Skip to main content

Lightweight Condition Parsing and Building of Evaluation Expressions

Project description


author: gk version: 190523


pycond: Lightweight Declarative Condition Expressions

Build Status codecovPyPI    version

Table Of Contents

What

You have a bunch of data, possibly streaming...

id,first_name,last_name,email,gender,ip_address
1,Rufe,Morstatt,rmorstatt0@newsvine.de,Male,216.70.69.120
2,Kaela,Scott,scott@opera.com,Female,73.248.145.44,2
(...)

... and you need to filter. For now lets say we have them already as list of dicts.

You can do it imperatively:

foo_users = [ u for u in users
              if ([u['gender'] == 'Male' or u['last_name'] == 'Scott') and
                  '@' in u['email']) ]

or you have this module assemble a condition function from a declaration like:

from pycond import parse_cond
cond = 'email contains .de and gender eq Male or last_name eq Scott'
is_foo = parse_cond(cond)

and then apply as often as you need, against varying state / facts / models (...):

foo_users = [ u for u in users if is_foo(state=u) ]

with roughly the same performance (factor 2-3) than the handcrafted python.

In real life performance is often better then using imperative code, due to pycond's lazy evaluation feature.

Why

When the developer can decide upon the filters to apply on data he'll certainly use Python's excellent expressive possibilities directly, e.g. as shown above through list comprehensions.
But what if the filtering conditions are based on decisions outside of the program's control? I.e. from an end user, hitting the program via the network, in a somehow serialized form, which is rarely directly evaluatable Python.

This is the main use case for this module.

Alternatives

But why yet another tool for such a standard job?

There is a list of great tools and frameworks where condition parsing is a (small) part of them, e.g. pyke or durable and many in the django world or from SQL statement parsers.

1. I just needed a very slim tool for only the parsing into functions - but this pretty transparent and customizable

pycond allows to customize

  • the list of condition operators
  • the list of combination operators
  • the general behavior of condition operators via global or condition local wrappers
  • their names
  • the tokenizer
  • the value lookup function

and ships as zero dependency single module.

All evaluation is done via partials and not lambdas, i.e. operations can be introspected and debugged very simply, through breakpoints or custom logging operator or lookup wrappers.

2. Simplicity of the grammar: Easy to type directly, readable by non programmers but also synthesisable from structured data, e.g. from a web framework.

3. Performance: Good enough to have "pyconditions" used within stream filters. With the current feature set we are sometimes a factor 2-3 worse but (due to lazy eval) often better, compared with handcrafted list comprehensions.

Mechanics

Parsing

pycond parses the condition expressions according to a set of constraints given to the parser in the tokenizer function. The result of the tokenizer is given to the builder.

import pycond as pc

cond = '[a eq b and [c lt 42 or foo eq bar]]'
cond = pc.to_struct(pc.tokenize(cond, sep=' ', brkts='[]'))
print(cond)
return cond

Output:

[['a', 'eq', 'b', 'and', ['c', 'lt', '42', 'or', 'foo', 'eq', 'bar']]]

Building

After parsing the builder is assembling a nested set of operator functions, combined via combining operators. The functions are partials, i.e. not yet evaluated but information about the necessary keys is already available:

f, meta = pc.parse_cond('foo eq bar')
assert meta['keys'] == ['foo']

Structured Conditions

Other processes may deliver condition structures via serializable formats (e.g. json). If you hand such already tokenized constructs to pycond, then the tokenizer is bypassed:

cond = [['a', 'eq', 'b'], 'or', ['c', 'in', ['foo', 'bar']]]
assert pc.pycond(cond)(state={'a': 'b'}) == True

Evaluation

The result of the builder is a 'pycondition', which can be run many times against a varying state of the system. How state is evaluated is customizable at build and run time.

Default Lookup

The default is to get lookup keys within expressions from an initially empty State dict within the module.

f = pc.pycond('foo eq bar')
assert f() == False
pc.State['foo'] = 'bar'
assert f() == True

(pycond is a shortcut for parse_cond, when meta infos are not required).

