[hddqsb]

It's not a big deal because there is no need to retry 20 times. The probability of getting a deleted article several times in a row is very low, so you could limit to e.g. 5 tries, and if all are deleted fall back to the next non-deleted article (or even a hard-coded article). The bias would be negligible assuming the proportion of deleted articles is low; to guarantee that it is low, one can periodically renumber to eliminated deleted articles (this can be done efficiently using the trick suggested by @munificent; but the naive O(n) approach would probably be good enough).