Pyccolo is a library for embedding source-level instrumentation in
Python (as opposed to bytecode-level). It aims to be ergonomic, composable,
and portable, by providing an intuitive interface, making it easy to layer
multiple levels of instrumentation, and allowing the same code to work across
multiple versions of Python (3.6 to 3.10), with few exceptions.
Pyccolo can be used (and has been used) to implement various kinds of dynamic analysis
tools and other instrumentation:
- Code coverage (see pyccolo/examples/coverage.py)
- Syntax-augmented Python (3.8 and up, see pyccolo/examples/null_coalesce.py)
- Dynamic dataflow analysis performed by nbsafety
- Tools to find unused imports at runtime (candidates for lazy importing)
- Tools to uncover semantic memory leaks
Below is a simple script that uses Pyccolo to print “Hello, world!” before
every statement that executes:
import pyccolo as pyc class HelloTracer(pyc.BaseTracer): @pyc.before_stmt def handle_stmt(self, *_, **__): print("Hello, world!") if __name__ == "__main__": with HelloTracer(): # prints "Hello, world!" 11 times pyc.exec("for _ in range(10): pass")
Instrumentation is provided by a tracer class that inherit from
pyccolo.BaseTracer. This class rewrites Python source code with instrumentation that
triggers whenever events of interest occur, such as when a
statement is about to execute. By registering a handler with the associated
event (with the
@pyc.before_stmt decorator, in this case), we can enrich our
programs with additional observability, or even alter their behavior
What is up with
A program’s abstract syntax tree is fixed at import / compile time, and when
our script initially started running, the tracer was not active, so unquoted
Python in the same file will lack instrumentation. It is possible to instrument
modules at import time, but only when the imports are performed inside a
tracing context. Thus, we must quote any code appearing in the same module
where the tracer class was defined in order to instrument it.
A core feature of Pyccolo is that its instrumentation is composable. It’s
usually tricky to use two or more
ast.NodeTransformer classes simultaneously
— sometimes you can just have one inherit from the other, but if they both
visit methods for the same AST node type, then typically you would
need to define a bespoke node transformer that uses logic from each base
transformer, handling corner cases to resolve incompatibilities. With Pyccolo,
you simply compose the context managers of each tracer class whose
instrumentation you wish to use, and everything usually Just
with tracer1: with tracer2: pyc.exec(...)
Compatibility with sys.settrace(…)
Pyccolo is designed to support not only AST-level instrumentation, but also
instrumentation involving Python’s built in tracing
To use it, you simply register handlers for one of the corresponding
Pyccolo events (
Here’s a minimal example:
import pyccolo as pyc class SysTracer(pyc.BaseTracer): @pyc.call def handle_call(self, *_, **__): print("Pushing a stack frame!") @pyc.return_ def handle_return(self, *_, **__): print("Popping a stack frame!") if __name__ == "__main__": with SysTracer(): def f(): def g(): return 42 return g() # push, push, pop, pop answer_to_life_universe_everything = f()
Note that we didn’t need to use
pyc.exec(...) in the above example, because Python’s built-in
tracing does not involve any AST-level transformations. If, however, we had registered handlers
for other events, such as
pyc.before_stmt, we would need to use
pyc.exec(...) to ensure those
handlers get called, when running code in the same file where our tracer class is defined.
What if I’m already using sys.settrace(…) with my own tracing function?
Pyccolo is designed to be composable, and should execute both your tracing function as well
as any handlers defined in any active Pyccolo tracers. For example Pyccolo’s unit tests for
return events work even when coverage.py
is active (and without breaking it), which also uses Python’s built-in tracing utilities.
Instrumenting Imported Modules
Instrumentation is opt-in for modules imported within tracing contexts. To determine whether
a module gets instrumented, the method
should_instrument_file(...) is called with the module’s
corresponding filename as input. For example:
Imports are instrumented by registering a custom finder / loader with
This loader ignores cached bytecode (which may possibly be uninstrumented), and avoids
generating new cached bytecode (which would be instrumented, possibly causing confusion
later when instrumentation is not desired).
Command Line Interface
You can execute arbitrary scripts with instrumentation enabled with the
pyc command line tool.
For example, to use the
NullCoalescer tracer defined in pyccolo/examples/null_coalesce.py,
you can call
pyc as follows, given some example script
# bar.py bar = None # prints `None` since bar?.foo coalesces to `None` print(bar?.foo)
baras a module (indeed,
pycperforms this internally when provided a file):
in case you were not already aware, Pyccolo is composable! 🙂
The above example demonstrates a tracer class that performs syntax augmentation on its
instrumented Python source to modify the default Python syntax. This feature is available
only on Python >= 3.8 for now and is lacking documentation for the moment, but you can
see some examples in the test_syntax_augmentation.py unit tests.
Pyccolo handlers can be registered for many kinds of events. Some of the more common ones are:
pyc.before_stmt, emitted before a statement executes;
pyc.after_stmt, emitted after a statement executes;
pyc.before_attribute_load, emitted in load contexts before an attribute is accessed;
pyc.after_attribute_load, emitted in load contexts after an attribute is accessed;
pyc.load_name, emitted when a variable is used in a load context (e.g.
bar = foo.baz);
pyc.return_, two non-AST trace events built-in to Python.
There are many different Pyccolo events, and more are always being added. See
for a full list.
Note that, for AST events, Python source is only transformed to emit some event
when there is at least one tracer active that has at least one handler
registered for that event. This prevents the transformed source from becoming
extremely bloated when only a few events are needed.
Every Pyccolo handler is passed four positional arguments:
- The return value, for instrumented expressions;
- The AST node (or node id, if using
- The stack frame, at the point where instrumentation kicks in;
- The event (useful when the same handler is registered for multiple events).
Some events pass additional keyword arguments, which I’m still in the process
of documenting, but the above four tend to suffice for most use cases.
Not every handler receives a return value; for example, this argument is always
pyc.after_stmt handlers. For certain handlers, the return value
can be overridden. For example, by returning a value in a
pyc.before_attribute_load, we override the object whose attribute is
accessed. If we return nothing or
None, then we do not override this object.
(If we actually want to override it as
None for some reason, then we can
pyc.Null.) For a particular event, handler return values compose with
other handlers defined on the same tracer class as well as with handlers
defined on other tracer classes.
Pyccolo instrumentation adds significant overhead to Python. In some
cases, this overhead can be partially mitigated if, for example, you only need
instrumentation the first time a statement runs. In such cases, you can
deactivate instrumentation after, e.g., the first time a function executes, or
after the first iteration in a loop for that respective function or loop, so
that further calls (iterations, respectively) use uninstrumented code with all
the mighty performance of native Python. This is implemented by activating
“guards” associated with the function or loop, as in the below example:
class TracesOnce(pyc.BaseTracer): @pyc.register_raw_handler((pyc.after_for_loop_iter, pyc.after_while_loop_iter)) def after_loop_iter(self, *_, guard, **__): self.activate_guard(guard) @pyc.register_raw_handler(pyc.after_function_execution) def after_func
NOW WITH OVER +8500 USERS. people can Join Knowasiak for free. Sign up on Knowasiak.com