openrsxl

openrsxl is a drop-in replacement for openpyxl 3.1.5 whose performance-critical parts are written in Rust.

• Same API - every module, class, function, signature, default, constant, exception and warning of openpyxl exists under the same name in openrsxl.
• Same results - the same values with the same Python types (an empty cell is None), the same errors for broken files, and byte-for-byte the same files when saving.
• Faster and lighter - load + save is 7-12x faster with 6-17x less memory, read-only streaming 4-13x faster (benchmarks).
• openrsxl.extended - openpyxl's API plus things openpyxl cannot do: streaming (read-only mode) access to formulas and their cached values, comments, hyperlinks, merged cells, dimensions, sheet properties, data validation, conditional formatting, tables and full mode's cells at empty positions - each behind its own flag.

Installation

pip install openrsxl

Pre-built wheels for CPython 3.10 - 3.14 (including free-threaded 3.14t) on Linux (glibc: x86_64, aarch64, i686, armv7, ppc64le, s390x; musl: x86_64, aarch64, armv7), macOS (x86_64, arm64) and Windows (x64, x86, arm64); no Rust toolchain needed. The only runtime dependency is et_xmlfile, like openpyxl. lxml, defusedxml, pillow, numpy and pandas are optional exactly as for openpyxl.

Quick start

Replace openpyxl by openrsxl in your imports - nothing else changes:

import openrsxl                                  # instead of: import openpyxl
from openrsxl.styles import Font, PatternFill    # instead of: from openpyxl.styles import ...

# read
wb = openrsxl.load_workbook("report.xlsx")
ws = wb["Sales"]
print(ws["A1"].value)           # same value and Python type as openpyxl
print(ws["Z999"].value)         # None for an empty cell, as in openpyxl
for row in ws.iter_rows(min_row=2, values_only=True):
    print(row)

# write
ws["B2"] = 42
ws["B2"].font = Font(bold=True)
ws["C2"] = "=B2*2"
ws.merge_cells("D2:E3")
wb.save("report-out.xlsx")      # byte-identical to what openpyxl writes

# stream a huge file with constant memory (openpyxl's read-only mode)
wb = openrsxl.load_workbook("huge.xlsx", read_only=True)
total = sum(row[3] or 0 for row in wb.active.iter_rows(min_row=2, values_only=True))
wb.close()

# write a huge file (openpyxl's write-only mode)
wb = openrsxl.Workbook(write_only=True)
ws = wb.create_sheet("log")
for i in range(1_000_000):
    ws.append([i, f"event {i}", i * 0.5])
wb.save("log.xlsx")

Existing code using openpyxl can also keep its imports and alias the module once (import openrsxl as openpyxl), or switch a whole application - including libraries that import openpyxl themselves, such as pandas' read_excel(engine="openpyxl") - with one call at start-up, before anything imports openpyxl:

import openrsxl.extended
openrsxl.extended.install_as_openpyxl()   # `import openpyxl` now imports openrsxl

from openpyxl.worksheet.formula import ArrayFormula   # openrsxl's class

Then isinstance checks written for openpyxl's classes hold for openrsxl's objects (without it, the classes of the two libraries are distinct, see below). Supported formats are openpyxl's: .xlsx, .xlsm, .xltx, .xltm.

See examples/quickstart.py for a complete, runnable example.

openrsxl.extended: streaming with full-mode features

openpyxl offers two ways to read a file:

• full mode (load_workbook(path)) gives everything (formulas, comments, hyperlinks, merged cells, ...) but keeps every cell in memory - a 100 MB file can take minutes and gigabytes;
• read-only mode (read_only=True) streams rows with constant memory but drops everything except values and styles.

openrsxl.extended is the same API as openrsxl (openrsxl.extended.styles, openrsxl.extended.load_workbook, ... are the same objects) whose load_workbook accepts extra keyword-only flags that add full-mode features to read-only streaming. Rows are still streamed, so memory does not grow with the number of cells; each feature is only paid for when enabled, and without any flag the result is exactly openpyxl's read-only mode.

from openrsxl.extended import load_workbook

wb = load_workbook("book.xlsx", read_only=True,
                   formula_and_value=True,   # formula and saved result in one pass
                   read_comments=True, read_hyperlinks=True, read_merged_cells=True)
ws = wb["Sheet1"]
for row in ws.iter_rows():
    for cell in row:
        cell.value            # follows data_only, exactly as in openpyxl
        cell.formula          # "=SUM(A1:A9)" (or None)
        cell.cached_value     # the result saved in the file, e.g. 45
        cell.hyperlink        # Hyperlink or None
        cell.comment          # Comment or None
print(ws.merged_cells.ranges)
wb.close()

