In such cases I do the following:

1. I reauthenticate the AGOL web connection through Tools > FME options > Web connections > right mouse click on connection and select 'reauthenticate'.
2. You can set the 'Features per request' lower to e.g. 500.

If it still fails, you might want to use another approach, because it will be hard to trace which features (sets) have been uploaded and which haven't.

I'm such cases you could prepare the sets in different tables or databases. Next you can create a new workspace with a WorkspaceRunner. The WorkspaceRunner can call the original workspace. With the new workspace you make sure every dataset is provided.

In the last scenario you'll have a good control of which sets have been written and which ones haven't. Now it is much easier to retry the ones that have failed.

Thank you @lars_de_vries! Will give it a shot.

It's a known issue. Believed to be caused by Amazon Cloudfront caching. Working with esri support and development to resolve. Follow this thread for updates: https://knowledge.safe.com/questions/84687/arcgis-online-readerwriter-error-code-503.html

Thanks for the update link