Openwhisk: actionProxy runtime input parameters limited to 128k

[jthomas]

Environment details:

IBM Bluemix (US-South).

Steps to reproduce the issue:

Create a new openwhisk/dockerskeleton action and use a JSON string with input parameters larger than 128KB.

$ printf 'a%.0s' {1..131072} | echo "\"$(cat)\"" | jq '{input: .}' > test_too_large.json
$  wsk action create argmax --docker openwhisk/dockerskeleton
$ wsk action invoke argmax -P test_too_large.json -b

This invocation fails with the following error logs...

2017-07-28T10:42:39.378372326Z stdout: [Errno 7] Argument list too long

This was discovered by an external user (https://github.com/apache/incubator-openwhisk/issues/2465).

Additional information you deem important:

This is happening because the actionProxy.py passes the input parameters as a JSON string through a command-line argument.
https://github.com/apache/incubator-openwhisk/blob/master/core/actionProxy/actionproxy.py#L131

Linux limits the maximum argument length to MAX_ARG_STRLEN which is 128KB on Linux.

Fixing this issue

Changing the actionProxy.py to pass input parameters using stdin rather than a command-line argument would remove this limit.

Here's an example of the change needed:

        try:
            input = json.dumps(args)
            p = subprocess.Popen(
                [self.binary],
                stdin=subprocess.PIPE,
                stdout=subprocess.PIPE,
                stderr=subprocess.PIPE,
                env=env)
        except Exception as e:
            return error(e)

        # run the process and wait until it completes.
        # stdout/stderr will always be set because we passed PIPEs to Popen
        (o, e) = p.communicate(input=input.encode())

https://gist.github.ibm.com/thomas6/99558d3c64bb8f571fe55c3cdd1b3817

I've built this change into an external image @ jamesthomas/custom_skeleton and been using it to resolve this issue in another project.

Do you want me to submit a PR? This would be a breaking change to existing users.

[rabbah]

This affects swift actions also although the cmd line is ignored for those actions. One plausible fix for swift is to override run or refactor it to allow for runtime specific code compensation.

[rabbah]

Here's a swift test that replicates the limitation https://github.com/apache/incubator-openwhisk/compare/master...rabbah:arglist?expand=1#diff-cd4d7cf77b5a623c67fc14b0de34cc1aR165

An option to fix the limit while not breaking existing actions: send all arguments over stdin, and for arguments <= 128K, also send them over argv[1]. We know already actions that otherwise rely on the input argument will fail for larger payloads.

[rabbah]

Fixed in the docker proxy https://github.com/apache/incubator-openwhisk-runtime-docker/pull/3