How to Split a Large CSV File

For a CSV up to 50 MiB, LocaCSV can split by column value, a fixed number of data rows, or a target file size entirely in your browser, then package up to 100 outputs in a ZIP archive.

For inputs over 50 MiB, compound keys, or scheduled workflows, a local Python script offers more control. Opening a very large CSV directly in Excel and cutting it apart manually is usually the least reliable option because worksheet limits, memory pressure, and copy-and-paste mistakes all become more likely as the file grows.

CSV is plain text, but the program opening it still has limits. An Excel worksheet has 1,048,576 rows. A file beyond that cannot fit on one sheet. Even below the row limit, imports can become slow because of wide records, long cells, type detection, quoting, available memory, or an incorrect delimiter.

• The file has more rows than an Excel worksheet can hold.
• The application expands the file in memory and the computer runs out of available resources.
• The delimiter is detected incorrectly, turning one record into an extreme number of columns.
• A simple line or comma split breaks quoted commas, escaped quotes, or line breaks inside a cell.
• The selected encoding does not match UTF-8, Windows-1252, or another legacy encoding used by the source system.
• Excel automatically converts IDs, leading zeros, long numbers, or date-like text.

The right tool depends on the split rule, file size, frequency, and whether you can maintain code. The table below summarizes the practical tradeoffs.

import csv
from pathlib import Path

source = Path("large.csv")
output_dir = Path("split")
output_dir.mkdir(exist_ok=True)

handles = {}
writers = {}

try:
    with source.open("r", encoding="utf-8-sig", newline="") as input_file:
        reader = csv.DictReader(input_file)
        key_column = "department"

        for row in reader:
            key = row.get(key_column, "") or "(blank)"
            if key not in writers:
                output = output_dir / f"group-{len(writers) + 1:03d}.csv"
                handle = output.open("w", encoding="utf-8-sig", newline="")
                writer = csv.DictWriter(handle, fieldnames=reader.fieldnames)
                writer.writeheader()
                handles[key] = handle
                writers[key] = writer
            writers[key].writerow(row)
finally:
    for handle in handles.values():
        handle.close()

LocaCSV parses commas, double quotes, escaped quotes, and line breaks inside cells as CSV structure. The preview is your best early warning: if columns do not line up there, correct the delimiter or encoding before running the split.

1. 1Open the CSV Splitter and select one CSV file.
2. 2Confirm whether the file has a header, then check the input encoding and delimiter. Do not continue if the preview is misaligned.
3. 3Choose Column value, Row count, or Target size as the split method.
4. 4For value mode, set the column and blank handling; for the other modes, enter data rows per file or a target size in MiB.
5. 5Confirm the output encoding, UTF-8 BOM option, and output delimiter, then start the split.
6. 6Review input rows, output and excluded rows, output-file count, and warnings before downloading the ZIP.

Value-based splitting creates a separate file for each category in one column—for example store, department, customer, or month. If a Store column contains London, Madrid, and Seoul, every row with the same value goes to the same output, producing three CSV files.

If headers are duplicated, absent, or padded with spaces, use the preview and one-based column number to confirm the target. When a value contains characters that Windows cannot use in a filename, or matches a Windows reserved name, LocaCSV converts the output name to a safe form.

• Rows with the same value stay in the same output file.
• Blank values can be excluded or written to a dedicated output.
• One column can be selected per run.
• If more than 100 unique outputs would be created, processing stops and no partial ZIP is published.
• Before exposing the ZIP, the tool checks that input rows equal output rows plus excluded rows.

File size alone does not predict how difficult a CSV is to process. Two 50 MiB files can behave differently when one has millions of short records and the other contains very long cells or many quoted line breaks.

The public LocaCSV Free limit for a single-input tool is one file up to 50 MiB. Device performance, browser behavior, CSV shape, and output volume can make a smaller file difficult as well. Renaming a larger file or passing a compressed archive does not bypass the limit.

• Row-count mode excludes the header from the number of data rows per file.
• Target-size mode uses actual bytes after output encoding, including the BOM and repeated header. A single oversized CSV record stays intact in its own file.
• There can be at most 100 output files; estimate unique values or row-based output count first.
• Keep the source unchanged, save the ZIP separately, and verify counts before replacing any downstream data.
• If the outputs are intended for Excel, check that each remains below 1,048,576 rows.

UTF-8 is common, while files from older Windows or Western European systems may use Windows-1252. LocaCSV auto-detects UTF-8 and Shift_JIS as candidates; CP949, Big5, GB18030, and Windows-1252 are manual choices. When detection is uncertain, use the preview rather than treating the guess as final.

For Excel, UTF-8 with a BOM is sometimes recognized more reliably. If a downstream system requires a legacy encoding, choose it only when every character can be represented. LocaCSV stops instead of silently replacing an unsupported character with a question mark.

• Check names, punctuation, currency symbols, and accented letters in the preview.
• Do not confuse a delimiter error with an encoding error: shifted columns usually point to comma, semicolon, tab, or pipe selection.
• If the source encoding is unknown, check the export settings of the system that created the file.
• Do not save a garbled preview and overwrite the source; start again from the original file.

LocaCSV runs CSV processing in a Web Worker inside your browser. The CSV body, filename, headers, cell values, and split values are not uploaded to an external server for processing. Results are assembled locally, and retained references are released after download or reset.

The public information site uses Cloudflare Web Analytics for page views. The processing site does not load advertising or analytics SDKs. It sends only identifier-free, first-party usage events limited to fields such as the operation, size bucket, duration bucket, and fixed error code. Events do not contain CSV contents, filenames, exact sizes, or row counts, and a failed event does not block processing or download.

For confidential data, also consider your organization’s policy, device security, browser extensions, and permissions on the download folder. Browser-local processing does not make the entire device automatically secure.