Passing Custom State

Use the state argument at evaluation:

assert pc.pycond('a gt 2')(state={'a': 42}) == True
assert pc.pycond('a gt 2')(state={'a': -2}) == False

Deep Lookup / Nested State

You may supply a path seperator for diving into nested structures like so:

m = {'a': {'b': [{'c': 1}]}}
assert (
    pc.pycond([('a', 'b', 0, 'c'), 'eq', 1], deep='.')(state=m)
    == True
)

assert pc.pycond('a.b.0.c', deep='.')(state=m) == True
assert pc.pycond('a.b.1.c', deep='.')(state=m) == False
assert pc.pycond('a.b.0.c eq 1', deep='.')(state=m) == True
# convencience argument for string conditions:
assert pc.pycond('deep: a.b.0.c')(state=m) == True

Custom Lookup And Value Passing

You can supply your own function for value acquisition.

  • Signature: See example.
  • Returns: The value for the key from the current state plus the compare value for the operator function.
# must return a (key, value) tuple:
model = {'eve': {'last_host': 'somehost'}}

def my_lu(k, v, req, user, model=model):
    print('user check. locals:', dict(locals()))
    return (model.get(user) or {}).get(k), req[v]

f = pc.pycond('last_host eq host', lookup=my_lu)

req = {'host': 'somehost'}
assert f(req=req, user='joe') == False
assert f(req=req, user='eve') == True

Output:

user check. locals: {'k': 'last_host', 'v': 'host', 'req': {'host': 'somehost'}, 'user': 'joe', 'model': {'eve': {'last_host': 'somehost'}}}
user check. locals: {'k': 'last_host', 'v': 'host', 'req': {'host': 'somehost'}, 'user': 'eve', 'model': {'eve': {'last_host': 'somehost'}}}

as you can see in the example, the state parameter is just a convention for pyconds' [title: default lookup function,fmatch=pycond.py,lmatch:def state_get function.

Lazy Evaluation

This is avoiding unnecessary calculations in many cases:

When an evaluation branch contains an "and" or "and_not" combinator, then at runtime we evaluate the first expression - and stop if it is already False. That way expensive deep branch evaluations are omitted or, when the lookup is done lazy, the values won't be even fetched:

evaluated = []

def myget(key, val, cfg, state=None, **kw):
    evaluated.append(key)
    # lets say we are false - always:
    return False, True

f = pc.pycond(
    '[a eq b] or foo eq bar and baz eq bar', lookup=myget
)
f()
# the value for "baz" is not even fetched and the whole (possibly
# deep) branch after the last and is ignored:
assert evaluated == ['a', 'foo']
print(evaluated)

Output:

['a', 'foo']

Building Conditions From Text

Condition functions are created internally from structured expressions - but those are hard to type, involving many apostropies.

The text based condition syntax is intended for situations when end users type them into text boxes directly.

Grammar

Combine atomic conditions with boolean operators and nesting brackets like:

[  <atom1> <and|or|and not|...> <atom2> ] <and|or...> [ [ <atom3> ....

Atomic Conditions

[not] <lookup_key> [ [rev] [not] <condition operator (co)> <value> ]
  • When just lookup_key is given, then co is set to the truthy function:
def truthy(key, val=None):
    return operatur.truth(k)

so such an expression is valid and True:

pc.State.update({'foo': 1, 'bar': 'a', 'baz': []})
assert pc.pycond('[ foo and bar and not baz]')() == True
  • When not lookup_key is given, then co is set to the falsy function:
m = {'x': 'y', 'falsy_val': {}}
# normal way
assert pc.pycond(['foo', 'eq', None])(state=m) == True
# using "not" as prefix:
assert pc.pycond('not foo')(state=m) == True
assert pc.pycond(['not', 'foo'])(state=m) == True
assert pc.pycond('not falsy_val')(state=m) == True
assert pc.pycond('x and not foo')(state=m) == True
assert pc.pycond('y and not falsy_val')(state=m) == False

Condition Operators

All boolean standardlib operators are available by default:

from pytest2md import html_table as tbl  # just a table gen.
from pycond import get_ops

for k in 'nr', 'str':
    s = 'Default supported ' + k + ' operators...(click to extend)'
    print(tbl(get_ops()[k], [k + ' operator', 'alias'], summary=s))
Default supported nr operators...(click to extend)
nr operatoralias
add+
and_&
eq==
floordiv//
ge>=
gt>
iadd+=
iand&=
ifloordiv//=
ilshift<<=
imod%=
imul*=
ior|=
ipow**=
irshift>>=
is_is
is_notis
isub-=
itruediv/=
ixor^=
le<=
lshift<<
lt<
mod%
mul*
ne!=
or_|
pow**
rshift>>
sub-
truediv/
xor^
itemgetter
length_hint
Default supported str operators...(click to extend)
str operatoralias
attrgetter
concat+
contains
countOf
iconcat+=
indexOf
methodcaller

Using Symbolic Operators

By default pycond uses text style operators.

