Analytics

Zulip has a cool analytics system for tracking various useful statistics that currently power the /stats page, and over time will power other features, like showing usage statistics for the various streams. It is designed around the following goals:

  • Minimal impact on scalability and service complexity.
  • Well-tested so that we can count on the results being correct.
  • Efficient to query so that we can display data in-app (e.g. on the streams page) with minimum impact on the overall performance of those pages.
  • Storage size smaller than the size of the main Message/UserMessage database tables, so that we can store the data in the main postgres database rather than using a specialized database platform.

There are a few important things you need to understand in order to effectively modify the system.

Analytics backend overview

There are three main components:

  • models: The UserCount, StreamCount, RealmCount, and InstallationCount tables (analytics/models.py) collect and store time series data.
  • stat definitions: The CountStat objects in the COUNT_STATS dictionary (analytics/lib/counts.py) define the set of stats Zulip collects.
  • accounting: The FillState table (analytics/models.py) keeps track of what has been collected for which CountStats.

The next several sections will dive into the details of these components.

The *Count database tables

The Zulip analytics system is built around collecting time series data in a set of database tables. Each of these tables has the following fields:

  • property: A human readable string uniquely identifying a CountStat object. Example: “active_users:is_bot:hour” or “messages_sent:client:day”.
  • subgroup: Almost all CountStats are further sliced by subgroup. For “active_users:is_bot:day”, this column will be False for measurements of humans, and True for measurements of bots. For “messages_sent:client:day”, this column is the client_id of the client under consideration.
  • end_time: A datetime indicating the end of a time interval. It will be on an hour (or UTC day) boundary for stats collected at hourly (or daily) frequency. The time interval is determined by the CountStat.
  • various “id” fields: Foreign keys into Realm, UserProfile, Stream, or nothing. E.g. the RealmCount table has a foreign key into Realm.
  • value: The integer counts. For “active_users:is_bot:hour” in the RealmCount table, this is the number of active humans or bots (depending on subgroup) in a particular realm at a particular end_time. For “messages_sent:client:day” in the UserCount table, this is the number of messages sent by a particular user, from a particular client, on the day ending at end_time.
  • anomaly: Currently unused, but a key into the Anomaly table allowing someone to indicate a data irregularity.

There are four tables: UserCount, StreamCount, RealmCount, and InstallationCount. Every CountStat is initially collected into UserCount, StreamCount, or RealmCount. Every stat in UserCount and StreamCount is aggregated into RealmCount, and then all stats are aggregated from RealmCount into InstallationCount. So for example, “messages_sent:client:day” has rows in UserCount corresponding to (user, end_time, client) triples. These are summed to rows in RealmCount corresponding to triples of (realm, end_time, client). And then these are summed to rows in InstallationCount with totals for pairs of (end_time, client).

Note: In most cases, we do not store rows with value 0. See Performance Strategy below.

CountStats

CountStats declare what analytics data should be generated and stored. The CountStat class definition and instances live in analytics/lib/counts.py. These declarations, along with any associated database queries, specify at a high level which tables should be populated by the system and with what data.

The core of a CountStat object is a parameterized raw SQL query, along with the respective parameter settings. A CountStat object + an end_time combine to give a full SQL query that aggregates data from the production database tables and inserts it into a *Count table.

Each CountStat object has the following fields. We’ll use the active_users:is_bot:day CountStat as a running example, which is a stat that keeps track of the number of active humans and active bots in each realm.

  • property: A unique, human-readable description, of the form “<english_description>:<subgroup_name>:<frequency>”. Example: “active_users:is_bot:day”.
  • zerver_count_query: A ZerverCountQuery object, which contains a
    • zerver_table: A table in zerver/models.py, to which filter_args are applied. E.g. UserProfile.
    • analytics_table: The *Count table where the data is initially collected. E.g. RealmCount.
    • query: A parameterized raw SQL string. E.g. count_user_by_realm_query.
  • filter_args: Filters the zerver_table. Example: {‘is_active’: True}, which restricts the UserProfiles under consideration to those with UserProfile.is_active = True .
  • group_by: The (table, field) being used for the subgroup. E.g. (UserProfile, is_bot).
  • frequency: How often to run the CountStat. Either ‘hour’ or ‘day’. E.g. ‘day’.
  • interval: Either ‘hour’, ‘day’, or ‘gauge’. If ‘hour’ or ‘day’, we’re interested in events that happen in the hour or day preceeding the end_time. If gauge, we’re interested in the state of the system at end_time. Example: ‘gauge’. (If ‘hour’, our example CountStat would instead be measuring the number of currently active users who joined in the last hour).

Note that one should be careful about making new gauge CountStats; see Performance Strategy below.

The FillState table

The default Zulip production configuration runs a cron job once an hour that updates the *Count tables for each of the CountStats in the COUNT_STATS dictionary. The FillState table simply keeps track of the last end_time that we successfully updated each stat. It also enables the analytics system to recover from errors (by retrying) and to monitor that the cron job is running and running to completion.

