It's especially weird in this case because the value-proposition of pandas is the ability to do things like aggregation in 3 lines of code and not have to iterate through every row.
The trouble with this code was the choice of the quadratic algorithm (which the idiomatic pandas version avoids). You could achieve a similar speed-up as OP for the pure Python code too, by dropping the doubly-nested loop.
The point stands as usual: better algo > clever tricks :) Personally, I find the code brevity that pandas offers does not outweigh the added complexity and cognitive load of its (IMO hard to remember and reason about) API.
EDIT: pure Python version
def gen_stats_python2(dataset_python):
aggr = {}
for product_id, product_order_num, quantity, price in dataset_python:
item = aggr.get(product_id, [0, []])
item[0] += quantity
item[1].append(price)
aggr[product_id] = item
return [
(int(product_id), len(prices), int(quantity), round(float(sum(prices)) / len(prices), 2))
for product_id, (quantity, prices) in aggr.items()
]
%timeit gen_stats_python(dataset_python)
1 loops, best of 3: 36.3 s per loop
%timeit gen_stats_python2(dataset_python)
10 loops, best of 3: 47.7 ms per loop def gen_stats(data):
start = time.time()
stats = []
data.sort(key=itemgetter(0))
for p_id, p_items in itertools.groupby(data, key=itemgetter(0)):
num_orders, _, total_quantity, total_price = map(sum, zip(*p_items))
p_id, num_orders, total_quantity = map(int, [p_id, num_orders/p_id, total_quantity])
stats.append([p_id, num_orders, total_quantity, round(float(total_price/num_orders),2)])
end = time.time()
working_time = end-start
return stats, working_time
Dataset size 55096 records
[[1, 8, 340, 22.55], [2, 4, 94, 15.83], [3, 2, 97, 3.89]]
[[1, 8, 340, 22.55], [2, 4, 94, 15.83], [3, 2, 97, 3.89]]
All results equal: True
Runtime (s)
gen_stats_slow: 28.489468574523926
gen_stats_fast: 0.06000018119812012Additionally, you might have a bug in the code: groupby() only groups consecutive keys, so your code will fail if the dataset is not already sorted by product_id.
I guess that's the lesson here -- use pandas to make sure you get a sane, fast, tested implementation of some common use-cases. That is, if you can remember its API, when it makes copies vs views under the hood, and all that gnarly stuff :-)
Looks like my gen_stats_eesmith is even faster. Radim's is about 400x faster, iClaudiusX is 450x faster, and eesmith is 580x faster.
Under pypy2-v5.8.0 those are 65x, 16x, and 91x, respectively. The original algorithm is 10x faster under pypy.
def gen_stats_python(dataset_python):
start = time.time()
product_stats = []
unique_products = set([x[0] for x in dataset_python])
for product_id in unique_products:
product_items = [x for x in dataset_python if x[0]==product_id ]
num_orders = len(product_items)
total_quantity = 0
total_price = 0
for row in product_items:
total_quantity += row[2]
total_price += row[3]
avg_price = float(total_price/num_orders)
product_stats.append([int(product_id),int(num_orders),int(total_quantity),round(avg_price,2)])
end = time.time()
working_time = end-start
return product_stats,working_time
def gen_stats_radim(dataset_python):
start = time.time()
aggr = {}
for product_id, product_order_num, quantity, price in dataset_python:
item = aggr.get(product_id, [0, []])
item[0] += quantity
item[1].append(price)
aggr[product_id] = item
return [
[int(product_id), len(prices), int(quantity), round(float(sum(prices)) / len(prices), 2)]
for product_id, (quantity, prices) in aggr.items()
], time.time()-start
def gen_stats_iclaudiusx(data):
start = time.time()
stats = []
data.sort(key=itemgetter(0))
for p_id, p_items in itertools.groupby(data, key=itemgetter(0)):
num_orders, _, total_quantity, total_price = map(sum, zip(*p_items))
p_id, num_orders, total_quantity = map(int, [p_id, num_orders/p_id, total_quantity])
stats.append([p_id, num_orders, total_quantity, round(float(total_price/num_orders),2)])
end = time.time()
working_time = end-start
return stats, working_time
def gen_stats_eesmith(dataset_python):
start = time.time()
product_stats = []
total_num_orders = defaultdict(int)
total_quantity_by_product_id = defaultdict(int)
total_price_by_product_id = defaultdict(float)
for product_id, order_id, quantity, price in dataset_python:
total_num_orders[product_id] += 1
total_quantity_by_product_id[product_id] += quantity
total_price_by_product_id[product_id] += price
for product_id in total_price_by_product_id:
num_orders = total_num_orders[product_id]
total_quantity = total_quantity_by_product_id[product_id]
total_price = total_price_by_product_id[product_id]
avg_price = float(total_price)/num_orders
product_stats.append([int(product_id),int(num_orders),int(total_quantity),round(avg_price,2)])
end = time.time()
working_time = end-start
return product_stats,working_time
reference_result = reference_dt = None
gen_stats_functions = (gen_stats_python, gen_stats_radim,
gen_stats_iclaudiusx, gen_stats_eesmith)
for gen_stats_function in gen_stats_functions:
result, dt = gen_stats_function(dataset_python)
if reference_result is None:
reference_result = result
