You are browsing a read-only backup copy of Wikitech. The live site can be found at wikitech.wikimedia.org
This private data lives in same server cluster that runs Wikimedia's production websites. This means you will need production shell access to get it (see also these notes on configuring SSH specifically for the purpose of working with the stats servers).
However, since this access gets you closer to both those production websites and this confidential data, it is not freely given out. First, you have to demonstrate a need for these resources. Second, you need to have a non-disclosure agreement with the Wikimedia Foundation. If you're a Foundation employee, this was included as part of your employment agreement. If you're a researcher, it's possible to be sponsored through a formal collaboration with the Wikimedia Foundation's Research team.
If you get this access, you must remember that this access is extremely sensitive. You have a duty to protect the privacy of our users. As Uncle Ben says, "with great power comes great responsibility." Always follow the rules outlined in the Acknowledgement of Server Access Responsibilities, which you have signed if you have access to this data.
In addition, keep in mind the following important principles:
- Be paranoid about personally identifiable information (PII). Familiarize yourself with the data you are working on, and determine if it contains any PII. It's better to double and triple check than to assume anything, but if you have any doubt ask the Analytics team (via IRC or email or Phabricator). Please see the data retention guidelines.
- Don't copy sensitive data (for example, data accessible only by the users in the analytics-privatedata-users) from its origin location to elsewhere (in HDFS or on any other host/support) unless strictly necessary. And most importantly, do it only if you know what you are doing. If you are in doubt, please reach out to the Analytics team first.
- Restrict access. If you do need to copy sensitive data somewhere, please make sure that you are the only one able to access the data. For example, if you copy Webrequest data from its location on HDFS to your /user/$your-username directory, make sure that the permissions are set to avoid everybody with access to HDFS to read the data. This is essential to avoid accidental leaks of PII/sensitive data or retention over our guidelines (https://meta.wikimedia.org/wiki/Data_retention_guidelines).
- Clean up copies of data. Please make sure that any data that you copied is deleted as soon as your work has been done.
If you ever have any questions or doubts, err on the side of caution and contact the Analytics team. We are very friendly and happy to help!
To get access, you submit a request on phabricator and tag
SRE-Access-Requests for shell access: Production shell access#Requesting access. You will need to specify what access group you need.
'analytics-*' groups have access to the Analytics Cluster (which mostly means Hadoop). 'statistics-*' groups get access to stat* servers for local (non distributed) compute resources. These groups overlap in what servers they grant ssh access to, but further posix permissions restrict access to things like MySQL, Hadoop, and files.
Here's a summary of groups you might need (as of 2016-10-18):
- Access to stat1006 and the credentials for the MariaDB slaves in
- Access to stat1006 for number crunching and accessing non private log files hosted there.
- Access to stat100 for the MariaDB slaves in
- WMDE specific group (mostly used for crons). Access to stat1007 and to MariaDB slaves in
- Access to stat1004 to connect to the Analytics/Cluster (Hadoop/Hive) (NO HADOOP PRIVATE DATA).
- Access to stat1004, stat1005 and stat1007 to connect to the Analytics/Cluster (Hadoop/Hive) and to query private data hosted there, including webrequest logs. Access to MariaDB slaves in
- If you want Hadoop access you probably want this. Please also request a Kerberos authentication principal following Analytics/Systems/Kerberos/UserGuide#Get_a_password_for_Kerberos if you hold an existing user in analytics-privatedata-users but not one for Kerberos, otherwise it is sufficient to request it in the task of your access request. To SRE: if you are adding a new user in analytics-privatedata-users, remember that from December 2019 Kerberos was enabled for Hadoop, so group membership is not enough to access the cluster. In order to create an Kerberos principal, please follow: Analytics/Systems/Kerberos#Create_a_principal_for_a_real_user.
Host access granted
|Access Groups||Access to stat1007||Access to stat1006||Access to stat1004||Access to stat1005||Access to notebook100|
Data access granted
|Access Groups||Hadoop access
(No private data)
The Analytics clients store data using redundant RAID configurations, but are not otherwise backed up. Your home directory on HDFS (
Once you have access to the production cluster, there are several servers which you can use to access the various private data sources and do general statistical computation. There are two types: the stat servers, designed for command-line use, and the SWAP servers, designed for Jupyter notebook use. Together, these are called the analytics clients, since they act as clients accessing data from various other databases.
|notebook1003||notebook1003.eqiad.wmnet||Hadoop, MariaDB, SWAP|
|notebook1004||notebook1004.eqiad.wmnet||Hadoop, MariaDB, SWAP|
You may need to access the internet from the analytics clients (for example, to download a Python script using
pip). By default, this will fail because the machines are tightly firewalled. You'll have to use the HTTP proxy.