Performance strategy

An important consideration with any analytics system is performance, since it’s easy to end up processing a huge amount of data inefficiently and needing a system like Hadoop to manage it. For the built-in analytics in Zulip, we’ve designed something lightweight and fast that can be available on any Zulip server without any extra dependencies through the carefully designed set of tables in Postgres.

This requires some care to avoid making the analytics tables larger than the rest of the Zulip database or adding a ton of computational load, but with careful design, we can make the analytics system very low cost to operate. Also, note that a Zulip application database has 2 huge tables: Message and UserMessage, and everything else is small and thus not performance or space-sensitive, so it’s important to optimize how many expensive queries we do against those 2 tables.

There are a few important principles that we use to make the system efficient:

  • Not repeating work to keep things up to date (via FillState)
  • Storing data in the *Count tables to avoid our endpoints hitting the core Message/UserMessage tables is key, because some queries could take minutes to calculate. This allows any expensive operations to run offline, and then the endpoints to server data to users can be fast.
  • Doing expensive operations inside the database, rather than fetching data to Python and then sending it back to the database (which can be far slower if there’s a lot of data involved). The Django ORM currently doesn’t support the “insert into .. select” type SQL query that’s needed for this, which is why we use raw database queries (which we usually avoid in Zulip) rather than the ORM.
  • Aggregating where possible to avoid unnecessary queries against the Message and UserMessage tables. E.g. rather than querying the Message table both to generate sent message counts for each realm and again for each user, we just query for each user, and then add up the numbers for the users to get the totals for the realm.
  • Not storing rows when the value is 0. An hourly user stat would otherwise collect 24 * 365 * roughly .5MB per db row = 4GB of data per user per year, most of whose values are 0. A related note is to be cautious about adding gauge queries, since gauge measurements are typically non-zero rather than being typically zero.

Backend Testing

There are a few types of automated tests that are important for this sort of system:

  • Most important: Tests for the code path that actually populates data into the analytics tables. These are most important, because it can be very expensive to fix bugs in the logic that generates these tables (one basically needs to regenerate all of history for those tables), and these bugs are hard to discover. It’s worth taking the time to think about interesting corner cases and add them to the test suite.
  • Tests for the backend views code logic for extracting data from the database and serving it to clients.

For manual backend testing, it sometimes can be valuable to use ./manage.py dbshell to inspect the tables manually to check that things look right; but usually anything you feel the need to check manually, you should add some sort of assertion for to the backend analytics tests, to make sure it stays that way as we refactor.

LoggingCountStats

The system discussed above is designed primarily around the technical problem of showing useful analytics about things where the raw data is already stored in the database (e.g. Message, UserMessage). This is great because we can always backfill that data to the beginning of time, but of course sometimes one wants to do analytics on things that aren’t worth storing every data point for (e.g. activity data, request performance statistics, etc.). There is currently a reference implementation of a “LoggingCountStat” that shows how to handle such a situation.

Analytics UI development and testing

Setup and Testing

The main testing approach for the /stats page UI is manual testing. For UI testing, you want a comprehensive initial data set; you can use manage.py populate_analytics_db to set up, login as the shylock user, and then go to /stats.

Adding or editing /stats graphs

The relevant files are:

  • analytics/views.py: All chart data requests from the /stats page call get_chart_data in this file. The bottom half of this file (with all the raw sql queries) is for a different page (/activity), not related to /stats.
  • static/js/stats/stats.js: The JavaScript and Plotly code.
  • templates/analytics/stats.html
  • static/styles/stats.css and static/styles/portico.css: We are in the process of re-styling this page to use in-app css instead of portico css, but there is currently still a lot of portico influence.
  • analytics/urls.py: Has the URL routes; it’s unlikely you will have to modify this, including for adding a new graph.

Most of the code is self-explanatory, and for adding say a new graph, the answer to most questions is to copy what the other graphs do. It is easy when writing this sort of code to have a lot of semi-repeated code blocks (especially in stats.js); it’s good to do what you can to reduce this.

Tips and tricks:

  • Use $.get to fetch data from the backend. You can grep through stats.js to find examples of this.
  • The Plotly documentation is at https://plot.ly/javascript/ (check out the full reference, event reference, and function reference). The documentation pages seem to work better in Chrome than in Firefox, though this hasn’t been extensively verified.
  • Unless a graph has a ton of data, it is typically better to just redraw it when something changes (e.g. in the various aggregation click handlers) rather than to use retrace or relayout or do other complicated things. Performance on the /stats page is nice but not critical, and we’ve run into a lot of small bugs when trying to use Plotly’s retrace/relayout.
  • There is a way to access raw d3 functionality through Plotly, though it isn’t documented well.
  • ‘paper’ as a Plotly option refers to the bounding box of the graph (or something related to that).
  • You can’t right click and inspect the elements of a Plotly graph (e.g. the bars in a bar graph) in your browser, since there is an interaction layer on top of it. But if you hunt around the document tree you should be able to find it.