node heap wtf

run.md

$ node --expose-gc server.js

in another terminal:

$ curl localhost:5001

output in first terminal shows that req.foo is still sitting in the heap:

{ before: 
   { nodes: 27143,
     time: Wed Mar 05 2014 13:59:20 GMT-0800 (PST),
     size_bytes: 4595904,
     size: '4.38 mb' },
  after: 
   { nodes: 25238,
     time: Wed Mar 05 2014 13:59:25 GMT-0800 (PST),
     size_bytes: 84183520,
     size: '80.28 mb' },
  change: 
   { size_bytes: 79587616,
     size: '75.9 mb',
     freed_nodes: 3430,
     allocated_nodes: 1525,
     details: 
      [ [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object],
        [Object] ] } }

server.js

var express = require('express');
var memwatch = require('memwatch');
var _ = require('underscore');

var app = express()

var checkHeap = function() {
  global.gc();
  var diff = snapshot.end();
  console.log(diff);
};

var snapshot = null;
app.use(function(req, res, next) {
  snapshot = new memwatch.HeapDiff();
  req.foo = _.range(10000000); // allocate a ton of space
  next();
})
app.use(app.router);

app.get("/", function(req, res, next) {
  res.write("hello world");
  res.end();
  setTimeout(checkHeap, 5000);
})

var server = app.listen(5001);

I simplified the server code a bit:
https://gist.github.com/defunctzombie/9380299

To help rule out more codepaths

Making multiple requests (after waiting for the snapshot from each one) does not yield linear memory growth.

Ok, so adding more gc() calls before the final memory check makes the issue go away. I suspect this is because gc() is more of a hint than an absolute. Someone with more v8 experience needs to comment on that.

hmmm i actually don't think is a problem. i read somewhere that v8 doesn't actually gc() until it starts running out of memory, so unless node is beginning to use a lot of memory, you won't see automatic garbage collection. thus, manually gc()ing multiple times solves the issue.

a real test should be whether the process dies when you continuously hit it with requests.

we could always kill the .next reference when the response emits a finish or close event just to be safe.