Introduction

Handling large datasets is a common challenge in data analysis and processing tasks. In this blog post, we will discuss our findings on comparing the performance of five different frameworks for concatenating CSV files into dataframes and analyzing their CPU consumption. The frameworks we evaluated are go-dataframe (Go), gota (Go), pandas (Python), petl (Python), and pyspark (Python).

Task Overview

The task at hand involves three primary steps

  1. Reading multiple CSV files and converting them into dataframes
  2. Concatenating these dataframes into a single dataframe
  3. Writing the consolidated data back to an output file

To compare the performance of these frameworks, we used benchmarking tools like Hyperfine and CMDBench. Hyperfine allowed us to measure the execution time for each framework, while CMDBench provided insights into the CPU and memory consumption during the task execution.

CSV File Description

Before diving into the comparison, let’s take a brief look at the structure of the CSV files we used for our analysis. These files contain records with columns such as first_name, last_name, email, ssn, job, country, phone_number, user_name, zipcode, invalid_ssn, credit_card_number, credit_card_provider, credit_card_security_code, and bban.

CPU Consumption Analysis

We evaluated the CPU consumption of each framework for different scenarios involving varying numbers of CSV files with 100,000 records each. Below are the summarized results:

2 Small Files

(10 runs 10 records each)
Framework Mean [s] Min [s] Max [s] Relative
go-dataframe 1.148 ± 0.081 1.070 1.288 7.74 ± 0.62
gota 0.752 ± 0.096 0.690 1.021 5.07 ± 0.68
pandas 0.485 ± 0.024 0.456 0.515 3.27 ± 0.20
petl 0.148 ± 0.006 0.137 0.156 1.00
pyspark 13.741 ± 2.424 11.671 20.077 92.61 ± 16.70

2 Files

(10 runs 100,000 records each)
Framework Mean [s] Min [s] Max [s] Relative
go-dataframe 1.689 ± 0.089 1.579 1.874 1.00
gota 2.282 ± 0.130 2.124 2.507 1.35 ± 0.10
pandas 2.816 ± 0.100 2.693 2.997 1.67 ± 0.11
petl 2.294 ± 0.049 2.257 2.404 1.36 ± 0.08
pyspark 14.142 ± 0.643 13.850 15.927 8.37 ± 0.58

5 Files

(10 runs 100,000 records each)
Framework Mean [s] Min [s] Max [s] Relative
go-dataframe 2.583 ± 0.204 2.349 3.086 1.00
gota 4.884 ± 0.316 4.462 5.374 1.89 ± 0.19
pandas 6.527 ± 0.349 6.209 7.235 2.53 ± 0.24
petl 8.117 ± 0.616 7.290 9.225 3.14 ± 0.34
pyspark 21.108 ± 2.345 18.462 26.281 8.17 ± 1.11

10 Files

(10 runs 100,000 records each)
Framework Mean [s] Min [s] Max [s] Relative
go-dataframe 3.847 ± 0.253 3.585 4.280 1.00
gota 9.127 ± 0.231 8.828 9.450 2.37 ± 0.17
pandas 14.185 ± 1.546 12.161 15.737 3.69 ± 0.47
petl 24.496 ± 3.328 21.238 30.702 6.37 ± 0.96
pyspark 23.638 ± 0.915 22.122 24.999 6.14 ± 0.47

25 Files

(10 runs 100,000 records each)
Framework Mean [s] Min [s] Max [s] Relative
go-dataframe 8.212 ± 0.559 7.521 9.105 1.00
gota 32.663 ± 1.640 30.131 35.154 3.98 ± 0.34
pandas 30.790 ± 0.572 30.060 31.684 3.75 ± 0.26
petl 107.274 ± 3.292 100.848 111.170 13.06 ± 0.98
pyspark 35.659 ± 0.719 34.524 37.100 4.34 ± 0.31

50 Files

(5 runs 100,000 records each)
Framework Mean [s] Min [s] Max [s] Relative
go-dataframe 16.669 ± 1.990 14.858 19.575 1.00
gota 91.149 ± 4.158 88.691 98.500 5.47 ± 0.70
pandas 60.217 ± 2.036 58.558 62.916 3.61 ± 0.45
pyspark 65.071 ± 1.078 64.086 66.612 3.90 ± 0.47

100 Files

(5 runs 100,000 records each)
Framework Mean [s] Min [s] Max [s] Relative
go-dataframe 26.075 ± 0.522 25.426 26.518 1.00
pandas 116.856 ± 0.334 116.315 117.220 4.48 ± 0.09

Memory Consumption Analysis

For the memory consumption analysis, we created graphs using CMDBench. Here are the key insights

2 Files

(100,000 records each)

go-dataframe gota pandas petl pyspark
go-dataframe gota pandas petl pyspark

5 Files

(100,000 records each)

go-dataframe gota pandas petl pyspark
go-dataframe gota pandas petl pyspark

10 Files

(100,000 records each)

go-dataframe gota pandas petl pyspark
go-dataframe gota pandas petl pyspark

25 Files

(100,000 records each)

go-dataframe gota pandas petl pyspark
go-dataframe gota pandas petl pyspark

50 Files

(100,000 records each)

go-dataframe gota pandas petl pyspark
go-dataframe gota pandas petl pyspark

100 Files

(100,000 records each)

go-dataframe pandas
go-dataframe pandas

Conclusion

From our analysis, we can draw several conclusions

  1. For small datasets, petl outperforms all other frameworks in terms of CPU consumption, closely followed by pandas.
  2. As the dataset size increases, pyspark consistently shows the highest CPU consumption, making it less suitable for large datasets in a single-node setup.
  3. Overall, go-dataframe outperforms all other frameworks in terms of CPU and memory consumption

Choosing the right framework depends on the specific requirements of your task, including dataset size, available resources, and the desired level of parallelization. We hope this analysis helps you make an informed decision when working with CSV concatenation and dataframe comparison tasks. Happy data processing!