The flags

All flags are keyword-only, default to False and need read_only=True (full mode already reads everything; formula_and_value raises ValueError without read_only=True, the other flags are ignored).

flag	what it adds (names and types of openpyxl's full mode)
formula_and_value	cell.formula: the formula as openpyxl reads it with data_only=False - a str starting with =, an ArrayFormula or a DataTableFormula - or None for cells without formula; cell.cached_value: the value openpyxl reads with data_only=True (the result Excel saved). Both in one pass, where openpyxl needs two full loads.
read_comments	cell.comment (Comment or None).
read_hyperlinks	cell.hyperlink (Hyperlink or None); as in full mode, an empty cell with a link gets the link target as its value.
read_merged_cells	ws.merged_cells (the ranges); cells covered by a range are MergedCell objects (value None); the top-left cell gets the borders of the range; range.start_cell is that cell.
read_dimensions	ws.row_dimensions and ws.column_dimensions (heights, widths, hidden, outline levels, styles).
read_sheet_properties	ws.sheet_properties, ws.views, ws.freeze_panes, ws.sheet_format, ws.print_options, ws.page_margins, ws.page_setup, ws.HeaderFooter, ws.auto_filter, ws.protection, ws.row_breaks, ws.col_breaks, ws.scenarios, ws.print_title_rows / print_title_cols / print_area.
read_data_validations	ws.data_validations.
read_conditional_formatting	ws.conditional_formatting.
read_tables	ws.tables.
create_empty_cells	a cell at every empty position, like the one full mode creates when the position is accessed: coordinate, row, column, value None, default style (and None for the enabled cell features), instead of read-only mode's shared, coordinate-less EmptyCell. Created on the fly (nothing is kept); rows keep read-only mode's shape.

Every value is what openpyxl's full mode returns for the same file - same attribute names, same types, same values. The only new names are cell.formula and cell.cached_value, because openpyxl has no attribute for them in any mode (it exposes the formula or the cached value as cell.value, depending on data_only).

Examples

Formulas and their results in one pass - cell.value keeps following data_only:

from openrsxl.extended import load_workbook

wb = load_workbook("book.xlsx", read_only=True, formula_and_value=True)
for row in wb.active.iter_rows(min_row=2, max_col=4):
    for cell in row:
        if cell.formula is not None:
            print(cell.coordinate, cell.formula, "=", cell.cached_value)
# with data_only=True, cell.value is the cached value (as in openpyxl) and
# cell.formula still gives the formula

Comments and hyperlinks:

from openrsxl.extended import load_workbook

wb = load_workbook("book.xlsx", read_only=True, read_comments=True, read_hyperlinks=True)
for row in wb.active.iter_rows():
    for cell in row:
        if cell.comment:                         # None if absent (empty and merged cells too)
            print(cell.coordinate, cell.comment.author, cell.comment.text)
        if cell.hyperlink:
            print(cell.coordinate, cell.hyperlink.target or cell.hyperlink.location)

Every position as a cell, as in full mode - read-only mode yields one shared EmptyCell without coordinate for every empty position:

from openrsxl.extended import load_workbook

wb = load_workbook("book.xlsx", read_only=True, create_empty_cells=True)
for row in wb.active.iter_rows():
    for cell in row:
        print(cell.coordinate, cell.value)       # empty positions too: None, default style

Merged cells:

from openrsxl.extended import load_workbook

wb = load_workbook("book.xlsx", read_only=True, read_merged_cells=True)
ws = wb.active
for rng in ws.merged_cells.ranges:
    print(rng.coord, rng.start_cell.value)       # the top-left cell, as full mode has it
for row in ws.iter_rows(values_only=True):
    print(row)                                   # covered cells are None

Sheet layout and rules:

from openrsxl.extended import load_workbook

wb = load_workbook("book.xlsx", read_only=True, read_dimensions=True, read_sheet_properties=True,
                   read_data_validations=True, read_conditional_formatting=True, read_tables=True)
ws = wb.active
print(ws.column_dimensions["A"].width, ws.row_dimensions[1].height, ws.freeze_panes)
for dv in ws.data_validations.dataValidation:
    print(dv.type, dv.formula1, dv.sqref)
for cf in ws.conditional_formatting:
    print(cf.sqref, [rule.type for rule in cf.rules])
for table in ws.tables.values():
    print(table.displayName, table.ref)

A feature that is not enabled does not exist, exactly like the attributes openpyxl's read-only cells do not have: cell.formula without formula_and_value=True raises AttributeError. With values_only=True, rows hold cell.value only (merged cells are None, link targets fill empty linked cells, as in full mode).

examples/extended_streaming.py is a runnable tour of every flag.

