[mistercow]

It still confers a negligible advantage. If the attackers have gotten into your data server, why shouldn't they be able to get into your application server just as easily? All you're doing is slightly inconveniencing the attacker before they can derive your salt wholesale.

Always remember: a secure system is secure even if the attacker can see your source code. Any work you do based on assuming the opposite is a waste of time.

[drakaal]

Really? How about because my database server could be something like DynamoDB that I don't own. Why would the execution and the Data server have the same vulnerabilities. The whole reason to have two servers is so that you can say. "The only thing that can talk to the database is a server on the 'Inside'", but injection attacks mean you could possibly exploit the data server with no access to the execution server.

[mistercow]

Even so, you're proposing a very low entropy salt. If the attacker can crack one or two weak passwords, they can start limiting their search space immediately, and crack more passwords. As they gain more examples to work with, they will be able to work out how your salt is derived, and then they can start looking for salt collisions based only on the data that is stored in the clear.

At that point, your attacker has not only the ability to derive the salt and attack each password individually, but potentially the ability to generate rainbow tables for subsets of users with identical (or largely identical) derived salts.

The point is that at best, you're adding a term to the overall hardness of the attack. When you use something like bcrypt, you are multiplying the hardness of the attack. There is no comparison between what you've proposed with MD5 or SHA-1 and simply using bcrypt on a random salt, even if the latter has your source code in the latter but not the former.