I believe the issue was twofold:
When done importing, the calling context is likely still valid, so if the output channel is not read enough and reach capacity, some event producer down the line can be blocked trying to send in that channel. When closing it, this send is still trying to proceed, which is illegal in go.
In rateLimitHandlerClient, there was a need to 2 different type of output channel: core.ExportResult and ImportEvent. To do so, the previous code was using a single channel type RateLimitingEvent and a series of goroutines to read/cast/send to the final channel. This could result in more async goroutine being stuck trying to send in an at-capacity channel. Instead, the code now use a simple synchronous callback to directly push to the final output channel. No concurrency needed anymore and the code is simpler.
Any of those fixes could have resolved the data race, but both fixes is more correct.
The Github bridge itself should not write anything. This commit removes
code writing to stdout and itroduces an event `ImportEventRateLimiting`
to `core.ImportResult` in order to inform about a rate limiting situation
of the Github GraphQL API. Now the communication with the user is
delegated to the various user interfaces.
The old implementation of the github bridge used maps to store several
channels holding data obtained from the Github API. Removing the maps and
simply packing data and channels together in a struct and passing it
through one single channel makes the program simpler in terms of
concurrency and, additionally, enables the garbage collector to free the
memory gradually without any additional provisions.
