Mako Debug (Bug Bash)

Onsite debug round on the open-source Python Mako template engine. Candidates clone the repo, run failing tests, and locate/fix 2-3 bugs using an IDE debugger. Heavy emphasis on debugger fluency — print-based debugging usually loses signal. Requirements You are given a local checkout of Mako (a Python templating library) — a debug-mako folder containing a version of the engine with hidden bugs. There are 3 broken tests; you must find, explain, and fix each one using the test suite. The round is timeboxed at 45–60 minutes: you download the code, work on your own machine, and share your screen while the interviewer watches. You must narrate your reasoning out loud the entire time — explaining your thoughts clearly is an explicitly graded dimension alongside finding the bugs. Mako's documentation is open during the round; the codebase is unfamiliar to most candidates so setting breakpoints, stepping, and reading the call graph is the core skill. Setup (Step 1 of the round — practice this before the interview): Open the debug-mako folder. Install deps: pip install -r bugsquash-requirements.txt. Run the full suite to see the big picture: python -m pytest test/ -v. Re-run a single failing test for clearer output, e.g. python -m pytest test/test_lookup.py::TestClass::test_name -v, then trace from entry point to the failure. Bugs commonly reported (different threads report different subsets): Missing check on whether a path argument is a folder vs a file before processing. A missing visit_* function so AST args are never written through to output. A ControlLine check that only filters comments where it should filter comments-only lines. A tag missing / template-compiler stretch bug in the last 10 minutes. Notes Multiple reports say signal is graded on debugger usage, not edits. Use a real IDE (PyCharm / VS Code) with breakpoints. One candidate's Python install was incompatible with the provided Mako repo and they fell back to print debugging — they were rejected even though they discussed fixes verbally. Interviewer usually provides almost no help during this round, by design. 2/3 bugs solved is widely reported as a passing bar; one report had the interviewer stop them after bug 2 saying "good enough." Some candidates report the interviewer pressing them on whether their fix is the right fix vs a fix that makes the test pass — be ready to argue the change. Mako is the canonical open-source Python templating library maintained by the SQLAlchemy project. The codebase ships a standard pyproject.toml / tox.ini layout, so pip install -e .[testing] followed by pytest reproduces the round's test runner locally. The 1.3.x release stream is current; older 1.1.x clones can mask bugs that no longer exist in main, so pin the version the interviewer hands you. The starter checkout is typically delivered as a downloadable archive (e.g. a Google Drive link) rather than a live repo clone — download, unzip, then cd debug-mako before installing. Mako template constructs the bugs touch Mako compiles templates into Python for speed, so a planted bug usually breaks one of these surface constructs at a specific pipeline stage. Knowing what each construct should produce tells you what the failing test is really asserting: Embedded Python — <% ... %> runs arbitrary statements inside the template. Variable interpolation — ${variable} substitutes a value into the output (interpolation bugs surface in the AST visitor / codegen path). Template inheritance — <%inherit file="base.html"/> lets templates share a layout. Reusable components — <%def name="greeting(name)">…</%def> defines callable functions inside a template. Control flow — line-prefixed % for item in items: / % endfor and % if …: map to Python loops/conditionals (the ControlLine comment-filter bug lives here). Module layout to orient in fast The bugs sit across the standard Mako source tree; knowing the responsibility split saves time when a failing test points you at one stage of the pipeline: mako/ ├── template.py # Main render entry point ├── lexer.py # Tokenizes template source ├── codegen.py # Turns parsed templates into Python code ├── ast.py # AST node definitions ├── _ast_util.py # AST mutation helpers ├── lookup.py # Finds/loads template files (folder-vs-file checks live here) ├── parsetree.py # Parse-tree node definitions (ControlLine etc.) ├── pygen.py # Generates Python source ├── pyparser.py # Parses Python expressions ├── runtime.py # Helpers used while the template runs ├── filters.py # Output filter functions ├── cache.py # Template caching system ├── exceptions.py # Custom error classes ├── cmd.py # Command-line entry points ├── compat.py # Python-version compatibility shims ├── util.py # General helpers test/ ├── test_template.py, test_lexer.py, test_lookup.py, test_ast.py, ... The reported bugs cluster in lookup.py (path folder-vs-file), the AST visitor code (ast.py / _ast_util.py, the missing visit_*), and parsetree.py / codegen.py (the ControlLine comment filter and the compiler stretch bug). Bug taxonomy to scan for Most planted bugs fall into a few shapes; pattern-matching against them speeds up the hunt: Wrong conditions — an if whose logic is inverted or off-by-one versus what it should test. Missing edge cases — code that breaks on unexpected input (e.g. a folder passed where a file is assumed). AST / parse-tree issues — a node not created, or a visitor method that never writes its node through. String / syntax errors — mishandled template text or token boundaries. Preparation Clone the official Mako repo and step through the AST / visitor code with a debugger before the interview; the mako/ast.py and mako/codegen.py modules are where most reported bugs live. Pre-install the same Python version + Mako repo locally, confirm pytest runs all tests, and confirm your IDE can hit breakpoints inside Mako's internals. Drill on reading unfamiliar Python by exploring an open-source library you've never touched: set a goal like "explain how requests.Session sends a redirect," trace it with a debugger. During the round, narrate your hypothesis before each step; interviewers want to hear the reasoning chain, not just the eventual diff.