Developer Guide

In general we prefer simplicity. We standardize on JavaScript/TypeScript (Node.js) and SQL (PostgreSQL) as the languages of implementation and try to minimize the number of complex libraries or frameworks being used. The website is server-side generated pages with minimal client-side JavaScript.

High level view

High level system structure

  • The questions and assessments for a course are stored in a git repository. This is synced into the database by the course instructor and DB data is updated or added to represent the course. Students then interact with the course website by doing questions, with the results being stored in the DB. The instructor can view the student results on the website and download CSV files with the data.
  • The majority of course content and configuration is done via plain text files in the git repository, which is the master source for this data.
  • All student data is all stored in the DB and is not pushed back into the git repository or disk at any point.

Unit tests and integration tests

  • Integration tests are stored in the apps/prairielearn/src/tests/ directory.
  • Unit tests are typically located next to the file under test, with the filename ending in .test.ts. For instance, tests for foo.ts would be in foo.test.ts in the same directory.
  • The tests are run by GitHub Actions on every push to GitHub.
  • The tests are mainly integration tests that start with a blank database, run the server to initialize the database, load the testCourse, and then emulate a client web browser that answers questions on assessments. If a test fails then it is often easiest to debug by recreating the error by doing questions yourself against a locally-running server.
  • If the PL_KEEP_TEST_DB environment is set, the test database (normally pltest_1, pltest_2, etc.) won't be removed when testing ends. This allows you inspect the state of the database whenever your testing ends. The database will get overwritten when you start a new test run.

Debugging server-side JavaScript

  • Use the debug package to help trace execution flow in JavaScript. To run the server with debugging output enabled:

    DEBUG=* make dev
  • To just see debugging logs from PrairieLearn you can use:

    DEBUG=prairielearn:* make dev
  • To insert more debugging output, import debug and use it like this:

    import debugfn from 'debug';
    const debug = debugfn('prairielearn:my-file');
    // in some function later
    debug('func()', 'param:', param);
  • UnhandledPromiseRejectionWarning errors are frequently due to improper async/await handling. Make sure you are calling async functions with await/.then()/.catch() as appropriate, and that async functions are not being called from callback-style code without a callbackify(). To get more information, Node can be run with the --trace-warnings flag. For example, node_modules/.bin/mocha --trace-warnings tests/index.js.

Debugging client-side JavaScript

  • Make sure you have the JavaScript Console open in your browser and reload the page.

Debugging SQL and PL/pgSQL

  • Use the psql commandline interface to test SQL separately. A default development PrairieLearn install uses the postgres database, so you should run:

    psql postgres
  • To debug syntax errors in a stored procedure, import it manually with \i filename.sql in psql.
  • To follow execution flow in PL/pgSQL use RAISE NOTICE. This will log to the console when run from psql and to the server log file when run from within PrairieLearn. The syntax is:

    RAISE NOTICE 'This is logging: % and %',
  • To manually run a function:

      the_sql_function (arg1, arg2);

HTML page generation

  • Express is used as the web framework.
  • All pages are server-side rendered and we try and minimize the amount of client-side JavaScript. Client-side JS should use vanilla JavaScript/TypeScript where possible, but third-party libraries may be used when appropriate.
  • Each web page typically has all its files in a single directory, with the directory, the files, and the URL all named the same. Not all pages need all files. For a real-world example, consider the page where users can accept the PrairieLearn terms and conditions, located at apps/prairielearn/src/ee/pages/terms. That directory contains the following files:

    • terms.ts: The main entry point for the page. It runs SQL queries and renders a template.
    • terms.sql: All SQL queries for the page.
    • terms.html.ts: The template for the page. Exports a function that returns an HTML document.
  • When possible, prefer explicitly passing individual typed properties to templates instead of adding properties to res.locals. However, res.locals may be used for data coming from middlewares that will be used on many pages.
  • Use @prairielearn/html to generate HTML pages. This uses HTML tagged-template literals to generate HTML, which in turn makes it easy to get full type-checking.
  • Reused templates are stored in the apps/prairielearn/src/components/ directory. These should generally accept an object with properties instead of being passed the full res.locals object.