Performance

Windows 10, Python 3.14, openpyxl 3.1.5 with lxml; each measurement in a fresh process, one at a time (docs/BENCHMARKS.md has every file and mode, and how to reproduce). Memory: the process' peak resident set size, with its increase over the baseline (before importing the library) in parentheses.

large_sample.xlsx - 98 MB, 25 sheets, 10 M cells	openpyxl	openrsxl
load + save	228.0 s - 4 617 MB (+4 599)	18.2 s - 291 MB (+273)
load + iterate all values	146.8 s - 4 594 MB (+4 576)	7.8 s - 359 MB (+340)
read-only: iterate all values	101.8 s - 149 MB (+131)	7.8 s - 119 MB (+100)
formula and cached value of every cell	full mode, loaded twice: > 270 s, > 4.5 GB	openrsxl.extended, one streaming pass: 11.3 s - 119 MB (+100)

100 000-row files written by openpyxl, and a 0.6 M cell workbook	speed-up	memory
load + save	7.0x - 12.3x	5.6x - 12x less
load + iterate all values	7.8x - 13.8x	3x - 11x less
read-only: iterate all values / cells	3.7x - 9.6x	less (43-49 MB vs 51-61 MB)
write-only: write 200 000 rows	2.6x (fill), same save	the same (43 MB)

The streaming features keep read-only mode's flat memory: all nine read_* / formula_and_value flags add ~2 MB to the 119 MB of large_sample.xlsx (whose peak is its shared strings table) and 10x the cells give the same peak. create_empty_cells adds no memory and ~0.2 µs per empty position.

Compatibility and limitations

openrsxl targets openpyxl 3.1.5 and reproduces its observable behaviour (values, types, errors, warnings, saved bytes - including openpyxl's bugs). What differs, in full in docs/COMPATIBILITY.md:

• Module names. Classes live in openrsxl.* (Font.__module__ == "openrsxl.styles.fonts"), so openpyxl and openrsxl objects cannot be mixed, and isinstance checks against openpyxl's classes are false - eg. for the ArrayFormula / DataTableFormula values of formula cells - unless openrsxl.extended.install_as_openpyxl() makes openpyxl's names point to openrsxl (above). Code that must accept both libraries' objects can test the attributes they share (value.t == "array", value.ref, value.text). Warnings carry the same category and message but other filename/lineno; tracebacks differ.
• openrsxl.__version__ is "3.1.5", the openpyxl version whose API is implemented (code checking the openpyxl version keeps working); the package's own version is importlib.metadata.version("openrsxl").
• ws._cells of loaded worksheets (a private attribute) is a dict-compatible CellStore (compact cells, Cell objects created on first access): isinstance(ws._cells, dict) is false, everything else behaves like the dict.
• CPython 3.10+. The Rust engine writes Python objects directly (CPython object layout). On other implementations (PyPy, GraalPy - built from source) it is disabled and openpyxl's Python code runs: same results, no speed-up. OPENRSXL_PURE_PYTHON=1 disables the engine on CPython too (debugging aid).
• Free-threaded Python (3.14t): the engine runs without the GIL; one worksheet used concurrently by several threads raises RuntimeError: Already borrowed instead of racing (openpyxl objects are not thread-safe either). Different workbooks in different threads work in parallel.
• Python fallbacks. Input the engine cannot reproduce with certainty (DOCTYPEs, non UTF-8 encodings, malformed XML, ...) is handled by openpyxl's own code; in read-only mode, warnings for elements before a malformed spot can be emitted twice.

openrsxl.extended (streaming features):

