Right, keep in mind i need to describe this by abstracting all possible confidential info:

I've been put in charge of a 10-year old transactional system of which the majority business logic is implemented at database level (triggers, stored procedures etc). Win2000 server, MSSQL 2000 Enterprise. No immediate plans for replacing/updating the system are being considered :(

The core process is a program that executes transactions - specifically, it executes a stored procedure with various parameters, lets call it sp_ProcessTrans. The program executes the stored procedure at asynchronous intervals. By itself, things work fine. But there are 30 instances of this program on remotely located workstations, all of them asynchronously executing sp_ProcessTrans and then retrieving data from the SQL server (execution is pretty regular - ranging 0 to 60 times a minute, depending on what items the program instance is responsible for) .

Performance of the system has dropped considerably with 10 yrs of data growth: the reason is the deadlocks and specifically deadlock wait times. The deadlock is on the Employee table. I have discovered:

So the reason for deadlocks is clear. I created a non-unique ordered clustered index on the field (field looks good, almost unique, NUM(7), very rarely changes). Immediate improvement in the test environment. The problem is that i cannot simulate the deadlocks in a test environment (I'd need 30 workstations; i'd need to simulate 'realistic' activity on those stations, so visualization is out).

I need to know if i must schedule downtime.
Creating an index shouldn't be a risky operation for MSSQL, but is there any danger (data corruption in transactions/select statements/extra wait time etc) to create this field index on the production database while the transactions are still taking place? (although i can select a time when transactions are fairly quiet through the 30 stations)

Are there any hidden dangers i'm not seeing (not looking forward to needing to restore the DB if something goes wrong, restoring would take a lot of time with 10yrs of data).