Depending on your use case and how much startup time you can tolerate... We need the ability to run 1,000s to 100,000s of ~3 second jobs as fast as possible, b

Depending on your use case and how much startup time you can tolerate... We need the ability to run 1,000s to 100,000s of ~3 second jobs as fast as possible, but only sporadically (~100 times a day with our current customer base). At that "low" frequency, can't justify the cost of keeping that many dedicated servers up 24/7 to meet this intermittent need, and we need sub-second startup times so we can't launch normal VMs on demand. However, the cost of servers on Google Compute Engine (GCE) is actually cheaper than lambda if you have big batches of jobs. The VMs could be launched via their autoscaler or Dataflow, and jobs could be dispatched via Pub/Sub. If you can secure enough preemptible GCE instances (which are of limited availability), these are even cheaper. GCE VMs launch in about a minute, compared to AWS's 3-7 minutes, so on-demand launching is actually doable for some use cases as long as you don't mind paying the 10 minute minimum. This also gives you access to huge resources (CPU and RAM) if needed. Specifically, ~1.1 million 3300 ms 1536 mb lambda jobs cost us $90.04, and GCE would be $50 on regular or $15 on preemptible GCE VMs for the same number of jobs in the /same total execution time/. After requesting a Lambda quota of 100,000 concurrent jobs, we were only allowed 1,000 ("for now", they say) -- thus it takes an hour for that many jobs to complete, so GCE$ = (1 hr) * (cost per vCPU/hr) * (1000 vCPUs). Note that we're using the max allowable memory on Lambda because that ~linearly decreased our execution time, but we use an average of 400 mb. My estimate also doesn't take into account the overhead time that comes from launching lambda jobs that might be reduced via the home-brew solution I described above. Someone else mentioned the issue with Lambda only having 0.10 vintage node. We have a somewhat complicated albeit elegant solution for communicating with a portable node 4.x binary via IPC. If AWS ever gets to allowing specified node versions, that would reduce our runtime marginally. (BTW, I put in a feature request to GCE for a lambda-like service here: https://code.google.com/p/google-compute-engine/issues/detail?id=253 https://code.google.com/p/google-compute-engine/issues/detai...)