My example shows when using DataFrame of float that querying might in certains cases be faster than using masks. When you look at the graph, the q̶u̶e̶r̶y̶ ̶f̶u̶n̶c̶t̶i̶o̶n̶ ̶p̶e̶r̶f̶o̶r̶m̶s̶ ̶b̶e̶t̶t̶e̶r̶ ̶w̶h̶e̶n̶ ̶t̶h̶e̶ ̶c̶o̶n̶d̶i̶t̶i̶o̶n̶ ̶i̶s̶ ̶c̶o̶m̶p̶o̶s̶e̶d̶ ̶o̶f̶ ̶1̶ ̶t̶o̶ ̶5̶ ̶s̶u̶b̶c̶o̶n̶d̶i̶t̶i̶o̶n̶s̶.
Edit (thanks to a_guest): mask function performs better when the condition is composed of 1 to 5 subconditions
Then, Is there any difference between the two methods since it tends to have the same trend over the number of subconditions.
The function used to plot my data:
import matplotlib.pyplot as plt
def graph(data):
t = [int(i) for i in range(1, len(data["mask"]) + 1)]
plt.xlabel('Number of conditions')
plt.ylabel('timeit (ms)')
plt.title('Benchmark mask vs query')
plt.grid(True)
plt.plot(t, data["mask"], 'r', label="mask")
plt.plot(t, data["query"], 'b', label="query")
plt.xlim(1, len(data["mask"]))
plt.legend()
plt.show()
The functions used to creates the conditions to be tested by timeit:
def create_multiple_conditions_mask(columns, nb_conditions, condition):
mask_list = []
for i in range(nb_conditions):
mask_list.append("(df['" + columns[i] + "']" + " " + condition + ")")
return " & ".join(mask_list)
def create_multiple_conditions_query(columns, nb_conditions, condition):
mask_list = []
for i in range(nb_conditions):
mask_list.append(columns[i] + " " + condition)
return "'" + " and ".join(mask_list) + "'"
The function to benchmark masking vs querying using a pandas DataFrame containing float:
def benchmarks_mask_vs_query(dim_df=(50,10), labels=[], condition="> 0", random=False):
# init local variable
time_results = {"mask": [], "query": []}
nb_samples, nb_columns = dim_df
all_labels = list('ABCDEFGHIJKLMNOPQRSTUVWXYZ')
if nb_columns > 26:
if len(labels) == nb_columns:
all_labels = labels
else:
raise Exception("labels length must match nb_columns" )
df = pd.DataFrame(np.random.randn(nb_samples, nb_columns), columns=all_labels[:nb_columns])
for col in range(nb_columns):
if random:
condition = "<" + str(np.random.random(1)[0])
mask = "df[" + create_multiple_conditions_mask(df.columns, col+1, condition) + "]"
query = "df.query(" + create_multiple_conditions_query(df.columns, col+1, condition) + ")"
print("Parameters: nb_conditions=" + str(col+1) + ", condition= " + condition)
print("Mask created: " + mask)
print("Query created: " + query)
print()
result_mask = timeit(mask, number=100, globals=locals()) * 10
result_query = timeit(query, number=100, globals=locals()) * 10
time_results["mask"].append(result_mask)
time_results["query"].append(result_query)
return time_results
What I run:
# benchmark on a DataFrame of shape(50,25) populating with random values
# as well as the conditions ("<random_value")
data = benchmarks_mask_vs_query((50,25), random=True)
graph(data)
What I get:
