This communicates more clearly for which versions of Node.js
the tests have run.
8 is not listed because tests are failing due to the nock HTTP mocking
library being incompatible with Node v8.
When proxyReq is aborted (by us - in response to a redirect response),
it can ultimately trigger an "error" event of type ECONNRESET.
This error was unconditionally propagated to the error handler:
https://github.com/nodejitsu/node-http-proxy/blob/v1.11.1/lib/http-proxy/passes/web-incoming.js#L134-L140
Our proxy error handler responds by writing a response + error code,
which prevents the response from being overwritten.
I found this after upgrading the nock library to 5.2.1+, which contains
510e31c73e
The redirect tests that were previously passing in Node 0.10.x were
failing, with errors like "Can't set headers after they are sent.".
The new implementation makes sure that the proxied response is not
forwarded to the original response.
- Add checkRateLimit option to the API.
- Extend the default server.js with environment variables
CORSANYWHERE_WHITELIST (re-using originWhitelist) and
CORSANYWHERE_RATELIMIT (using the new checkRateLimit option)
to make it easy to enforce usage limits.
- Document that Heroku doesn't want open proxies.
- CI & 100% test coverage
- Do not crash when a malformed HTTP response is received.
- Remove getHandler (its documented form was broken anyway).
- Allow custom help file to be set.
- Add option to redirect same-origin requests instead of proxying them.
Some clients try to use CORS Anywhere, even for same-origin requests...
Add a new setting "redirectSameOrigin" to not waste server resources on
proxying such requests.
Fixes#42
It is not supported, and its current form was not documented.
If you need to proxy a request without listening on a port, just
dispatch a 'request' event on the return value of createServer().
Notes:
- npm run test also runs the memory tests, which might be flaky on
some systems.
- npm run test-coverage only runs the unit tests (it can't run the
memory test anyway because the memory tests create a Node child
process, and istanbul cannot cover that).
- Support proxying through proxy (`getProxyForUrl` option, defaulting to
proxy configuration from the environment variables, see
https://github.com/Rob--W/proxy-from-env) (#37).
- Update gTLD list.
Node.js has forbidden header names with spaces in
6192c9892f.
The next error is thrown when the test sets the "some header" header:
Uncaught TypeError: Header name must be a valid HTTP Token ["some header"]
To fix it, replace spaces with dashes.
Changes:
- Updated to node-http-proxy 1.11.1 from 1.3.0
- Add httpsOptions configuration to listen as a https server.
- Update gTLD list, and allow requests to unlisted TLDs if the
scheme is explicitly set (#14).
- Added loads of unit tests and automatic memory leak tests.
Using the performNRequests, I collected the following statistics
before choosing the maximum allowed "leaked" memory.
Node.js 0.12.2,
Using the http module ('use-http-instead-of-cors-anywhere'):
Memory usage delta: 132800 (100 requests of 50 kb each, 250ms)
Memory usage delta: 110144 (100 requests of 1 kb each, 172ms)
Memory usage delta: 709936 (1000 requests of 1 kb each, 902ms)
Memory usage delta: 865104 (10000 requests of 1 kb each, 7073ms)
Memory usage delta: 930416 (100000 requests of 1 kb each, 62856ms)
Using CORS Anywhere:
Memory usage delta: 356784 (100 requests of 50 kb each, 1004ms)
Memory usage delta: 355248 (100 requests of 1 kb each, 641ms)
Memory usage delta: 1326856 (1000 requests of 1 kb each, 3338ms)
Memory usage delta: 1462584 (10000 requests of 1 kb each, 21186ms)
Memory usage delta: 1473624 (100000 requests of 1 kb each, 211202ms)
Clearly, there is a small leak, but it is not proportional/linear
in terms of the number of requests, so the observed "leak" is probably
not an issue. Furthermore, the "leak" also occurs with the plain
http module.
After setting fixed limits, I ran the tests on Node.js 0.10.25 and
observed that the tests failed due to the too low limits, so I
incremented the limits (400 -> 550, 1500 -> 2000).
All xfwd and tests fails with this change, but that's because the
tests themselves were incorrect (using port 80/443 instead of the
ephemeral ports created for the tests).
