It can work for some usage patterns, but it’s important to understand the trade offs Hosted gives up easy hardware replacement, easier redundancy, 24/7 monitor

It can work for some usage patterns, but it’s important to understand the trade offs Hosted gives up easy hardware replacement, easier redundancy, 24/7 monitoring etc But a hybrid approach worked well for one place I worked at. Their work pattern was: Job scheduled by scheduler and pushed into queue (all AWS systems) Local cluster of 20 machines pull from that queue, compute the job, and store the result back. It worked because: 1) if a job failed the scheduler would re-schedule it 2) the local machines were treated as disposable, if one broke it was repaired or thrown away 3) deployment to the local nodes was simply copying a binary and config file 4) the latency didn’t matter as most effort of the jobs was compute 5) the bandwidth didn’t matter because the jobs were small and didn’t require much data 6) the tasks were embarrassingly parallel, so throughput scaled linearly with nodes Sometimes it can work, but without knowing the specifics of your compute jobs the best I can do are the above generalisations