HTML style

  • Use Bootstrap as the style. As of 2024-06-05 we are using v4.
  • Local CSS rules go in public/stylesheets/local.css. Try to minimize use of this and use plain Bootstrap styling wherever possible.
  • Buttons should use the <button> element when they take actions and the <a> element when they are simply links to other pages. We should not use <a role="button"> to fake a button element. Buttons that do not submit a form should always start with <button type="button" class="btn ...">, where type="button" specifies that they don't submit.

SQL usage

  • Write raw SQL rather than using a ORM library. This reduces the number of frameworks/languages needed.
  • Prefer implementing complex logic in TypeScript instead of inside queries.
  • Use the SQL convention of snake_case for names. Also use the same convention in JavaScript for names that are the same as in SQL, so the question_id variable in SQL is also called question_id in JavaScript code.
  • Use uppercase for SQL reserved words like SELECT, FROM, AS, etc.
  • SQL code should not be inline in JavaScript files. Instead it should be in a separate .sql file, following the Yesql concept. Each filename.js file will normally have a corresponding filename.sql file in the same directory. The .sql file should look like:

    -- BLOCK select_question
      id = $question_id;
    -- BLOCK insert_submission
      submissions (submitted_answer)

From JavaScript you can then do:

import { loadSqlEquiv, queryRow } from '@prairielearn/postgres';

import { QuestionSchema } from './lib/db-types.js';

const sql = loadSqlEquiv(import.meta.url);

const question = await queryRow(sql.select_question, { question_id: 45 }, QuestionSchema);
  • To keep SQL code organized it is a good idea to use CTEs (WITH queries). These are formatted like:

      first_preliminary_table AS (
          -- first preliminary query
      second_preliminary_table AS (
          -- second preliminary query
      -- main query here
      first_preliminary_table AS fpt,
      second_preliminary_table AS spt;

DB stored procedures (sprocs)

  • Stored procedures are created by the files in sprocs/. To call a stored procedure from JavaScript, use code like:

    const workspace_id = 1342;
    const message = 'Startup successful';
    await sqldb.callAsync('workspaces_message_update', [workspace_id, message]);
  • The stored procedures are all contained in a separate database schema with a name like server_2021-07-07T20:25:04.779Z_T75V6Y. To see a list of the schemas use the \dn command in psql.
  • To be able to use the stored procedures from the psql command line it is necessary to get the most recent schema name using \dn and set the search_path to use this quoted schema name and the public schema:

    set search_path to "server_2021-07-07T20:25:04.779Z_T75V6Y",public;
  • During startup we initially have no non-public schema in use. We first run the migrations to update all tables in the public schema, then we call sqldb.setRandomSearchSchemaAsync() to activate a random per-execution schema, and we run the sproc creation code to generate all the stored procedures in this schema. This means that every invocation of PrairieLearn will have its own locally-scoped copy of the stored procedures which are the correct versions for its code. This lets us upgrade PrairieLearn servers one at a time, while old servers are still running with their own copies of their sprocs. When PrairieLearn first starts up it has search_path = public, but later it will have search_path = "server_2021-07-07T20:25:04.779Z_T75V6Y",public so that it will first search the random schema and then fall back to public. The naming convention for the random schema uses the local instance name, the date, and a random string. Note that schema names need to be quoted using double-quotations in psql because they contain characters such as hyphens.
  • For more details see sprocs/array_and_number.sql and comments in server.js near the call to sqldb.setRandomSearchSchemaAsync().

DB schema (simplified overview)

  • The most important tables in the database are shown in the diagram below (also as a PDF image).

Simplified DB Schema

  • Each table has an id number that is used for cross-referencing. For example, each row in the questions table has an id and other tables will refer to this as a question_id. For legacy reasons, there are two exceptions to this rule:

