Due to the order of the previous logic in ScanReader, an error not caused by EOF would effectively get ignored since an error will always be returned with `isPrefix` equal to false, thus triggering the first break, and error checking is not performed beyond that point.
Thus, canceling an installation process for a server while this process was running would hang the routine and cause the loop to run endlessly, even with a canceled context.
If my debugging is correct, this should address pterodactyl/panel#3903 in its entirety by addressing a few areas where it was possible for a channel to lock up and cause everything to block
This also removes server process termination logic when a server is breaching the output limits. It simply continues to efficiently throttle the console output.
This change fixespterodactyl/panel#3921 by implementing logic to drop the oldest message in a channel and push the newest message onto the channel when the channel buffer is full.
This is distinctly different than the previous implementation which just dropped the newest messages, leading to confusing behavior on the client side when a large amount of data was sent over the connection.
Up to 10ms per channel is allowed for blocking before falling back to the drop logic.
This new package has significant better resource usage, and we do a _lot_ of JSON parsing in this application, so any amount of improvement becomes significant
This also fixes an improperly written server deletion listener to look at the correct context cancelation.
Theoretically this should help address the issues in pterodactyl/panel#3596 but I'm not really sure how that happens, and theres no steps for reproduction.
* Cleanup server sync logic to work in a single consistent format
Previously we had a mess of a function trying to update server details from a patch request. This change just centralizes everything to a single Sync() call when a server needs to update itself.
We can also eventually update the panel (in V2) to not hit the patch endpoint, rather it can just be a generic endpoint that is hit after a server is updated on the Panel that tells Wings to re-sync the data to get the environment changes on the fly.
The changes I made to the patch function currently act like that, with a slightly fragile 2 second wait to let the panel persist the changes since I don't want this to be a breaking change on that end.
* Remove legacy server patch endpoint; replace with simpler sync endpoint
If you have two env variables (for example ONE_VARIABLE and ONE_VARIABLE_NAME) ONE_VARIABLE_NAME has prefix ONE_VARIABLE and will be skipped.
Co-authored-by: Jakob <dev@schrej.net>
If a request to upload a file part to S3 fails for any 5xx reason it will begin using an exponential backoff to keep re-trying the upload until we've reached a minute of trying to access the endpoint.
This should resolve temporary resolution issues with URLs and certain S3 compatiable systems such as B2 that sometimes return a 5xx error and just need a retry to be successful.
Also supports using the server context to ensure backups are terminated when a server is deleted, and removes the http call without a timeout, replacing it with a 2 hour timeout to account for connections as slow as 10Mbps on a huge file upload.
