Depending on the nature of the jobs, you can follow the usual approach and use the first two letters or some other (evenly distributed!) criteria to distribute the jobs among subdirectories. This makes scanning for fresh jobs harder though. Alternatively, move jobs that are "in processing" into a separate directory which is not scanned. That will reduce the load that idle jobs produce while scanning for work to do.

If you have to have a high throughput and can't batch your 100k requests into jobs of (say) 100 items or so, I'd look at premade solutions or maybe just at dedicating a database machine which serves as the central job directory.