• cell.formula and cell.cached_value are new names (see above); a disabled feature raises AttributeError.
• Rows and columns are those of read-only mode: cells outside the sheet's <dimension> are not produced (cells created by a feature - a merged or linked empty cell - extend max_row / max_column), cell positions follow the <row> elements, and a sheet that claims a huge dimension yields that many (empty) cells, as in openpyxl (use iter_rows(max_row=..., max_col=...) or ws.reset_dimensions()).
• Empty positions are read-only mode's coordinate-less EmptyCell unless create_empty_cells=True, which creates full mode's cell for each of them (about 0.2 µs per empty position, no memory).
• Cell data that makes full mode fail in load_workbook (eg. an invalid style index) behaves as in read-only mode: the error is raised when the faulty data is used - when the style is read, or while iterating if a feature needs it (the style of a merged range's corner cell).
• Merged ranges are instances of a MergedCellRange subclass (isinstance holds); start_cell is computed on first access; cells and their links / comments are new objects on every iteration.
• Not streamed (yet): images, charts, drawings, pivot tables, ws.legacy_drawing; rich_text=True behaves as in read-only mode.

Security

openrsxl reads untrusted files the way openpyxl does, and is tested for it:

• The Rust XML tokenizer knows only the five predefined entities and character references; it never resolves external entities or touches the network or the file system. Documents with a DOCTYPE are parsed by openpyxl's own code path (expat with its entity-expansion limits, lxml with resolve_entities=False, defusedxml when installed and enabled) - exactly as openpyxl.
• Hostile input is tested in every load mode against openpyxl (tests/test_security.py): XXE in every part, billion laughs, 100 000-level element nesting, 100 000 comments around the root, numbers and coordinates beyond every integer limit, path traversal in relationships, duplicate zip members - and the fuzzing corpus of Apache POI (its crash test cases and XML bombs). Malformed input never panics or crashes the interpreter: it raises openpyxl's exception.
• Like openpyxl, decompression is not size-limited: a "zip bomb" is read as far as the data asks for (read-only and streaming modes keep memory flat).

How it works

openrsxl = openpyxl's object model (ported, MIT licensed - see NOTICE) + a Rust engine for everything that touches cell data:

openpyxl (pure Python)	openrsxl
xml.etree iterparse of worksheets	streaming Rust XML tokenizer (src/xml.rs)
WorkSheetParser.parse_cell / bind_cells	src/sheet.rs
Worksheet._cells dict of Cell objects	CellStore (src/store.rs): compact cells, Cell objects created on access
shared strings table	src/strings.rs, src/text.rs
shared-formula Translator / Tokenizer	src/formula.rs (compiled templates)
from_excel date conversion	src/dates.rs
<sheetData> serialisation (lxml and ElementTree output)	src/writer.rs
read-only row streaming	RowReader in src/lib.rs
(extended) sheet parts around <sheetData>	src/prescan.rs

Like calamine / python-calamine, cells of a loaded worksheet are kept in compact Rust form (10 bytes per cell, shared strings by index, formulae de-duplicated into position-independent templates) and Python objects are only created when needed; save() and iter_rows(values_only=True) work directly on the compact form. Whenever the engine meets input it cannot reproduce with certainty it hands over to openpyxl's original Python code, so the result is always openpyxl's.

How it is tested

openpyxl itself is the test oracle (a test dependency only): every test compares openrsxl with openpyxl.

• openpyxl's complete test-suite, ported (2 600 tests), with lxml and with ElementTree;
• 2 600+ differential tests: every public name and signature, ~375 one-feature reader fixtures in 5 load modes, every value type written, hostile input, openrsxl.extended against full mode, hypothesis-driven fuzzing of XML, numbers, formulae and random sequences of API calls on loaded workbooks;
• a corpus of ~1 960 real-world files from the test-suites of XlsxWriter (Excel-made files), Apache POI, ClosedXML, calamine, pandas, exceljs, python-calamine and openpyxl, compared in 10 modes each (tools/corpus.py): every cell, style and worksheet attribute, warnings, errors and the saved package byte for byte; openrsxl.extended against openpyxl's full mode.

pip install -e .[test]                 # or: maturin develop --release
pytest                                 # differential, security, fuzz tests
cd tests/upstream && pytest            # openpyxl's own test-suite
python tools/corpus.py <directory>     # whole-file differential run

docs/TESTING.md is the verification report of this release (corpus, results, security review); see docs/DEVELOPMENT.md for the layout and tools.

License

MIT. openrsxl contains code derived from openpyxl 3.1.5 (MIT licensed, copyright (c) 2010 openpyxl); see NOTICE.