http.batch socket: too many open files

[ppcano]

The http.batch does not describe how many resources could be loaded in parallel in the same batch statement.

1. @liclac
   What could it be one possible max of requests in an http.batch statement? We need to know this to improve HAR converter WIP #291 . I have tested a little bit more and the error is also related to the number of VUs and sleep periods.

2. For users experiencing the same error, could we help them with a better error message or avoiding this error? Thoughts?

...
WARN[0008] Request Failed            error="Get https://d3dfcx9bcohi6j.cloudfront.net/static/images/logo-harvard.06e80d50d253.png: dial tcp 54.230.96.198:443: socket: too many open files"
WARN[0008] Request Failed            error="Get https://d3dfcx9bcohi6j.cloudfront.net/static/images/logo-sephora.d6de678a1214.png: dial tcp 54.230.96.198:443: socket: too many open files"
WARN[0008] Request Failed            error="Get https://d3dfcx9bcohi6j.cloudfront.net/static/images/logo-sfdc.51c6a0da92cf.png: dial tcp 54.230.96.198:443: socket: too many open files"
WARN[0008] Request Failed            error="Get https://d3dfcx9bcohi6j.cloudfront.net/static/images/logo-blue-medium.a122a7ddd06f.png: dial tcp 54.230.96.198:443: socket: too many open files"
WARN[0008] Request Failed            error="Get https://d3dfcx9bcohi6j.cloudfront.net/static/images/insights-arrow.9e9a1171b89b.png: dial tcp 54.230.96.198:443: socket: too many open files"
...

[liclac]

There should definitely be a limit to how many requests are executed in parallel.

[liclac]

Note: the current solution doesn't do per-host limiting, because that adds a significant amount of complexity for something I honestly can't see being very useful.

If you have a use case where a granular limit trumps a general one, please comment.

[ppcano]

@liclac

What will the expected behavior be for example, if the user sets the batch option to 10 and add 20 requests in a batch?

[liclac]

All of them get ran in parallel.

[robingustafsson]

Wait, why would all 20 requests, in @ppcano's example above, run in parallel?

Per-host limiting is needed as a tweakable knob to more realistically simulate the connection management of a browser (common use case when testing websites that need to support browsers with varying configs for number of connections per host).

I guess this CL (as a solution to golang/go#13957) will make implementing this easy when it lands (hopefully in Go 1.10, Feb next year) :)

[liclac]

Oh, I misread it. At most 10 requests will be ran in parallel, a new one is started every time a previous one is finished.

[liclac]

I'll toss in a batchPerHost option, the problem is that it's gonna need some major refactoring of how the module works to avoid parsing URLs multiple times, and more complexity is not what this module is really wanting for right now.

[liclac]

batchPerHost added. 06488b8