded_data fields to the schema. An error is raised when trying to access those fields for data before the release date.|
|2015-03-30||task T91749, task T91793, task T89396||0.0.2||Add x_cache, user_agent_map and x_analytics_map fields to the schema. An error is raised when trying to access those fields for data before the release date. Note: user_agent_map and x_analytics_map are to replace original user_agent and x_analytics fields, but field replacement breaks schema backward compatibility. We will deploy this chamge when we'll be confident enough that no more backward incompatible change are needed soon.|
|2015-04-10||task T95178, task T94584||0.0.3||Add ts, access_method, agent_type and is_zero fields to the schema. An error is raised when trying to access those fields for data before the release date. Note: The agent_type on.y contains spider and user values for the moment. A detailed analysis on user_agent and usage patterns should be overtaken to provide new/refined values. |
[agent_type field is bugged in this version and is fixed in 0.0.4.]
|2015-05-07||0.0.4||Add referer_class field to the schema, with possible values internal, external or unknown. An error is raised when trying to access those fields for data before the release date. Agent_type bug is corrected, and some improvements are made in the computation of is_pageview and access_method fields.|
|2015-06-11||task T99932, task T99918, task T96044||0.0.5||Add normalized_host and pageview_info fields to the schema, and update user_agent_map adding wmf_app_version. normalized_host is created splitting the uri host as a struct with project_class (wikipedia, wikidata for instance), project (en, commons for instance), qualifiers (an array of in-between qualifiers, like m or zero for instance), and tld (org most of the time). pageview_info is set only when is_pageview is true, and is a map with keys project (en.wikipedia, commons.wikimedia for instance), language_variant (zh-hant for instance), and page_title when possible.|
|2015-08-13||task T107764, task T108833, task T108717||0.0.6||Correct bug in the use of webrequest uri hosts to compute if pageview or not. The change affect 0.2% of pageviews by removing false positives. The 'bits' webrequest_source partition and associated automatic jobs have been removed (every data that was in bits is now text partition), and a new 'maps' webrequest_source partition has been created (data for this partition starts on 2015-08-11T00:00:00).|
|2015-08-03, 2015-08-10||task T106581||Two separate Kafka upgrade attempts caused significant data loss. A bug in Kafka caused data not to be compressed properly. Once the bug was fixed, we spent a lot of time patching and repairing what data we could. Several hourly webrequest partitions had significant loss and could not be restored. This loss was either caused by Kafka being too overloaded to accept produce requests, or by a recovery attempt in which a single broker for whom no replicas were in sync was restarted. http://debugging.wmflabs.org/#refinery-graphs-tab. See also: Incident_documentation/20150803-Kafka,Incident_documentation/20150810-Kafka|
|2015-08-27||task T109256, task T108866||0.0.7||Correct two bugs:
|2015-08-31||0.0.8||arbcom-*.wikimedia.org are not pageviews anymore.|
|2015-09-16||task T109383, task T106134, task T108598||0.0.9||Update ua-parser to an up-to-date version (better user-agent classification), improve spider tagging by enhancing the regular expression used in addition of ua-parser, update pageview definition to prevent requests having 'preview' set in x_analytics header to be counted as pageviews or appPageviews. |
|2015-11-19||task T119054||0.0.10||Improve spider tagging adding user agent being |
|2015-12-01||task T116023||0.0.11||Add the mediawiki page_id for this hit when available. For redirects this could be the page_id of the redirect or the page_id of the target. This may not always be set, even if the page is actually a pageview. Upgrade pageview_definition to better filter non-interesting path.|
|2015-12-15||There was an irrecoverable raw_webrequest data loss for the hour 17 UTC of: 25.6% (misc), 19.5% (mobile), 19.1% (text), 39.1% (upload). the loss was due to the enabling of IPSec. This encrypts varniskafka traffic between caches in remote datacenters and the Kafka brokers in eqiad. During a period of about 40ish minutes, no webrequest logs from remote datacenters were successfully produced to Kafka.|