    • Tables that reference the pl_courses table use course_id instead of pl_course_id.
    • The users table has a user_id primary key instead of id.
  • Each user is stored as a single row in the users table.
  • The pl_courses table has one row for each course, like TAM 212.
  • The course_instances table has one row for each semester ("instance") of each course, with the course_id indicating which course it belongs to.
  • Every question is a row in the questions table, and the course_id shows which course it belongs to. All the questions for a course can be thought of as the "question pool" for that course. This same pool is used for all semesters (all course instances).
  • Assessments are stored in the assessments table and each assessment row has a course_instance_id to indicate which course instance (and hence which course) it belongs to. An assessment is something like "Homework 1" or "Exam 3". To determine this we can use the assessment_set_id and number of each assessment row.
  • Each assessment has a list of questions associated with it. This list is stored in the assessment_questions table, where each row has a assessment_id and question_id to indicate which questions belong to which assessment. For example, there might be 20 different questions that are on "Exam 1", and it might be the case that each student gets 5 of these questions randomly selected.
  • Each student will have their own copy of an assessment, stored in the assessment_instances table with each row having a user_id and assessment_id. This is where the student's score for that assessment is stored.
  • The selection of questions that each student is given on each assessment is in the instance_questions table. Here each row has an assessment_question_id and an assessment_instance_id to indicate that the corresponding question is on that assessment instance. This row will also store the student's score on this particular question.
  • Questions can randomize their parameters, so there are many possible variants of each question. These are stored in the variants table with an instance_question_id indicating which instance question the variant belongs to.
  • For each variant of a question that a student sees they will have submitted zero or more submissions with a variant_id to show what it belongs to. The submissions row also contains information the submitted answer and whether it was correct.

DB schema (full data)

DB Schema

DB schema conventions

  • Tables have plural names (e.g. assessments) and always have a primary key called id. The foreign keys pointing to this table are non-plural, like assessment_id. When referring to this use an abbreviation of the first letters of each word, like ai in this case. The only exceptions are aset for assessment_sets (to avoid conflicting with the SQL AS keyword), top for topics, and tag for tags (to avoid conflicts). This gives code like:

    -- select all active assessment_instances for a given assessment
      assessments AS a
      JOIN assessment_instances AS ai ON (ai.assessment_id =
    WHERE = 45
      AND ai.deleted_at IS NULL;
  • We (almost) never delete student data from the DB. To avoid having rows with broken or missing foreign keys, course configuration tables (e.g. assessments) can't be actually deleted. Instead they are "soft-deleted" by setting the deleted_at column to non-NULL. This means that when using any soft-deletable table we need to have a WHERE deleted_at IS NULL to get only the active rows.

DB schema modification

See migrations/

Database access

  • DB access is via the @prairielearn/postgres package. This wraps the node-postgres library.
  • For single queries we normally use the following pattern, which automatically uses connection pooling from node-postgres and safe variable interpolation with named parameters and prepared statements:

    const questions = await queryRows(
      { course_id: 45 },

Where the corresponding filename.sql file contains:

-- BLOCK select_questions_by_course
  course_id = $course_id;
  • For queries where it would be an error to not return exactly one result row:

    const question = await queryRow(sql.block_name, QuestionSchema);
  • Use explicit row locking whenever modifying student data related to an assessment. This must be done within a transaction. The rule is that we lock either the variant (if there is no corresponding assessment instance) or the assessment instance (if we have one). It is fine to repeatedly lock the same row within a single transaction, so all functions involved in modifying elements of an assessment (e.g., adding a submission, grading, etc) should call a locking function when they start. All locking functions are equivalent in their action, so the most convenient one should be used in any given situation:

    Locking function Argument
    assessment_instances_lock assessment_instance_id
    instance_questions_lock instance_question_id
    variants_lock variant_id
    submission_lock submission_id
  • To pass an array of parameters to SQL code, use the following pattern, which allows zero or more elements in the array. This replaces $points_list with ARRAY[10, 5, 1] in the SQL. It's required to specify the type of array in case it is empty:

    await sqldb.queryAsync(sql.insert_assessment_question, {
      points_list: [10, 5, 1],
    -- BLOCK insert_assessment_question
      assessment_questions (points_list)
  • To use a JavaScript array for membership testing in SQL use unnest() like:

    const questions = await sqldb.queryRows(
      { id_list: [7, 12, 45] },
    -- BLOCK select_questions
      id IN (
  • To pass a lot of data to SQL a useful pattern is to send a JSON object array and unpack it in SQL to the equivalent of a table. This is the pattern used by the "sync" code, such as sprocs/sync_news_items.sql. For example:

    let data = [
      { a: 5, b: 'foo' },
      { a: 9, b: 'bar' },
    await sqldb.queryAsync(sql.insert_data, {
      data: JSON.stringify(data),
    -- BLOCK insert_data
      my_table (a, b)
      jsonb_to_recordset($data) AS (a INTEGER, b TEXT);
  • To use a JSON object array in the above fashion, but where the order of rows is important, use ROWS FROM () WITH ORDINALITY to generate a row index like this:

    -- BLOCK insert_data
      my_table (a, b, order_by)
      (jsonb_to_recordset($data) AS (a INTEGER, b TEXT))
      ORDINALITY AS data (a, b, order_by);

Asynchronous control flow in JavaScript

  • New code in PrairieLearn should use async/await whenever possible.
  • Use the async library for complex control flow or when mixing Promise-based and callback-based code.

