I think I have an explanation for the anomaly at 3. I queried not just based on first digit, but also on number of digits.
Here is the query I used:
SELECT LEFT(reputation, 1) AS first_digit
, LENGTH(reputation) AS num_digits
, COUNT(*)
FROM users
GROUP BY first_digit, num_digits
ORDER BY first_digit ASC, num_digits DESC
After massaging that a bit, you get this as your raw data:
1st Digit 5 digits 4 digits 3 digits 2 digits 1 digit
1 190 2504 7049 12838 52486
2 30 876 2973 7180 0
3 11 485 2150 4634 2623
4 7 274 1578 3074 0
5 0 157 1114 2336 301
6 1 128 899 1972 4
7 1 84 687 1481 146
8 0 62 618 1305 0
9 1 40 446 1141 447
The 1-digit reps throw things off. In particular, there is a large number of users with rep of 3, because that is your rep if you ask one question, get no upvotes, and accept some answer. Also, due to a bug, if a new user receives a downvote, and then the downvote is revoked, they will have rep=3.
Let's look at that in chart form. First, here is all the data. Clearly the users with rep of 1 are anomalous.

Now, let's filter out those users with rep=1. Now we can see that there are anomalous spikes for rep=3 and rep=9. (Rep=9 happens with 1 upvote and one downvote.)

Finally, lets filter out all users with one-digit rep. We get a result that is much closer to an ideal Benford's Law distribution now.

0
feels left out of your analysis.