Per discussion in:

  http://groups.google.com/group/modwsgi/browse_frm/thread/e7111816e70d236a

Apache implements an optimisation when keep alive is requested, which results in file buckets 
referencing a file containing > 255 and < ~8000 bytes, not being immediately flushed out but 
held over to when request actually ends or subsequent request on same connection occurs.

This causes problems for mod_wsgi as the file descriptor wrapped by wsgi.file_wrapper could 
have been closed by the time that Apache tries to use it. The end result is a completely empty 
response and an error on the Apache error logs (but only if LogLevel is debug) of:

  (9)Bad file descriptor: core_output_filter: writing data to the network

The affected code in mod_wsgi is in Adapter_output_file():

    APR_BRIGADE_INSERT_TAIL(bb, b);

    b = apr_bucket_eos_create(r->connection->bucket_alloc);
    APR_BRIGADE_INSERT_TAIL(bb, b);

    Py_BEGIN_ALLOW_THREADS
    rv = ap_pass_brigade(r->output_filters, bb);
    Py_END_ALLOW_THREADS

As the optimisation only occurs in Apache when EOS bucket is detected immediately after the file 
bucket, fix is to insert a flush bucket.

    APR_BRIGADE_INSERT_TAIL(bb, b);

    b = apr_bucket_flush_create(r->connection->bucket_alloc);
    APR_BRIGADE_INSERT_TAIL(bb, b);

    b = apr_bucket_eos_create(r->connection->bucket_alloc);
    APR_BRIGADE_INSERT_TAIL(bb, b);

    Py_BEGIN_ALLOW_THREADS
    rv = ap_pass_brigade(r->output_filters, bb);
    Py_END_ALLOW_THREADS