  • ops_use_symbolic switches processwide to symbolic style only.
  • ops_use_symbolic_and_txt switches processwide to both notations allowed.
pc.ops_use_symbolic()
pc.State['foo'] = 'bar'
assert pc.pycond('foo == bar')() == True
try:
    # this raises now, text ops not known anymore:
    pc.pycond('foo eq bar')
except:
    pc.ops_use_symbolic_and_txt(allow_single_eq=True)
    assert pc.pycond('foo = bar')() == True
    assert pc.pycond('foo == bar')() == True
    assert pc.pycond('foo eq bar')() == True
    assert pc.pycond('foo != baz')() == True

Operator namespace(s) should be assigned at process start, they are global.

Extending Condition Operators

pc.OPS['maybe'] = lambda a, b: int(time.time()) % 2
# valid expression now:
assert pc.pycond('a maybe b')() in (True, False)

Negation not

Negates the result of the condition operator:

pc.State['foo'] = 'abc'
assert pc.pycond('foo eq abc')() == True
assert pc.pycond('foo not eq abc')() == False

Reversal rev

Reverses the arguments before calling the operator

pc.State['foo'] = 'abc'
assert pc.pycond('foo contains a')() == True
assert pc.pycond('foo rev contains abc')() == True

rev and not can be combined in any order.

Wrapping Condition Operators

Global Wrapping

You may globally wrap all evaluation time condition operations through a custom function:

l = []

def hk(f_op, a, b, l=l):
    l.append((getattr(f_op, '__name__', ''), a, b))
    return f_op(a, b)

pc.run_all_ops_thru(hk)  # globally wrap the operators

pc.State.update({'a': 1, 'b': 2, 'c': 3})
f = pc.pycond('a gt 0 and b lt 3 and not c gt 4')
assert l == []
f()
expected_log = [('gt', 1, 0.0), ('lt', 2, 3.0), ('gt', 3, 4.0)]
assert l == expected_log
pc.ops_use_symbolic_and_txt()

You may compose such wrappers via repeated application of the run_all_ops_thru API function.

Condition Local Wrapping

This is done through the ops_thru parameter as shown:

def myhk(f_op, a, b):
    return True

pc.State['a'] = 1
f = pc.pycond('a eq 2')
assert f() == False
f = pc.pycond('a eq 2', ops_thru=myhk)
assert f() == True

Using ops_thru is a good way to debug unexpected results, since you can add breakpoints or loggers there.

Combining Operations

You can combine single conditions with

  • and
  • and not
  • or
  • or not
  • xor by default.

The combining functions are stored in pycond.COMB_OPS dict and may be extended.

Do not use spaces for the names of combining operators. The user may use them but they are replaced at before tokenizing time, like and not -> and_not.

Nesting

Combined conditions may be arbitrarily nested using brackets "[" and "]".

Via the brkts config parameter you may change those to other separators at build time.

Tokenizing Details

Brackets as strings in this flat list form, e.g. ['[', 'a', 'and' 'b', ']'...]

Functioning

The tokenizers job is to take apart expression strings for the builder.

Separator sep

Separates the different parts of an expression. Default is ' '.

pc.State['a'] = 42
assert pc.pycond('a.eq.42', sep='.')() == True

sep can be a any single character including binary.

Bracket characters do not need to be separated, the tokenizer will do:

# equal:
assert (
    pc.pycond('[[a eq 42] and b]')()
    == pc.pycond('[ [ a eq 42 ] and b ]')()
)

The condition functions themselves do not evaluate equal - those had been assembled two times.

Apostrophes

By putting strings into Apostrophes you can tell the tokenizer to not further inspect them, e.g. for the seperator:

pc.State['a'] = 'Hello World'
assert pc.pycond('a eq "Hello World"')() == True

Escaping

Tell the tokenizer to not interpret the next character:

pc.State['b'] = 'Hello World'
assert pc.pycond('b eq Hello\ World')() == True

Building

Autoconv: Casting of values into python simple types

Expression string values are automatically cast into bools and numbers via the public pycond.py_type function.

