# Working with Large Data Sets¶

Big data is something of a buzzword in the modern world. While statsmodels works well with small and moderately-sized data sets that can be loaded in memory–perhaps tens of thousands of observations–use cases exist with millions of observations or more. Depending your use case, statsmodels may or may not be a sufficient tool.

statsmodels and most of the software stack it is written on operates in memory. Resultantly, building models on larger data sets can be challenging or even impractical. With that said, there are 2 general strategies for building models on larger data sets with statsmodels.

## Divide and Conquer - Distributing Jobs¶

If your system is capable of loading all the data, but the analysis you are attempting to perform is slow, you might be able to build models on horizontal slices of the data and then aggregate the individual models once fit.

A current limitation of this approach is that it generally does not support patsy so constructing your design matrix (known as exog) in statsmodels, is a little challenging.

A detailed example is available here.

 DistributedModel(partitions[, model_class, …]) Distributed model class DistributedResults(model, params) Class to contain model results

If your entire data set is too large to store in memory, you might try storing it in a columnar container like Apache Parquet or bcolz. Using the patsy formula interface, statsmodels will use the __getitem__ function (i.e. data[‘Item’]) to pull only the specified columns.

import pyarrow as pa
import pyarrow.parquet as pq
import statsmodels.formula.api as smf

class DataSet(dict):
def __init__(self, path):
self.parquet = pq.ParquetFile(path)

def __getitem__(self, key):
try: