I put in the batching point at: 1) over a certain time-limit, e.g. 100ms 2) over a certain number of commands, e.g. 2000 3) encountering a specific 'flush' comm

I put in the batching point at: 1) over a certain time-limit, e.g. 100ms 2) over a certain number of commands, e.g. 2000 3) encountering a specific 'flush' command, that I injected sometimes to simply the logic Those were the numbers I, by trial and error, determined for my server. But remember, I didn't do SELECTs; I put everything I might SELECT into memory. So my numbers are rather application-specific perhaps. Its important to measure the average and worst case response times, not the best-case. Under light load, best-case times are from blocking calls, which is why conventional ORM looks good enough in prototyping ;)