This can be prevented by setting the autoconv parameter to False or by using Apostrophes:

pc.State['a'] = '42'
assert pc.pycond('a eq 42')() == False
# compared as string now
assert pc.pycond('a eq "42"')() == True
# compared as string now
assert pc.pycond('a eq 42', autoconv=False)() == True

If you do not want to provide a custom lookup function (where you can do what you want) but want to have looked up keys autoconverted then use:

for id in '1', 1:
    pc.State['id'] = id
    assert pc.pycond('id lt 42', autoconv_lookups=True)

Context On Demand And Lazy Evaluation

Often the conditions are in user space, applied on data streams under the developer's control only at development time.

The end user might pick only a few keys from many offered within an API.

pycond's ctx_builder allows to only calculate those keys at runtime, the user decided to base conditions upon: At condition build time hand over a namespace for all functions which are available to build the ctx.

pycon will return a context builder function for you, calling only those functions which the condition actually requires.

pc.ops_use_symbolic_and_txt(allow_single_eq=True)

# Condition the end user configured, e.g. at program run time:
cond = [
    ['group_type', 'in', ['lab', 'first1k', 'friendly', 'auto']],
    'and',
    [
        [
            [
                [
                    ['cur_q', '<', 0.5],
                    'and',
                    ['delta_q', '>=', 0.15],
                ],
                'and',
                ['dt_last_enforce', '>', 28800],
            ],
            'and',
            ['cur_hour', 'in', [3, 4, 5]],
        ],
        'or',
        [
            [
                [
                    ['cur_q', '<', 0.5],
                    'and',
                    ['delta_q', '>=', 0.15],
                ],
                'and',
                ['dt_last_enforce', '>', 28800],
            ],
            'and',
            ['clients', '=', 0],
        ],
    ],
]

# Getters for API keys offered to the user, involving potentially
# expensive to fetch context delivery functions:
# Signature must provide minimum a positional for the current
# state:
class ApiCtxFuncs:
    def expensive_but_not_needed_here(ctx):
        raise Exception("Won't run with cond. from above")

    def group_type(ctx):
        raise Exception(
            "Won't run since contained in example data"
        )

    def cur_q(ctx):
        print('Calculating cur_q')
        return 0.1

    def cur_hour(ctx):
        print('Calculating cur_hour')
        return 4

    def dt_last_enforce(ctx):
        print('Calculating dt_last_enforce')
        return 10000000

    def delta_q(ctx):
        print('Calculating (expensive) delta_q')
        time.sleep(0.1)
        return 1

    def clients(ctx):
        print('Calculating clients')
        return 0

if sys.version_info[0] < 3:
    # we don't think it is a good idea to make the getter API stateful:
    p2m.convert_to_staticmethods(ApiCtxFuncs)

f, nfos = pc.parse_cond(cond, ctx_provider=ApiCtxFuncs)
# this key stores the context builder function
make_ctx = nfos['complete_ctx']

# now we get (incomplete) data..
data1 = {'group_type': 'xxx'}, False
data2 = {'group_type': 'lab'}, True

t0 = time.time()
for event, expected in data1, data2:
    assert pc.pycond(cond)(state=make_ctx(event)) == expected

print('Calc.Time', round(time.time() - t0, 4))
return cond, ApiCtxFuncs

Output:

Calculating clients
Calculating cur_hour
Calculating cur_q
Calculating (expensive) delta_q
Calculating dt_last_enforce
Calculating clients
Calculating cur_hour
Calculating cur_q
Calculating (expensive) delta_q
Calculating dt_last_enforce
Calc.Time 0.206

But we can do better - we still calculated values for keys which might be only needed in dead ends of a lazily evaluated condition.

Lets avoid calculating these values, remembering the custom lookup function feature.

pycond does generate such a custom lookup function readily for you, if you pass a getter namespace as lookup_provider:

# we let pycond generate the lookup function now:
f = pc.pycond(cond, lookup_provider=ApiCtxFuncs)

# Same events as above:
data1 = {'group_type': 'xxx'}, False
data2 = {'group_type': 'lab'}, True

t0 = time.time()
for event, expected in data1, data2:
    # we will lookup only once:
    assert f(state=event) == expected

print(
    'Calc.Time (only one expensive calculation):',
    round(time.time() - t0, 4),
)

Output:

Calculating cur_q
Calculating (expensive) delta_q
Calculating dt_last_enforce
Calculating cur_hour
Calculating clients
Calc.Time (only one expensive calculation): 0.1005

The output demonstrates that we did not even call the value provider functions for the dead branches of the condition.

Auto generated by pytest2md, running test_tutorial.py

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pycond-190523.tar.gz (21.1 kB view hashes)

Uploaded Source

Built Distribution

pycond-190523-py3-none-any.whl (15.8 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page