How to load a large xlsx file with Apache POI?

I have a large .xlsx file (141 MB, containing 293413 lines with 62 columns each) I need to perform some operations within.

I am having problems with loading this file (OutOfMemoryError), as POI has a large memory footprint on XSSF (xlsx) workbooks.

This SO question is similar, and the solution presented is to increase the VM's allocated/maximum memory.

It seems to work for that kind of file-size (9MB), but for me, it just simply doesn't work even if a allocate all available system memory. (Well, it's no surprise considering the file is over 15 times larger)

I'd like to know if there is any way to load the workbook in a way it won't consume all the memory, and yet, without doing the processing based (going into) the XSSF's underlying XML. (In other words, maintaining a puritan POI solution)

If there isn't tough, you are welcome to say it ("There isn't.") and point me the ways to a "XML" solution.


Solution 1:

I was in a similar situation with a webserver environment. The typical size of the uploads were ~150k rows and it wouldn't have been good to consume a ton of memory from a single request. The Apache POI Streaming API works well for this, but it requires a total redesign of your read logic. I already had a bunch of read logic using the standard API that I didn't want to have to redo, so I wrote this instead: https://github.com/monitorjbl/excel-streaming-reader

It's not entirely a drop-in replacement for the standard XSSFWorkbook class, but if you're just iterating through rows it behaves similarly:

import com.monitorjbl.xlsx.StreamingReader;

InputStream is = new FileInputStream(new File("/path/to/workbook.xlsx"));
StreamingReader reader = StreamingReader.builder()
        .rowCacheSize(100)    // number of rows to keep in memory (defaults to 10)
        .bufferSize(4096)     // buffer size to use when reading InputStream to file (defaults to 1024)
        .sheetIndex(0)        // index of sheet to use (defaults to 0)
        .read(is);            // InputStream or File for XLSX file (required)

for (Row r : reader) {
  for (Cell c : r) {
    System.out.println(c.getStringCellValue());
  }
}     

There are some caveats to using it; due to the way XLSX sheets are structured, not all data is available in the current window of the stream. However, if you're just trying to read simple data out from the cells, it works pretty well for that.

Solution 2:

A improvement in memory usage can be done by using a File instead of a Stream. (It is better to use a streaming API, but the Streaming API's have limitations, see http://poi.apache.org/spreadsheet/index.html)

So instead of

Workbook workbook = WorkbookFactory.create(inputStream);

do

Workbook workbook = WorkbookFactory.create(new File("yourfile.xlsx"));

This is according to : http://poi.apache.org/spreadsheet/quick-guide.html#FileInputStream

Files vs InputStreams

"When opening a workbook, either a .xls HSSFWorkbook, or a .xlsx XSSFWorkbook, the Workbook can be loaded from either a File or an InputStream. Using a File object allows for lower memory consumption, while an InputStream requires more memory as it has to buffer the whole file."

Solution 3:

The Excel support in Apache POI, HSSF and XSSF, supports 3 different modes.

One is a full, DOM-Like in-memory "UserModel", which supports both reading and writing. Using the common SS (SpreadSheet) interfaces, you can code for both HSSF (.xls) and XSSF (.xlsx) basically transparently. However, it needs lots of memory.

POI also supports a streaming read-only way to process the files, the EventModel. This is much more low-level than the UserModel, and gets you very close to the file format. For HSSF (.xls) you get a stream of records, and optionally some help with handling them (missing cells, format tracking etc). For XSSF (.xlsx) you get streams of SAX events from the different parts of the file, with help to get the right part of the file and also easy processing of common but small bits of the file.

For XSSF (.xlsx) only, POI also supports a write-only streaming write, suitable for low level but low memory writing. It largely just supports new files though (certain kinds of append are possible). There is no HSSF equivalent, and due to back-and-forth byte offsets and index offsets in many records it would be pretty hard to do...

For your specific case, as described in your clarifying comments, I think you'll want to use the XSSF EventModel code. See the POI documentation to get started, then try looking at these three classes in POI and Tika which use it for more details.

Solution 4:

POI now includes an API for these cases. SXSSF http://poi.apache.org/spreadsheet/index.html It does not load everything on memory so it could allow you to handle such file.

Note: I have read that SXSSF works as a writing API. Loading should be done using XSSF without inputstream'ing the file (to avoid a full load of it in memory)