Awesome

pytrace - a fast python tracer

pytrace records function calls, arguments and return values.

traces aid debugging, profiling and obviate logging.

pytrace has been tested on python 2.7 and python 3.2. (should support 2.6 and up)

pytrace has been tested on os x and several linux distributions.

Follow @alonhorev on twitter for updates.

pytrace depends on sqlite and a C implementation of protocol buffers.

on debian/ubuntu: sudo apt-get install libsqlite3-dev libprotobuf-c0-dev

on fedora: sudo yum install libsqlite3x-devel sqlite-devel python-devel protobuf-c-devel

on mac (sqlite is included): brew install protobuf-c or port install protobuf-c

install pytrace:

pip install pytrace

Invoke pytrace with your script:

$ pytrace foo.py --bar

Invoke the reader from the same directory by executing pytrace with no arguments:

$ pytrace

The reader can be invoked while the script is running. providing 'online' debugging capabilities.

The collected data can be viewed in an interactive reader. The reader supports less-like key bindings.

trace reader

The reader can search for regular expressions:

The reader can filter traces using a python expressions. The following fields can be used for filters:

time - int
tid - int
module - string
func - string
arg (argument name) - string. special arguments are 'return value' and 'exception' used to filter function return values and exceptions.
value (argument value) - string
type (argument type) - string

Field types:

int fields - supports algebric operators (>, <, >=, <=, ==). e.g: time > '2012/08/15 01:23:45'.
string fields: support string comparison (==, !=). strings comparison supports sql 'like' syntax. for example: module == 'proj%' filters modules starting with 'proj'.

trace filter

The reader corrects queries that don't match anything:

filter correction

The database is saved in the current working directory and is named traces.sqlite.

In order to not run out of disk space, The database will be truncated to a fixed number of traces (currently hard coded to 10000).

Hot functions can be skipped using a decorator:

from pytrace import notrace

@notrace
def hot():
	pass

Trace specific packages:

$ export TRACE_MODULES=/Users/alon/project

You can specify a colon : separated list of folders as well.

pytrace can be broken down to three parts:

a trace generator - using python's built-in tracing mechanism (sys.settrace) function calls are translated to binary trace records saved in memory as protocol buffers (http://code.google.com/p/protobuf-c/).
a trace dumper - runs in a separate thread/process, collects traces from memory and dumps them to a sqlite database.
a trace reader - reads traces from the database.

The separation of trace generation and dumping has several advantages:

speed - the overhead on the traced code is minimal, serialization to protocol buffers is faster then inserting to a database.
prioritization - the trace records are saved to a lock-less circular buffer. The generator has priority over the dumper, meaning, if the dumper doesn't keep up it will lose traces (the generator never blocks).
Python utilizes two cores! - the dumper thread does not touch python objects, only trace records that are saved as binary strings. therefore, it doesn't acquire the global interpreter lock.
Trace aggregation from several processes - By using shared memory the trace data can be shared between a dumper process and one or more generator processes.