If you wish to publish a dataset or report from one of the analytics clients, you can place it in the
/srv/published-datasets. You can find more information on Analytics/Web publication.
On stat1005 we have deployed an AMD GPU for T148843. The long term plan is to make it available for all the users logging in, but for the moment its access is restricted to the POSIX group
gpu-testers to better test it (and avoid usage contention etc..). Please reach out to the Analytics team if you wish to get added to the group to test the GPU for your use case.
The Analytics MariaDB cluster contains copies of the production MediaWiki databases (both actively-used mainstream projects and small internal-facing wikis, like various projects' Arbitration Committees).
- As of November 2019, Hadoop is authenticated via Kerberos. See Kerberos User guide and Hadoop testing cluster.
Hadoop is our storage system for large amounts of data. The easiest way to query the Hadoop data is through Hive, which can be accessed from stat1007 and stat1004 - simply type
beeline in the terminal, switch to the
wmf database, and input your query.
At the moment there are no recommended Hive access packages for R or Python. In the meantime, the best way to get data out of the system is to treat it as you would the Analytics slaves; through the terminal, type:
beeline -f my_query.hql > file_name.tsv
For information about writing HQL to query this data, see the Hive language manual.
Data sets and data streams can be found in Category:Data_stream
Data Dashboards. Superset and Turnilo (previously called Pivot)
You need a wikitech login that is in the "wmf" or "nda" LDAP groups. If you don't have it, please create a task like https://phabricator.wikimedia.org/T160662
Before requesting access, please make sure you:
- have a functioning Wikitech login. Get one: https://toolsadmin.wikimedia.org/register/
- are an employee or contractor with wmf OR have signed an NDA
Depending on the above, you can request to be added to the wmf group or the nda group. Please indicate the motivation on the task about why you need access and ping the analytics team if you don't hear any feedback soon from the Opsen on duty.
MediaWiki application data
You can do a lot of work with the data stored by MediaWiki in the normal course of running itself. This includes data about:
- Users' edit counts (consult the
- Edits to a particular page (consult the
revisiontable, joined with the
pagetable if necessary)
- Account creations (consult the
You can access this data using the replica MariaDB databases. These are accessible from the stat100* machines, as detailed below.
For an overview of how the data is laid out in those databases, consult the database layout manual.
There are a few things that aren't available from the databases replicas. The main example of this is the actual content of pages and revisions. Instead, you can access them through the API or in the XML dumps, which are both described below.
A subset of this application data, which doesn't present privacy concerns, is also publicly accessible through the API (except for private wikis, which you shouldn't really need to perform research on anyway!). A good way to understand it, and to test queries, is Special:ApiSandbox, which provides a way of easily constructing API calls and testing them. The output includes "Request URL" - a direct URL for making that query in the future, that should work on any and all Wikimedia production wikis.
If you're interested in common API tasks, and don't feel like reinventing the wheel, there are a number of Python-based API wrappers and MediaWiki utilities. Our very own Aaron Halfaker maintains MediaWiki Utilities, which includes a module dedicated to API interactions. There's no equivalent for R yet.
Every month, XML snapshots of the databases are generated. Since they're generated monthly, they're always slightly outdated, but make up for it by being incredibly cohesive (and incredibly large). They contain both the text of each revision of each page, and snapshots of the database tables. As such, they're a really good way of getting large amounts of diffs or information on revisions without running into the query limits on the API.
Aaron's MediaWiki-utilities package contains a set of functions for handling and parsing through the XML dumps, which should drastically simplify dealing with them. They're also stored internally, as well as through dumps.wikimedia.org, and can be found in
/mnt/data/xmldatadumps/public on stat1007.
One analytics-specific source of data is EventLogging. This allows us to track things we're interested in as researchers that MediaWiki doesn't normally log. Examples include:
- A log of changes to user preferences;
- A/B testing data;
- Clicktracking data.
These datasets are stored in the
event_sanitized Hive databases, subject to HDFS access control.
An important piece of community-facing data is information on our pageviews; what articles are being read, and how much? This is currently stored in our Hadoop cluster, which contains aggregated pageview data as well as the mostly-raw database of web requests. See the detailed documentation here.
When you have IP addresses - be they from the RequestLogs, EventLogging or MediaWiki itself - you can do geolocation. This can be a very useful way of understanding user behaviour and evaluating how our ecosystem works. We currently use the MaxMind geolocation services, which are accessible on both stat1006 and stat1007: a full guide to geolocation and some examples of how to do it can be found on the 'geolocation' page.
- Other groups including
statistics-web-usersare for people doing system maintenance and administration, so you don't need them just to access data.