Using async route handlers with ExpressJS

  • Express can't directly use async route handlers. Instead we use express-async-handler like this:

    import asyncHandler from 'express-async-handler';
      asyncHandler(async (req, res, next) => {
        // can use "await" here

Security model

  • We distinguish between authentication and authorization. Authentication occurs as the first stage in server response and the authenticated user data is stored as res.locals.authn_user.
  • The authentication flow is:

    1. We first redirect to a remote authentication service (e.g. SAML SSO, Google, Microsoft).

    2. The remote authentication service redirects back to a callback URL, e.g. /pl/oauth2callback for Google. These endpoints confirm authentication, create the user in the users table if necessary, set a signed pl_authn cookie in the browser with the authenticated user_id, and then redirect to the main PL homepage. This cookie is set with the HttpOnly attribute, which prevents client-side JavaScript from reading the cookie.

    3. Every other page authenticates using the signed browser pl_authn cookie. This is read by middlewares/authn.js which checks the signature and then loads the user data from the DB using the user_id, storing it as res.locals.authn_user.

  • Similar to unix, we distinguish between the real and effective user. The real user is stored as res.locals.authn_user and is the user that authenticated. The effective user is stored as res.locals.user. Only users with role = TA or higher can set an effective user that is different from their real user. Moreover, users with role = TA or higher can also set an effective role and mode that is different to the real values.
  • Authorization occurs at multiple levels:

    • The course_instance checks authorization based on the authn_user.
    • The course_instance authorization is checked against the effective user.
    • The assessment checks authorization based on the effective user, role, mode, and date.
  • All state-modifying requests must (normally) be POST and all associated data must be in the body. GET requests may use query parameters for viewing options only.

State-modifying POST requests

  • Use the Post/Redirect/Get pattern for all state modification. This means that the initial GET should render the page with a <form> that has no action set, so it will submit back to the current page. This should be handled by a POST handler that performs the state modification and then issues a redirect back to the same page as a GET:
      asyncHandler(async (req, res) => {
        if (req.body.__action == 'enroll') {
          await queryAsync(sql.enroll, {
            course_instance_id: req.body.course_instance_id,
            user_id: res.locals.authn_user.user_id,
        } else {
          throw new error.HttpStatusError(400, `unknown __action: ${req.body.__action}`);
  • All data modifying requests should come from form elements like:

    <form name="enroll-form" method="POST">
      <input type="hidden" name="__action" value="enroll" />
      <input type="hidden" name="__csrf_token" value="${__csrf_token}" />
      <input type="hidden" name="course_instance_id" value="56" />
      <button type="submit" class="btn btn-info">Enroll in course instance 56</button>
  • The res.locals.__csrf_token variable is set and checked by early-stage middleware, so no explicit action is needed on each page.

Logging errors

  • We use Winston for logging to the console and to files:

    import { logger } from '@prairielearn/logger';'This is an info message');
    logger.error('This is an error message');
    // This will be logged to the log file, but not to the console:
    logger.verbose('This is a verbose message');
  • All logger functions have a mandatory first argument that is a string, and an optional second argument that is an object containing useful information. It is important to always provide a string as the first argument.

Coding style

ESLint and Prettier are used to enforce consistent code conventions and formatting throughout the codebase. See .eslintrc.js and .prettierrc.json in the root of the PrairieLearn repository to view our specific configuration. The repo includes an .editorconfig file that most editors will detect and use to automatically configure things like indentation. If your editor doesn't natively support an EditorConfig file, there are plugins available for most other editors.

For Python files, ruff is used for autoformatting and enforcing code conventions, and Pyright is used for static typechecking. See pyproject.toml in the root of the PrairieLearn repository to view our specific configuration. We encourage all new Python code to include type hints for use with the static typechecker, as this makes it easier to read, review, and verify contributions.

To lint the code, use make lint. This is also run by the CI tests.

To automatically fix lint and formatting errors, run make format.