|2015-12-18||There was another irrecoverable raw_webrequest data loss between 21:54 and 22:15. A puppet change was merged in which an important firewall rule dealing with IPSec was lost. This kept all varnishkafkas in remote datacenters from producing to Kafka during this time.|
|2016-02-03||task T115919||0.0.12||Referer classification has an updated version providing external search engine details (when any).|
|2016-03-09||task T125731||0.0.13||Improve spider tagging regular expression to match the bots that follow the conventions specified in https://meta.wikimedia.org/wiki/User-Agent_policy. Remove tagging user agents containing "WikimediaBot" as bot.|
|2016-03-23||task T130399||0.0.14||Add x_analytics header inclusive filtering (consider pageview if header contains <code>pageview=1</code>), update ua-parser to an up-to-date version, correct a bug in search engine referer classification.|
|2016-04-11||task T131824, task T128612||0.0.15||Correct x_analytics header inclusive filtering (only apply |
|2016-05-16||task T135168||0.0.16||Correct pageview definition to flag mobile app pageviews when x_analytics header has |
|2016-08-12||task T142410||0.0.16||Correct cache_status header|
|2017-02-09||task T156628, task T155141, task T157528||0.0.17||Update pageview definition to remove previews (POST with "action=submit" in query is now excluded from pageviews). Add DSXS (self-identified bot) to bot regex. Add namespace_id field.|
|2017-03-21||task T142410||0.0.18||Update pageview definition to stop filtering out test.wikipedia.org from pageviews.|
|2017-06-13||Task T118557||0.0.18||Removed |
|2017-08-23||Task T168874, Task T171760||0.0.19||Add |
|2017-11-03||Task T177199||Changed |
|2018-02-15||task T167907||0.0.20||Add ISP data from MaxMind GeoIP2 ISP database to webrequest.|
|2018-04-10||Task T191940||For about 54min (from 22:46 - 23:40) on April 10, a significant portion of global traffic was not reaching our sites (incident documentation). The validity of the webrequest data itself appears to not have been affected.|
|2018-04||task T187014||From February 6 to April 16, 2018, the geolocation data for traffic from Opera browsers on mobile web was incorrect (wrongly labeled them as coming from the US, "Unknown" and some other countries rather than their true origin).|
|2018-05-02||task T191714||Add Ecosia and Startpage to list of search engines.|
|2018-05-21||Task T193578||Updated UA parsing source regexes. Mainly Windows NT (Windows 2000 to the present) requests now appear in the same OS family, as "Windows". Improves identification of page previews from browsers and devices, and correctly identifies devices previously marked as "Generic Android". Many requests that were marked wrongly as IE7 before are now (mostly) correctly identified as IE11.|
|2018-06-21||task T192779||Wikipedia Android app starts sending referer data (production release: 2018-06-27)|
|2018-07-26||task T188776||Update pageview definition to include foundation.wikimedia.org host|
|2018 -09-09||onwards||task T211077||Referrer class change for Chrome Mobile|
|2018-10-09||Task T170606||Add accept header|
|2018-11-05||Task T208752||Issues with data coming from upload that might result in 'fake' traffic for some webp files|
|2019-01-07||Task T212862, Task T153821||Updated IPs for Labs. Wikitech added to PageviewDefinition.|
|2019-04-01||Task: T219842||Kafka jumbo failover produces some data loss for hour 22 UTC.|
|2019-06-04||task T224451||From 2019-04-25 until 2019-06-04, pages with "+" in the title were omitted from pageviews.|
|2019-06-05||task T224187||From now on, user agent strings longer than 400 chars are not parsed.|
|2019-06-28||task T225792||doc.wikimedia.org host is not flagged as pageview anymore|
|2019-09-18||task T212854||0.0.22||Update user-agent parser to up-to-date version. Main differences are |
|2019-11-07||task T233661||Added tls and tls_map fields|
|2020-01-13||task T239625||Added several common search engines to the list of hard-coded search engine referers. Added a predictive referer classifier to classify non-hard-coded search engines. Introduced support for Android app referers.|
|2020-07-01||task T256514||0.0.23||Update pageview definition for new android mobile content consumption.|
|2020-10-20||task T236740||0.0.24||Remove latitude, longitude, and postal code from |
|2020-11-16||task T267008||Grow the number of buckets by which the table is bucketed from 64 to 256. Also apply a change supposedly applied in February 2017: make |
|2022-01-26||task T299397||Add low entropy Sec-CH-UA request headers: |
|2022-01-26||task T300164||Data loss has been incurred from June 4th 2021 to January 26th 2022 due to a varnishkafka install problem on some of the caching hosts.
