Simple comparative benchmarks for CSV parsing libraries.
The "file" benchmarks are IO-based: reading from a file into a list or vector of rows which is then forced with a deepseq. Conveniently, the differences between each library are reproducibly distinct.
The script generates a file of 1,000 rows, of 20 fields with [a-zA-z0-9] and
[," \r\n\t], ensuring that the full range of CSV is tested:
stack exec generate > in.csv
| Case | Allocated | Max | Live | GCs |
|---|---|---|---|---|
| cassava/decode/Vector ByteString | 8,187,688 | 9,328 | 20,584 | 4 |
| cassava/decode/[ByteString] | 10,424,808 | 9,256 | 20,392 | 6 |
| sv/Data.Sv/parseDecodeFromFile | 13,893,912 | 1,191,464 | 1,256,184 | 10 |
| lazy-csv/parseCsv/[ByteString] | 17,851,008 | 1,121,544 | 1,185,336 | 16 |
| csv-conduit/readCSVFile/[ByteString] | 27,012,448 | 1,204,504 | 1,244,456 | 25 |
| csv-conduit/readCSVFile/Vector ByteString | 27,973,968 | 1,231,656 | 1,303,896 | 26 |
| csv-conduit/readCSVFile/[String] | 34,319,272 | 2,828,712 | 4,134,544 | 32 |
| csv/Text.CSV/parseCSVFromFile | 141,654,888 | 3,376,504 | 8,840,696 | 140 |
| Name | |
|---|---|
| cassava/decode/Vector ByteString | 1.666 ms |
| cassava/decode/[ByteString] | 2.191 ms |
| lazy-csv/parseCsv/[ByteString] | 3.034 ms |
| sv/Data.Sv/parseDecodeFromFile | 3.838 ms |
| csv-conduit/readCSVFile/[ByteString] | 8.016 ms |
| csv-conduit/readCSVFile/Vector ByteString | 8.482 ms |
| csv-conduit/readCSVFile/[String] | 12.47 ms |
| csv/Text.CSV/parseCSVFromFile | 29.56 ms |