Spot price is per-instance-hour: one single-threaded step wastes a 64-core VM

When you size a burst EC2 VM for a parallel data-conversion pipeline, profile each stage's parallelism first. In one FSDB-to-columnar conversion, the parallel shard stage kept ~45-65 of 64 cores busy for 12 min. The single-threaded pack/encode stage that followed took 3x longer on one core while the other 63 idled, but billing is per instance-hour, so the idle cores cost the same. CPU-seconds x per-core price (~2 cents/core-hour on a 64-vCPU spot VM) gives a tiny 'compute cost' that misleads; the real bill is wall-clock. Before launching big VMs, make the serial tail parallel (per-signal block encode across threads, or emit encoded blocks directly from the shards and skip the intermediate records). Also launch the VM only once the job is ready to start, not ahead of time.