svnwrap: use custom StringIO class in get_file() The wrappers were calling ra.get_file() with a cStringIO object. Empirically, svn 1.7.5 is writing 16kB blocks to the stream object, and cStringIO reallocates its internal buffer and doubles its size whenever it is filled. With large committed files this requires two large memory blocks at the same time. SimpleStringIO implements the mimimum StringIO interface used by ra.get_file() but instead stores all the blocks and "join" them at the end. It means more fragmentation but requires only one large block, without overallocation. Also, 16kB blocks should be friendly to most allocators. In practice, this simple change let me convert a revision containing multiple moderately large files, the largest being around 450MB, with a 32-bits Windows setup, python 2.7, swig svn 1.7.5, in stupid mode, while it was previously aborting with "not enough memory". The same revision still fails in replay mode.

#!/bin/sh
mkdir temp || exit 1
cd temp
svnadmin create repo
svn co file://`pwd`/repo wc
pushd wc
mkdir -p project/trunk
svn add project
svn ci -m 'trunk'
cd project/trunk
echo a > a
mkdir narf
svn add a narf
svn ci -m 'file and empty dir'
popd
svnadmin dump repo > ../empty_dir_in_trunk_not_repo_root.svndump
echo 'dump in empty_dir_in_trunk_not_repo_root.svndump'
echo 'you can probably delete temp now'