[Web-SIG] Emulating req.write() in WSGI

[Graham Dumpleton]

On 29 June 2010 05:01, Aaron Fransen <aaron.fransen at gmail.com> wrote:
> One of the nice things about mod_python is the req.write() function.

One thing I should warn you about req.write() in Apache is that for
streaming data as you seem to be using it, it will accumulate memory
against a request for each write call and that will not be reused,
albeit it will be released again at the end of the request.

The problem here isn't actually in mod_python but in the underlying
Apache ap_rwrite() call.

What this function does is that for each call to it, it creates what
is called a bucket to hold the data to be written. The memory for this
bucket is allocated from the per request memory pool each time. This
bucket is then passed down the Apache output filter chain and
eventually the data gets written out.

Now, because the code doesn't attempt to reuse the bucket, that memory
then remains unused, but still allocated against the memory pool, with
the memory pool only being destroyed at the end of the request.

The outcome of this is that if you had a long running request which
continually wrote out response data in small bits using req.write(),
for each call there is a small increase in amount of memory taken from
the per request memory pool with it not being reused. Thus if the
request were running for a very long time, you will see a gradual
increase in overall memory usage of the process. When the request
finishes, the memory is reclaimed and reused, but you have by then
already set the high ceiling on ongoing process memory in use.

Anyway, thought I should just warn you about this. In part this issue
may even be why mod_python got a reputation for memory bloat in some
situations. That is, the fundamental way of returning response data
could cause unnecessary increase in process size if called many times
for a request.

Graham