The current bidding strategy is good for finding machines that are cheap, but not machines that will stay around for longer. By bidding higher, we can have machines stick around longer, but we do not want too many bids at a high price, otherwise our cost exposure is too high. Implement a "1/n" bidding strategy to limit the overall cost, increase spot lifetimes, and still pay a low price. Let B be the maximum bid for a given instance-type (for given zone). Bid for machines with the following pattern: MachineA -> bid B MachineB -> bid B/2 MachineC -> bid B/3 etc... In the face of price increases, AWS will shutdown the lesser bids, freeing up money to continue funding the higher bids. For any AWS price point, we will not spend more than B, plus we effectively let AWS map the price to the number of machines at our disposal. This is not a good strategy for ETL machines, where we take advantage of the last free hour, and we desire an AWS-initiated shutdown.
You need to log in before you can comment on or make changes to this bug.