GM Version checking doesn't work

[jiivan]

To reproduce, run pytest -vs --maxfail=1 --lf scripts/concent_acceptance_tests/force_payment/test_payment.py
Result:

DEBUG    [golem.network.concent.client       ] send_to_concent(): POST 'https://staging.concent.golem.network/api/v1/send/' hdr: {'Content-Type': 'application/octet-stream', 'X-Golem-Messages': '2.21.1'} 
DEBUG    [golem.network.concent.client       ] Headers received from Concent: {'Server': 'nginx/1.15.0', 'Date': 'Fri, 04 Jan 2019 09:02:59 GMT', 'Content-Type': 'application/json', 'Content-Length': '192', 'Connection': 'keep-alive', 'Concent-Version': 'v0.10.2-101-gbecec05d', 'Concent-Golem-Messages-Version': '2.17.0', 'X-Frame-Options': 'SAMEORIGIN'} 
WARNING  [golem.network.concent.client       ] Concent request failed with status 400 and response: '{"error": "Golem Message contains wrong fields. Should be an instance of <class \'golem_messages.message.tasks.SubtaskResultsAccepted\'> not <class \'bool\'> [subtask_results_accepted_list:True]"}'

Analysis:
Local GM version is 2.21.1, remote version is 2.17.0. Concent service acts as if those versions are compatible, hence it doesn't understand new message nesting protocol.

[PaweuB]

Pivotal:
https://www.pivotaltracker.com/story/show/163016112

[kbeker]

Sanity check needs to be added on django side (middleware). There are 3 cases:

• If X-Golem-Messages is not added or it's None in header:
  Pass and do nothing,
• If X-Golem-Messages is added and it contains correct GM version:
  Pass and do nothing,
• If X-Golem-Messages is added but contains wrong version of GM or it's contains version which could not be parsed:
  Respond with error

[cameel]

Just to clarify: this check should be performed by the router that kicks in before you even get to the cluster. But currently that can only be enabled if we have exactly two clusters. Otherwise routing is disabled and there are no checks. On staging we have only a single version right now. I've added an issue to make it work the same way no matter whether we have one or two versions available: golemfactory/concent-deployment#289. When the requested version is not available, the router will now respond with HTTP 404.

The change proposed by @kbeker above is an additional sanity check in the Concent app - if you, for whatever reason, manage to get through to a Concent cluster with an incompatible version of GM in your X-Golem-Messages header, Concent should respond with ServiceRefused. This would indicate a cluster misconfiguration so you should never get this in production.

@kbeker The header is called X-Golem-Messages and contains GM version used by the Golem client. That's the one we should check. Concent-Golem-Messages-Version is a completely different header Concent puts on its HTTP responses and contains GM version used by Concent.

Also, this should be obvious, but just in case: note that (as stated in the #857 blueprint) versions do not need to be exactly the same. It's enough if they're compatible:

Versions are considered compatible if they share the minor and major version number. E.g 2.18.5 is compatible with 2.18.1 but not with 2.17.5 or 3.0.0.

Note that the way we do it in the router matches semver only for release versions but it does not really matter because this mechanism is meant only for production. We just look at first two numbers and ignore the rest: golemfactory/concent-deployment#258 (comment). It's best if the Python code checks it exactly the same way.

[kbeker]

@kbeker The header is called X-Golem-Messages and contains GM version used by the Golem client. That's the one we should check. Concent-Golem-Messages-Version is a completely different header Concent puts on its HTTP responses and contains GM version used by Concent.

@cameel You're right, my bad, edited.

[cameel]

Update: on the standup today we agreed that instead of ServiceRefused, Concent should return HTTP 404 when the version does not match.

[pawelkisielewicz]

Fixed in concent and should work properly. It was fixed in #1057, didn't close automatically