I think we currently don't have any functions in time-sensitive streams (they're rather new, so they are used for new features such as Oyster automated refunds, which can be applied a few days later if need be), so when I say we haven't had any performance issues with them, it has to be taken with a pinch of salt.

I think if we were using microservices, they would have substantially more logic in them than the individual functions have, so they would have a lower network overhead. They wouldn't scale automatically, so we'd have to have quite a few machines on all the time, which would cost a bit more.

The main benefit we enjoy from functions is the ability to change code with zero downtime, low risk and minimum disruptions to the whole service.

We've found the interfaces in Azure also to be great - we can write end-to-end tests that can poll service bus to check when all messages are delivered to assert against any final case.

Do you know of any write ups about TFL's infrastructure/processes? Would love to read about it.

I'd also like to buy whoever lead the TFL API's move from XML to JSON a coffee|beer!