reference_dt = dt
assert result == reference_result, (gen_stats_function.__name__,
result[:2], reference_result[:2])
print(gen_stats_function.__name__, dt, reference_dt/dt) def gen_stats_tech2(dataset_python):
start = time.time()
product_stats = []
# totals = (num_orders, quantity, price)
totals = defaultdict(lambda: (0, 0, 0.0))
for product_id, order_id, quantity, price in dataset_python:
o_count, o_quant, o_price = totals[product_id]
totals[product_id] = (o_count + 1, o_quant + quantity, o_price + price)
for product_id, product_info in totals.items():
num_orders, total_quantity, total_price = product_info
avg_price = total_price / num_orders
product_stats.append(
[product_id, num_orders, int(total_quantity),
round(avg_price, 2)]
)
end = time.time()
working_time = end - start
return product_stats, working_time def gen_stats_tech2(dataset_python):
start = time.time()
totals = defaultdict(lambda: (0, 0, 0.0))
for product_id, order_id, quantity, price in dataset_python:
o_count, o_quant, o_price = totals[product_id]
totals[product_id] = (o_count + 1, o_quant + quantity, o_price + price)
product_stats = [
[product_id, num_orders, total_quantity,
round(total_price / num_orders, 2)]
for product_id, (num_orders, total_quantity, total_price)
in totals.items()
]
end = time.time()
working_time = end - start
return product_stats, working_time
Most improvements are from use of some of the builtins like dictionary's .items method rather than making a getitem call per loop, removal of references to globals (int) since the values stored already have an integer value, removing N calls to append and letting the list-comp manage, and removing a store-then-unpack.
Sadly I don't think the top loop can be reduced in a similar manner because it's self-referential. Still, that made for a fun hour or so, I was really hoping to hit 1,000x improvement :/
Interestingly, the O(n) Python is still about 4x slower than the O(n) Pandas -- the same relative difference as in the original O(n^2) implementations from the blog post.
So despite the mishaps of this "Senior IT Business Analyst" choosing an unsuitable algorithm, his relative graphs pretty much stand.
This comparison was never about concrete (optimal) implementations but about comparing exactly the same implementations, doing exactly the same operations. Such approach would be the same for comparing performance of any languages, like C and Java. This is not "the best algo" race.
So, after a day of discussion, the result is the following: regardless if the implementation is quadratic or linear, the relative speed of all 3 technologies is (roughly) the same. I could not agree more as this is primary school math: dividing by the same denominator still means that proportions are the same :)
I plan to update my blog post, based on many interesting comments you have all provided. I definetely think this analysis should be more complex. Thank you for your insights.
I have many great pieces of code for inspiration, but I am struggling with providing exactly the same implementations. So it may cost me some time to provide all 3 optimized implementations that are exactly relevant to each other.
I'll be a bit pedantic. The original essay says "pure Python", which is not a technology but a language. There are several implementations of Python, which I'll argue is the "technology". The PyPy implementation of the pure Python algorithm is 10x faster than the CPython implementation.
According to the original essay, that's close to the CPython/NumPy performance, and faster than the CPython/Pandas version.
It's tempting to omit PyPy because "nobody uses it in data science". That's a self-fulfilling argument.
PyPy is doing amazing work to support both NumPy and Pandas, but it's limited by funding. Why aren't data science people pumping money into PyPy? It would give a huge performance boost even for naive, throw-it-together algorithms.
If you don't know why PyPy can do for you, how do you learn if/when you should use it?
My point is a minor one. "Python" isn't a technology. "CPython" and "PyPy" are technologies. Or you can try MicroPython, Iron Python, Jython, and others.
In any case, you'll see the best idiomatic Pandas code is reported to be 500x faster than the original, while the best idiomatic Python is reported to be 950x faster. The proportions are not the same, so it's not like "dividing by the same denominator".
The rank order is unchanged, though that's a weaker observation than what you wrote.
But this was really not your best piece (technically; it may have met its marketing / traffic goals).
The power of pandas comes - partly - from vectorized operations. This means that instead of iterating through every row it processes the data in data sets transforming the operations into matrix/vector manipulation. For matrix manipulation it uses low lewel high performat packages borrowed from c, numpy, etc.
Writing "pandas code" not using the pandas operations is like eating soup with a fork. It works but it is pretty slow.
I'm not sure what you imagine happens inside pandas, but I assure you it's not magic.