Where a naive split goes wrong
The obvious approach is to open the file in a text editor, count ten thousand lines, cut, paste into a new file, repeat. Three things break.
The second file has no header row, so the import tool reads the first data row as column names and either rejects the file or imports ninety-nine hundred rows with the first customer's email as a column heading. Somewhere around line 6,412 a field contains a line break inside quotes, an address with a second line, and the cut landed in the middle of it, so one record is split across two files and both are malformed. And the editor saved the new file in a different encoding from the original, so every accented name in the second half arrives as a pair of wrong characters.
A split done by a tool that understands CSV avoids all three. The CSV splitter reads the file by records rather than by lines, copies the header onto every part, and writes each part in the same encoding it read.
Rows are not lines
RFC 4180 allows a field to contain commas, double quotes and line breaks, provided the field is wrapped in double quotes and any quote inside it is doubled. A record can therefore span several physical lines. A tool that counts newlines to find record boundaries will miscount on any file with multi-line addresses or notes fields, and the miscount is silent: the row count looks plausible and one record in a thousand is corrupt.
The row count the splitter reports before cutting is the test. If it matches what the source system said it exported, the parser handled the quoting. If it is higher, something is being split that should not be, and the file deserves a look in the CSV viewer before anything is uploaded.
The header on every part
A chunk of data with no column names is a file nobody downstream can use, and most import tools map columns by header rather than by position. Every part must begin with the same header row as the original. The splitter does this by default. If a part is later edited by hand, check that the header survived the edit, because a deleted first row is the most common reason a second batch maps its columns wrongly.
Set the rows-per-part slightly under the import limit rather than exactly at it. Some tools count the header as a row; some count the trailing newline as one. Nine thousand nine hundred is safer than ten thousand.
Encoding, the invisible problem
A CSV is bytes, and the bytes for an accented character differ between UTF-8 and the legacy Windows encodings. Excel on Windows has historically written and read the legacy encoding unless told otherwise, which is why a UTF-8 file with a name like Müller opens in Excel as something unprintable, and why a file saved from Excel can confuse a UTF-8 import.
The fix in both directions is to be explicit. Export from Excel using the CSV UTF-8 option rather than plain CSV, or run the workbook through the Excel to CSV converter, which writes UTF-8. When importing a CSV into Excel, use the data import dialog and choose the encoding rather than double-clicking the file. A split tool should not change encoding at all, and should not add or remove the byte-order mark at the start of the file, because some importers treat it as part of the first column name.
Before splitting: deduplicate and check
A ninety-thousand-row export frequently contains the same record several times, from repeated exports appended together or a join that multiplied rows. The duplicate line remover collapses exact duplicates before the split, which can remove a part or two entirely and keeps the import tool from rejecting the batch for duplicate keys.
If the destination wants JSON rather than CSV, an API or a document store, the CSV to JSON converter produces it from each part, and the same header-on-every-part rule is what makes the keys consistent across the batch. Open one part after splitting, confirm the columns line up and an accented name reads correctly, and only then upload the rest.
Questions people ask
Use a tool that parses CSV records rather than counting lines, so quoted fields with line breaks stay intact. Choose rows per part or number of parts, keep the header row on every part, and check that the reported row count matches the source export before cutting.
Yes. Import tools map columns by their names, and a part without a header either fails or treats the first data row as column names. A CSV splitter copies the header onto every part automatically; a manual cut in a text editor does not.
The file was written in one encoding and read in another, usually UTF-8 versus a legacy Windows encoding. Export from Excel with the CSV UTF-8 option, and when opening a CSV in Excel use the import dialog to pick the encoding rather than double-clicking.
Slightly under it, such as 9,900. Some tools count the header row toward the limit and some count a trailing blank line. A small margin avoids the last part being rejected for being one row over.
If it is higher, a field containing a line break was probably split in two by a line-based parser. If it is lower, the file may have been truncated when it was downloaded or saved. Open it in a CSV viewer and look at the rows around the discrepancy before uploading.
Not with a tool that runs on the device. The file is read, parsed and written back as parts locally, which matters for customer lists and financial exports. The practical limit is the browser's memory, which handles files of hundreds of megabytes on an ordinary laptop.

