How coverage.py works

For advanced use of coverage.py, or just because you are curious, it helps to understand what’s happening behind the scenes.

Coverage.py works in three phases:

Execution: Coverage.py runs your code, and monitors it to see what lines were executed.
Analysis: Coverage.py examines your code to determine what lines could have run.
Reporting: Coverage.py combines the results of execution and analysis to produce a coverage number and an indication of missing execution.

The execution phase is handled by the coverage run command. The analysis and reporting phases are handled by the reporting commands like coverage report or coverage html.

As a short-hand, I say that coverage.py measures what lines were executed. But it collects more information than that. It can measure what branches were taken, and if you have contexts enabled, for each line or branch, it will also measure what contexts they were executed in.

Let’s look at each phase in more detail.

Execution

At the heart of the execution phase is a trace function. This is a function that the Python interpreter invokes for each line executed in a program. Coverage.py implements a trace function that records each file and line number as it is executed.

For more details of trace functions, see the Python docs for sys.settrace, or if you are really brave, How C trace functions really work.

Executing a function for every line in your program can make execution very slow. Coverage.py’s trace function is implemented in C to reduce that overhead. It also takes care to not trace code that you aren’t interested in.

When measuring branch coverage, the same trace function is used, but instead of recording line numbers, coverage.py records pairs of line numbers. Each invocation of the trace function remembers the line number, then the next invocation records the pair (prev, this) to indicate that execution transitioned from the previous line to this line. Internally, these are called arcs.

As the data is being collected, coverage.py writes the data to a file, usually named .coverage. This is a SQLite database containing all of the measured data.

Plugins

Of course coverage.py mostly measures execution of Python files. But it can also be used to analyze other kinds of execution. File tracer plugins provide support for non-Python files. For example, Django HTML templates result in Python code being executed somewhere, but as a developer, you want that execution mapped back to your .html template file.

During execution, each new Python file encountered is provided to the plugins to consider. A plugin can claim the file and then convert the runtime Python execution into source-level data to be recorded.

Dynamic contexts

When using dynamic contexts, there is a current dynamic context that changes over the course of execution. It starts as empty. While it is empty, every time a new function is entered, a check is made to see if the dynamic context should change. While a non-empty dynamic context is current, the check is skipped until the function that started the context returns.

Analysis

After your program has been executed and the line numbers recorded, coverage.py needs to determine what lines could have been executed. Luckily, compiled Python files (.pyc files) have a table of line numbers in them. Coverage.py reads this table to get the set of executable lines, with a little more source analysis to leave out things like docstrings.

The data file is read to get the set of lines that were executed. The difference between the executable lines and the executed lines are the lines that were not executed.

The same principle applies for branch measurement, though the process for determining possible branches is more involved. Coverage.py uses the abstract syntax tree of the Python source file to determine the set of possible branches.

Reporting

Once we have the set of executed lines and missing lines, reporting is just a matter of formatting that information in a useful way. Each reporting method (text, HTML, JSON, annotated source, XML) has a different output format, but the process is the same: write out the information in the particular format, possibly including the source code itself.