POST operations need to handle intermittent network failure.

[samoconnor]

The GitHub doc says that they have been careful to use HTTP methods with semantics appropriate to the API operations.

HTTP.jl retries most request automatically in the event of network failure. However, POST operations are not generally idempotent and cannot be automatically retried. GitHub.jl should implement system-state-aware retry loops around POST requests. i.e. when a POST operation throws an HTTP.IOError GitHub.jl should retry the operation in a way that avoids duplicate changes to system state.

In the case of the create_status operation that is the subject of JuliaWeb/HTTP.jl#220, POST /repos/:owner/:repo/statuses/:sha, it seems that double-posting would create two statuses, https://developer.github.com/v3/repos/statuses/, so in the event of a failure, the GitHub.jl create_status function should do a GET to check if the status was created before retrying (or, if duplicate statuses are not considered harmful to overall system correctness, just set retry_non_idempotent=true).

GitHub.jl/src/repositories/statuses.jl

Lines 30 to 33 in beb3659

	@api_default function create_status(api::GitHubAPI, repo, sha; options...)
	result = gh_post_json(api, "/repos/$(name(repo))/statuses/$(name(sha))"; options...)
	return Status(result)
	end

See also JuliaWeb/HTTP.jl#214.

This issue may apply in other places that use gh_post_json.

[ararslan]

Help with this would be very much appreciated. This is causing significant issues with Nanosoldier, which is critical community infrastructure.

[KristofferC]

Do you know what postrequest that is failing? If it is an idempotent one then we could just run with retry_non_idempotent=true for that request?

[samoconnor]

@KristofferC the request that is failing is create_status. As described above "it seems that double-posting would create two statuses", i.e. it is not idempotent.

I would like to help out, but I need to see more detailed logs of what is going on with Nanosoldier: JuliaWeb/HTTP.jl#220 (comment)
JuliaWeb/HTTP.jl#220 (comment).

[samoconnor]

See log analysis: JuliaWeb/HTTP.jl#220 (comment)

To deal with the immediate practical problem in the Nanosoldier's use case I recommend modifying GitHub.jl to:

• set retry_non_idempotent=true, restoring the old behaviour of blindly retrying POST requests; or
• set idle_timeout=30 or idle_timeout=20 to remove/reduce the incidence of attempting to reuse a connection that has in-fact become unusable; or
• (as a last resort) set reuse_limit=0 to disable connection re-use completely.

(To be reliable, GitHub.jl will still need to implement state-aware retry of non-idepmontent operations, but the changes above should improve failure rates in the short term).

[samoconnor]

I think this probably should stay open per #106 (comment): "still need to implement state-aware retry of non-idepmontent operations"