`createGetComponent` doesn't prevent leaking data between requests about hypernova

[alexindigo]

@goatslacker It seems more than documentation bug, running things in VM implies isolation.
Seems like having cache in front of the VM creates the issue.

[goatslacker]

I marked it as a bug + documentation because it can be both. Either we fix the issue or we amend the documentation to make it explicit that using createVM and createGetComponent will leak.

It's a tough situation because if you don't have a caching layer then you're adding server-render time to your total time. It's not cheap to eval code.

There are a few things you can do to make things faster AND get the benefits of isolation. You'd have to basically pack all the code into a single bundle (webpack/browserify works) and then you can re-eval that inside a VM. I don't have numbers with me but that shouldn't add too much latency.

[alexindigo]

Yep, I'm planning to try different options too. Not sure how bundling "all the code" would help though. My immediate candidates to try are: regular vm thing, code-revaluation and hacky thing of new Function combined with with :))) but need to setup sound load testing environment first and get the base line.

[goatslacker]

Bundling all the code helps because a lot of the time spent creating a new VM is in resolving the require paths, crawling through the file system, reading all the code and evaluating it (which creates a separate new module). When you bundle all the code there are less files that need to be resolved and thus less code to iterate through. It's just a single bundle VM.

Don't take my word for it though, try it out yourself :)

[alexindigo]

I was going to preload all the files (either parsed or not) at start time in any case, I/O during request is not my thing :)

[goatslacker]

I've tried that before as well, caching all the files and then just re-evaluating the VM on each request. I forget what my findings were. It might be wise to revisit that idea.

Right now having things shared between each VM isn't that big of a deal given the tradeoffs. Rendering is still a synchronous affair so any potential data that could be leaked is wiped at the start of each new request. This will be something to look after though as we (hopefully soon) move towards either streaming or async rendering.

[alexindigo]

I want to know more about async rendering :)

[morganpackard]

I'm going to take a crack at this. I think we can cache the function returned by vm.runInNewContext and just repeatedly call it to initialize a clean module once per request.

I did some profiling on module initialization, measure the time to load, parse, and run react.js.

// parse the wrapped module code
const executedWrapped = runInNewContext(wrapped);

// initialize module
const initializedReact = executedWrapped(localExports, require, localModule, __localFilename, __localDirname);

I recorded the time to parse, and the time for the first module initialization, then the average time for 10k subsequent initializations. Once the module has been initialized once, subsequent initializations are much cheaper, but not free.

time to interpret the initial string (but not initialize the module): 35ms
time for initial module initialization: 59ms
average time for subsequent module initialization: 2.9ms

In this test, subsequent initializations (after the expensive initial one) take about 5% of the time as the the initial round.

[morganpackard]

I have a solution that's working. Here are some key points on what my fork does:

• Adds an option in vm.run to allow you to use module wrapper functions instead of caching the module exports. This allows you to safely use 'stateful' modules, reseting the state between requests.
• Allows you to pass in globals on a per-request basis if you're using the cached wrapper function option. There's now a single global context object which is shared between all modules. This object is "cleaned" between requests. Without this option, the behavior is not changed.

I'd love to move toward a PR for this code if you'd be up for taking a look. It still needs a little polishing, but I'd be happy to submit a work in progress PR if you're interested in seeing it sooner rather than later.