Summary of the data loss analysis:
|2022-09-06||task T315329||0.0.25||Add Brave, Petal, VK to list of search engines and expand predicted other to include suche domains (German).|
|2023-01-09||task T326658||Data loss has been incurred on January 8th 2023 between the 8th to14th hours due to issues encountered by Eqsin datacenter on 2023/01/08.
Below is an analysis of the data loss:
At 2023-1-8-8 hour we lost an average of 8119100 records (2.8391253%) of valid webrequest text data.
At 2023-1-8-9 hour, we lost an average of 8313842 records (2.663%) of valid webrequest text data.
At 2023-1-8-10 hour, we lost an average of 8336440 requests (2.55%) of valid webrequest text data.
At 2023-1-8-11 hour, we lost an average of 10172609 requests (2.916%) of valid webrequest text data.
At 2023-1-8-12 hour, we lost an average of 11262617 requests (3.09%) of valid webrequest text data.
2023-1-8-13 hour, we lost an average of 11974996 requests (3.081%) of valid webrequest text data.
2023-1-8-14 hour, we lost an average of 10958911 requests (2.717%) of valid webrequest text data.
In summary, on 2023/01/8 between the 8th and 14th hours, we lost an average of 2.83% of webrequest-text data (from eqsin datacenter only).
|2023-02-16||task T327074||0.0.26||Add referer_data field to the schema|
|TBD (20230306)||task T327073||0.0.27||Generation of the partition from Airflow+Spark (in place of Oozie+Hive)|
In almost all cases, you should not query this "unrefined" data, because the "refined" data in
wmf.webrequest is richer and much faster to query.
If you do need query this raw data in Hive, you will first need to add the json "serde" (serializer-deserializer) to your path:
ADD JAR /usr/lib/hive-hcatalog/share/hcatalog/hive-hcatalog-core.jar
You can do this in Spark by invoking the Spark command line tool as follows:
/usr/bin/spark2-shell --jars /usr/lib/hive-hcatalog/share/hcatalog/hive-hcatalog-core.jar.
In addition to the core data, two tables are created with detailed statistics about the hourly partition loaded.
wmf_raw.webrequest_sequence_stats: contains details for each varnish-kafka source of duplicates / missing data.
wmf_raw.webrequest_sequence_stats_hourly: contains hourly aggregated percent_missing and percent_loss for each webrequest source.
The statistics in those two tables are computed using a 'sequence' number, a counter generated by each varnishkafka process incrementing its value for each line it outputs. This sequence number allows to track for duplicates and missing data using some arithmetics (max, min, distinct). For details, see the code for webrequest_sequence_stats and webrequest_sequence_stats_hourly.
Querying the statistics
A typical query of these statistics is:
SELECT year, month, day, hour, webrequest_source, percent_lost, percent_duplicate FROM wmf_raw.webrequest_sequence_stats_hourly WHERE year = XXXX AND month = XX AND day = XX AND hour = XX AND webrequest_source IN (XXX) ORDER BY year, month, day, hour, webrequest_source LIMIT 1000;
For an example of querying to diagnose a data loss issue, see T211000.
If you want to look at API use, it makes more sense to use the Mediawiki API request data stream which is logged directly from MediaWiki when it responds to API requests. The benefits of that data include:
- Includes internal API requests not routed through the Varnish servers
- Data on API requests is not mixed in with data on regular web requests.
- Contains detailed data on the content of POST API requests (they are logged in webrequests, but without the request bodies).
RAW IP usage for IP Data
The following data streams (available as Data Lake tables) are derived from webrequest:
- Pageview hourly
- Pageview actor
- Projectview hourly
- Unique devices
- Browser general
- Mediawiki API request
- Mobile apps session metrics
- Mobile apps uniques
- Interlanguage navigation
See also Analytics/Data Lake/Traffic#Datasets.
- Research into referrer counts: task T195880
- It doesn't include some internal (within-cluster) API requests which are not routed through the Varnish servers.