Question-rendering control flow

  • The above files are all called/included by each of the top-level pages that needs to render a question (e.g., pages/instructorQuestionPreview, pages/studentInstanceQuestion, etc). Unfortunately the control-flow is complicated because we need to call lib/question-render.js during page data load, store the data it generates, and then later include the components/QuestionContainer.html.ts template to actually render this data.
  • For example, the exact control-flow for pages/instructorQuestion is:

    1. The top-level page pages/instructorQuestion/instructorQuestion.js code calls lib/question-render.getAndRenderVariant().

    2. getAndRenderVariant() inserts data into res.locals for later use by components/QuestionContainer.html.ts.

    3. The top-level page code renders the top-level template pages/instructorQuestion/instructorQuestion.html.ts, which then includes components/QuestionContainer.html.ts.

    4. components/QuestionContainer.html.ts renders the data that was earlier generated by lib/question-render.js.

Question open status

  • There are three levels at which “open” status is tracked, as follows. If open = false for any object then it will block the creation of new objects below it. For example, to create a new submission the corresponding variant, instance_question, and assessment_instance must all be open.

    Variable Allow new instance_questions Allow new variants Allow new submissions

Errors in question handling

  • We distinguish between two different types of student errors:

    1. The answer might be not be gradable (submission.gradable = false). This could be due to a missing answer, an invalid format (e.g., entering a string in a numeric input), or a answer that doesn't pass some basic check (e.g., a code submission that didn't compile). This can be discovered during either the parsing or grading phases. In such a case the submission.format_errors object should store information on what was wrong to allow the student to correct their answer. A submission with gradable = false will not cause any updating of points for the question. That is, it acts like a saved-but-not-graded submission, in that it is recorded but has no impact on the question. If gradable = false then the score and feedback will not be displayed to the student.

    2. The answer might be gradable but incorrect. In this case submission.gradable = true but submission.score = 0 (or less than 1 for a partial score). If desired, the object can be set to give information to the student on what was wrong with their answer. This is not necessary, however. If is set then it will be shown to the student along with their submission.score as soon as the question is graded.

  • There are three levels of errors that can occur during the creation, answering, and grading of a question:

    Error level Caused Stored Reported Effect
    System errors Internal PrairieLearn errors On-disk logs Error page Operation is blocked. Data is not saved to the database.
    Question errors Errors in question code issues table Issue panels on the question page variant.broken_at != null or submission.broken == true. Operation completes, but future operations are blocked.
    Student errors Invalid data submitted by the student (unparsable or ungradable) submission.gradable set to false and details are stored in submission.format_errors Inside the rendered submission panel The submission is not assigned a score and no further action is taken (e.g., points are changed for the instance question). The student can resubmit to correct the error.
  • The important variables involved in tracking question errors are:

    Variable Error level Description
    variant.broken_at Question error Set to NOW() if there were question code errors in generating the variant. Such a variant will be not have render() functions called, but will instead be displayed as This question is broken.
    submission.broken Question error Set to true if there question code errors in parsing or grading the variant. After submission.broken is true, no further actions will be taken with the submission.
    issues table Question error Rows are inserted to record the details of the errors that caused variant.broken != null or submission.broken == true to be set to true.
    submission.gradable Student error Whether this submission can be given a score. Set to false if format errors in the submitted_answer were encountered during either parsing or grading.
    submission.format_errors Student error Details on any errors during parsing or grading. Should be set to something meaningful if gradable = false to explain what was wrong with the submitted answer.
    submission.graded_at None NULL if grading has not yet occurred, otherwise a timestamp.
    submission.score None Final score for the submission. Only used if gradable = true and graded_at is not NULL. None Feedback generated during grading. Only used if gradable = true and graded_at is not NULL.
  • Note that submission.format_errors stores information about student errors, while the issues table stores information about question code errors.
  • The question flow is shown in the diagram below (also as a PDF image).

    Question flow

JavaScript equality operator

You should almost always use the === operator for comparisons; this is enforced with an ESLint rule.

The only case where the == operator is frequently useful is for comparing entity IDs that may be coming from the client/database/etc. These may be either strings or numbers depending on where they're coming from or how they're fetched. To make it abundantly clear that ids are being compared, you should use the idsEqual utility:

import { idsEqual } from './lib/id';

console.log(idsEqual(12345, '12345'));
// > true

"Modern" queries that use Zod validation will automatically coerce all IDs to strings. If you're confident that data on both sides of the comparison is coming from a Zod-validated query, you can use the === operator directly.